Ten days ago I had an experience with Gemini 3.8 flash that made me wonder if I was being routed to a different model under test. I was trying to use rocm with llama.cpp on my 128gb Strix Halo but could only get it to run Vulkan. I pasted the error message into agy and it proceeded to attach GDB to my GPU driver, reverse-engineer the kernel queue ioctl interface, and author an LD_PRELOAD C shim to get ROCm llama.cpp working on my Strix Halo. My jaw was hanging open the whole time.
completely offtopic but is rolling with rocm worth it? I spend a fair bit monthly on rental gpus for projects and going to upgrade at home instead, AMD has some solid winners here pricewise but get conflicting reports about using it for ML in 2026.
once upon a time it seemed unthinkable to use anything but nvidia but seems to have come a long way since I last looked, probably would be just pytorch and gemma 31B
I get the feeling the situation is only going to improve longer term so might be a good time to just do it
3.8 Flash is just quite good, and so is the Antigravity harness.
I use a mix of Fable 5.1, Opus 5.5, and Gemini 3.8 Flash and Gemini holds it's own. Especially in writing, frontend, and sysadmin work. agy for configuring a NixOS system has been truly incredible.
What basic features are missing from agy? I've been using it and cli-cc + web-cc for months (among a few other random harnesses to test here and there) and they all seem roughly comparable to me.
I actually just cancelled Ultra also because I couldn't subscribe to a YouTube Family plan while I had it active (Google... :[) but trying to use Codex as a replacement while I testdrive Astra makes me yearn for agy again.
In the olden times, aka like two years ago, AI chats would just stop working or just start slicing off the oldest parts of the context to fit the model's window.
That said, compaction feels like an idea that should work reasonably well, but across all of the major providers and agent tools I've used has never actually produced compelling results, to where if I see I'm getting close to the token limit I prefer to start putting a bow on the project and readying it for a fresh start. Even when I provide a detailed compaction prompt it usually focuses on the wrong stuff.
It certainly has compaction (since the public launch I assume) and I HATE it. I have some remedies but nothing perfect yet. It never retains ALL the crucial bits. If a conversation runs into two compactions it is often a sign that I have to abandon it and retain whatever I can, to form a seed prompt for an adjacent conversation.
That is really the biggest beef I have with agy over others, the forced auto compaction at the 250k token threshold (3.8-flash), while the model itself (via API) would be fine with a 1M context window. Even if the model is great, restricting context to 250k tokens (and auto compacting no matter what) limits certain applications and workflows somewhat.
agy cli does not have auto mode. I've tried and tried and tried to work with agy cli sandbox-mode and just failed.
agy --dangerously-skip-permissions
in my experience is the only workable solution that doesn't ask confirmation for every step. And I hate working in YOLO mode. Seemingly the Antigravity GUI had some features added in a recent release, but a) I don't want to work with the GUI and b) it was poorly implemented as I couldn't get it to work. VS Code plugins are allowed with subscriptions, but is not the CLI experience of Claude Code I want.
gemini-cli supported 'pre-write diff tabs' (y/n) in external editors like vscode. In Claude Code I heavily use 'pre-write diff tabs' for documentation and miss it sincerely in agy cli.
IMHO Gemini 3.8 flash is fast and good enough, but the agy-suite is below par to say it nice. Someone else in this thread calls agy a terrible harness which is probably more accurate.
Auto mode means that another model reviews tool calls to attempt to disallow less safe ones. It's different from bypass permissions mode which typically just doesn't filter at all.
I use a variety of models for various subagents. I don't want to change my harness every time I change models, or be beholden to companies for something the open source community can handle better.
I use Antigravity but for some reason, `agy` in the command line feels very bad/incapable of doing things. I can't quite explain it but the most common issue I run into it is just hanging on being unable to finish a tool call
Can confirm, I was doing a routine internet search thing for a curiosity 3 days ago (about the only thing I used Gemini for) and was surprised by how suddenly thorough and quality the response seemed, almost overnight.
Adding my anecdote, because it amused me: I finished wiring up the compute/sensor box for my robot, ssh'd in and told agy "I have a Livox Mid 360 Lidar connected to this Jetson orin nano, setup a full environment with docker, cuda, ros2, foxglove and get it all working so I can see the lidar output". It did all the local config for the lidar, setup docker and the ROS2 environment, then told me "open up this url in foxglove" and sure enough everything worked. Whole thing used up 6% of my weekly limit.
For getting redroid running on my Linux system, 3.8 Flash decided to binary patch a .so file instead of getting the AOSP source code and patch/build it properly.
And I saw it do this twice, once for Android 14 and once for Android 16.
I think this is just within 3.8 flash's capabilities.
Astra also really loves reverse engineering binaries. I guess it's one of those things that isn't that complicated but is super tedious, and tedium means nothing to AI.
Can confirm - I am HEAVY claude user, but always like to check with AGY and CODEX in between. AGY with Gemini 3.8 flash cooked last couple of times and CODEX is basically out of the mix for me
WHY ARE ALL OPENAI MODELS SO CHATTY - i thought claude kept going on, then i literally put it in claude.md that summarize your thinking in 200 words or less and tell me in points what you did and what's next. Did the same for CODEX - nope still keeps effing going on and on and on
My experience with Gemini 3.8 Flash has been awful; it gives me the most hallucinations out of the major models. I'm not using it for coding, but general research on different topics.
I've been tinkering with Gemini for several months and I think it's great. The most complex things I've had it do is create a rust emulator from a compiled game, as well as create a buildroot linux image, trouble shoot problems etc.
Gemini is honestly amazing sometimes. If they didn't force you to use a terrible harness, charge too much for way too little, and generally act like customers are a giant problem to be avoided I'm sure Google could take over the AI market.
what is so terrible with their harness? I've been using gemini cli, now use agy, Pi agent harness, and agent (cursor), and my only real issue with agy was the permission handling, but other than that, it was ok.
Also I hope you don’t have children, eat meat, travel, have a car, run AC, buy things in other countries and such. Those things all take way way way more natural resources.
All your examples are private goods: excludable and rival. If one person uses a unit, that prevents others from using them.
Patches to open source software are public goods. Your using them doesn’t prevent others from using them. So if you spend resources creating a public good, it’s in everyone’s interest to share it.
If action X takes a million times more resources than action Y, it's silly to focus on or highlight action Y. Seriously: if you are a regular meat eater, your choices use several orders of magnitude more water than even a heavy LLM user. A quip from a comic doesn't somehow erase that or make it irrelevant.
I had a similar but less impressive experience recently with Muse Spark 1.3.
Asked pi agent it to identify the main hero sprite size of game I was running. It had a ton of shader effects so it was hard to determine.
It used some cli tools to identify that it was a game made with Godot, decompiled the executable but data was encrypted, broke the encryption after writing a brute force tool to test keys extracted from the exe, then proceeded to extract the game gd scripts and assets, only to answer the question of the sprite size.
The important take away here: the leapfrogging we’ve seen this year doesn’t seem to be a temporary thing. The famous theory of Dario Amodei was that AI was this winner-takes-all field where the first team to get a head start would never cede ground back. The term he liked to use was, “concentrating”. This is yet another datapoint that he was wrong about that. AI seems more distributed amongst neoclouds and traditional hyperscalers, FAANG and startups, GPUs and ASICs than it did this time a year ago.
Google has TPUs, a frontier model, a completely separate and lucrative revenue stream they can call on at will, and teams working on multiple different language modeling strategies simultaneously. Did I mention the vast and ominous data centers that already serve a significant fraction of the internet? If that ain't a moat, then what exactly is a moat?
Recently there's been a lot of talk about how Google missed the AI boat. Now today with Gemini 4 they're back in the boat. But give it a few months, people will be counting Google out again. This has been a repeating pattern for well over a year now.
Not to mention - they have the internet already indexed (they have a local copy), all books, and youtube. Beyond that, everyone "connects" to Gmail, but google HAS Gmail.
> Then why have they been lagging behind OpenAI and Anthropic
Because they're not desperate. Slow and steady wins the race, at this rate all Google has to do is wait for OpenAI and Anthropic to exhaust themselves on aggressive training, then they can casually amble along right past them.
This is basically what people were saying about Microsoft versus Google and other upstarts in the early 2000s.
Slow and steady wins the race. How could they lose to something on the web, when Microsoft owns the web browser itself? Everything runs on Windows and IE. They can just relax and wait for competitors to exhaust themselves, then quickly build their own version. Isn’t that how Netscape lost. Etc.
Now Google is the new Microsoft, just like Microsoft became the new IBM.
> This is basically what people were saying about Microsoft versus Google and other upstarts in the early 2000s.
They would have been right if Google marquee product was an Office Suite
Google was the AI company before AI companies were a thing. The comparison with Microsoft and IBM are misplaced because they failed to capture new territory; ML/AI is Google's stomping grounds. The criticism that Google is bad at consumer chatbots is true, but that's not where the real future value lays.
For some of us OpenAI is already the new Google. I prefer chatgpt even when I'm searching for web links. Google Search had already been deteriorating terribly even before the rise of LLMs.
Also because they are Google. Google being Google: unreliable (they could kill a product anytime), too much of a platform risk (all products in a single place, get banned and lose the company), lack of support unless you really pay a big bill (into the 7 figures), among other… things.
Incredible how the idea of Google being such bad choice has become so entrenched that people prefer the offerings of two companies that might not exist anymore in a few years over Google‘s similar offering.
>
Then why have they been lagging behind OpenAI and Anthropic for most of the last few years, and only briefly been at the frontier?
One possible explanation: because Google is a little bit more frugal and focuses on how to make providing AI models financially feasible - combined with some willingness to burn money so that they don't strongly fall behind on their AI models.
On the other hand, OpenAI and Anthropic at least formerly concentrated on building and providing the best models that they could with concerns about financial feasibility taking a backseat.
Just to be clear: I do have the impression that by now (likely because of pressure from investors) OpenAI and Anthropic take these financial concerns more seriously, but nevertheless Google's vs OpenAI's/Anthropic's "DNAs" concerning on what to focus on differ.
Google never really has to outpace the competitors (other than to have some relevance) but they have a very large group of business customers using them for business process work in Gmail, Docs, etc.
Clearly they will win when the models are close enough to frontier to be good enough, but are long-term cheap for buy. I.e. they will aim to make it a commodity.
In theory MS has the same opportunity (plus they have GitHub so, you know, dev eco system too) but seem be blowing the strategy.
Anthropic and OpenAI are having to race to the top on ability entirely to keep their name in the media and in front of us all (which costs: hence more recently trying to pivot away from model releases and more into controversy/danger). The main cost is in training and so this strategy is much much more expensive and this will play out either as a huge cost hike or a forced slow down in pace.
I believe essentially Google is betting on that & I think it's probably the right strategy.
> Then why have they been lagging behind OpenAI and Anthropic for most of the last few years, and only briefly been at the frontier?
Because it's not an existential battle for Google. If OAI or Anthropic disappear from the absolute frontier for ~8 months the news cycle and churn will diminish them to the second rate. Google is processing near 4 quadrillion tokens every month, that's - I'm sure - significantly more than OAI or Anthropic, because Google is interested more so in their flash models and getting these competitive, which they are.
From a business perspective a frontier model does not make much sense anymore if you are not a startup. Neither for Amazon, nor for Google. Their clouds need models that are fast and perform well in their agent frameworks nothing were a frontier model excels at.
Most Google products even use flash lite underneath, so their frontier model is mostly used for distillation.
>
From a business perspective a frontier model does not make much sense anymore if you are not a startup. Neither for Amazon, nor for Google. Their clouds need models that are fast and perform well in their agent frameworks nothing w[h]ere a frontier model excels at.
A good consideration; just one point from my side: as far as I am aware (but I may be wrong), Gemini is not known to perform well in an agentic framework.
This is no contradiction to your other claims, quite the opposite: perhaps (or even likely) Google wants to avoid that their models become a commodity in some (agentic?) application where the middleman who actually writes this application gets a disproportionate of the money that the customer of the application pays for it.
Well at the moment we do not use Gemini in an agentic framework. But as said the flash and flash lite families are basically their driving force in some of their applications in Google cloud, like document ai and its ocr capabilities are probably better when it comes to business documents than any other (at least in perf to cost to speed).
We also drive Gemini lite in our application where customer can use it to generate simple automation, like an agentic framework but way way smaller scale. And while it struggles in more context heavy operations it still is a beast when feeding it one or two pdf documents and asking questions about them and it’s hella fast.
> A good consideration; just one point from my side: as far as I am aware (but I may be wrong), Gemini is not known to perform well in an agentic framework.
I used it for a month over the summer, right before they were going through the migration to antigravity. It was a fine workhorse IMO, no complaints from me.
I don't think that other revenue stream is completely separate. It weighs on them as they need to think about tradeoffs. Classical search is going away sooner or later so they need to replace that with AI powered search.
Data centers are important but a few others also has them: Amazon, Microsoft, Meta. SpaceX will likely be in/at the top I AI dedicated precessing power in 2027 as well.
I don't see the moat. I see a company with a lot of other commitments that is not the best at delivering consumer facing products. They have some good cards but so do others.
> Google, Microsoft, or Amazon are more likely to be the AI leaders than OpenAI or Anthropic.
If not now, then when will these companies be AI leaders?
Even Google, with its staggering advantages in cash, compute, real estate, training data, and having basically invented the field only manages to briefly claim a 1-2 week lead once or twice a year.
The financials for Anthropic and OpenAI are likely borderline suicidal, google and co are publicly traded. Moreover, all innovations downstream to them dont they? Why not just stay slightly behind, especially given many have stake in those other companies?
There are many companies that have data centers. They are conceptually easy to build. An ASIC is difficult enough that if you make one someone will leapfrog you while you are still making it (at least so far), though once you have one your costs will be enough lower than the competition that you can perhaps undercut them.
If Moore's law continues, then in less than 10 years today's state of the art model will be able to run on a cell phone. How much smarter do we actually need AI to be? Would it still require datacenters and custom hardware?
Moore's law stalled ~2015. Unfortunately, no way current models will run on the <100W thermal budget of a cell phone. Printing the weights directly into a chip would help efficiency a lot, but not enough.
They probably said the same thing about social media back in the day.
I'm sure the thinking out there, and hence investment, is all about how to tether the user to the most addictive, network-effected, incredibly deep, server-side, moat-able version of AI possible.
For now, for cloud training. but for consumers, nvidia vs amd reasonably close - the moat there is thin and shrinking. I suspect AMD will surprise us. nvidia has no motes in china, which may be a new source of (gpu) chip design. Huawei's Ascend 910C is about a generation behind... again: for now.
point is: moats dry up. I see nvidia's shrinking as a real possibility.
The whole winner-take-all idea seems entirely based around Singularity/Rationalism and would require massive advances that we probably aren't close to at all.
yeah this would be one interpretation where "winner take all" could still be right. though with the recent accelerating release cadence, doesn't it seem like the head start / lead has been shrinking over time?
My personal theory is (assuming there really is no moat) whoever starts the latest with developing AI models might actually win as they should be able to develop a competitive product with significant less resources and initial investment resulting in a higher ROI. AI might even become a commodity.
Based on historical developments the cost of compute will go down again eventually, decreasing the cost of training AI models of the same quality as today even further. That part is what I would be the most certain about.
> Governments should conduct safety testing on sufficiently capable AI models, including successors to GLM-5.3. Without high-quality evaluations from independent sources, the impact of these capabilities might not become fully clear to model developers until it is too late. As AI developers across the world build increasingly capable open-weight models, we hope they work to appropriately safeguard these capabilities and prevent misuse.
I for one do not think my government is up to the task of designing or implementing such a system
You say that because the 'most' existing models have done is hack governments and companies. Can't you think of worse things a model could do; accidentally or by instruction?
help people with suicide and school shootings like ChatGPT already has
OpenAi is alledged to have been monitoring these internally and not contacting authorities. Lawsuits have been filed, I see gross negligence without the gory details
I have for more concerns around human-chatbot maladies than I do around the cyber security stuff. For example, why hack grandma when you can get her to do something willingly through impersonation. How do we prove authenticity in a post truth world?
The problem is twofold. One, even a monopoly AI provider wouldn't have pricing power against its suppliers. Its suppliers are energy, semiconductors, and real estate. Semiconductors maybe they could get some leverage on but energy and real estate have plenty of other buyers. Two, there's still no evidence of a runaway scenario (ie a small lead turns into a big lead over time) and there's still no evidence that there's some resource that you can deny everyone else that they can't build your product also. You can't hoard energy, compute, memory, data, human talent, or customers.
The net effect is that the most likely scenario is if one big lab fails, they will likely all fail. Their revenues are all correlated.
To go to your dotcom comparison, the winner will be the ones picking through the assets that were written down by orders of magnitude and trying new products with the technology until one sticks to the wall. But I don't know if a dramatic crash is guaranteed either.
>
The problem is twofold. One, even a monopoly AI provider wouldn't have pricing power against its suppliers. Its suppliers are energy, semiconductors, and real estate. Semiconductors maybe they could get some leverage on but energy and real estate have plenty of other buyers.
Concerning the leverage on energy and real estate: don't forget that the AI companies have quite a lot of choice where to build their data centers. So AI companies have lots of opportunities to play several parties off against each other (in particular also for real estate and energy).
"but it can stay there so long as the balance sheet doesn't deteriorate."
Uhm, what? LOL.
People dont value firms based on balance sheets fella. Have you taken a basic valuation class?
Tesla is a nice stock for traders - they like the volatility. Nobody holds Tesla as stock for investing. If you were to truly value it on an intrinsic value basis you'd have to bring in failure risk.
I suspect this is going to end up like most services provided e.g. cloud stuff, balkanized between a couple major players and an assortment of DIY or less popular options if you don't like those ecosystems, plus some UX/DX focused wrappers that use the big players under the hood.
I think that would be a pretty satisfactory outcome compared to one hypercompany consuming trillions of dollars of the world economy.
“Divide the world” sounds ominous. Here’s another scenario to consider:
Internet access is not really unlimited, but for many people with fiber at home, it effectively is and we pay a flat rate.
Perhaps by the end of next year, most programmers will stop thinking about metered access for AI? For many people, the cheaper models (about as good as today’s frontier models) will be good enough.
Which might sound good, but the downside is that it will also be easier to build an AI botnet without the users paying for it noticing. Particularly when people are running AI inference on their own hardware.
Hardware is still insanely hard to get a hold of, and the stuff that's being built doesn't really work for home use. Maybe if it crashes Nvidia will adjust the hardware flow.
My guess is even if the AI market busts there is still a massive demand for hardware as models are solving all kind of problems now.
But ya, lots of hardware everywhere not managed well is how you get sovereign AI.
Or, like airlines, the ones that are left will have great technology but be not so great from a business and financial perspective. To me AI seems like a commodity service.
THe problem with analogies is that they are imperfect.
I would argue those who already rule the world, will continue to do so.
What happens to OAI and Anthropic? No idea, probs go bust. Google just has to offer a half-decent offering in the long run and have a cost-advantage and it'll eventually knock OAI and Anthropic out as firms figure out what combination of models they want to be best for their economics and generating returns. Enterprises trust google over OAI and Anthropic. A clear signal of this was the Apple deal.
Dont forget those sweet returns fellas! CEO's are hired to make the owners wealthier. That is not gone.
I guess if one of them hits singularity, it could in theory just wipe out all the rest, seeing how they keep escaping and hacking into other systems :)
>
I guess if one of them hits singularity, it could in theory just wipe out all the rest
The story that some AI company might reach singularity and then "everything will be different" is another science-fiction story that executives of AI companies love to tell to justify the staggering amount of necessary investments and cash burn. :-)
I find it quite unique how many people buy into this. Its the worlds most blatant conflict of interest, I dont even know why Sam and Dario bother doing interviews
I think some in the AI industry drank their own Kool-Aid. They believed that if they had the best model and the most compute, they could tell the model, "Make a better model." And it would, and the next one could make its replacement, and so on.
So far, that's not exactly how it's played out. Humans are still necessary for the leaps in capability or efficiency. A model can grind on a problem to eke out the most performance, and models can synthesize data and iterate on various techniques to find the optimal combination. But, seems like humans still have to provide the real thinking, and the talent and drive for doing that is not concentrated in one company or city or even one country. And, (surprisingly) a lot of the people involved are in it for advancing the field more than making another billion dollars, so they're publishing their research.
So, yeah, the moat isn't deep. Even the compute moat, that OpenAI, Musk, and a bunch of other also-rans (like Oracle) bet the farm on, isn't really panning out. The Chinese makers just spent their effort on making models vastly more efficient, since they couldn't do anything about having an order of magnitude less compute available.
But that's the whole point of the singularity. Right now the models use a lot of human effort and ingenuity to improve the models, but about a year ago it was 100% human. We'll see in another year, but if this pace continues I doubt there will be more than a handful of people who can contribute more than the models.
> I think some in the AI industry drank their own Kool-Aid. They believed that if they had the best model and the most compute, they could tell the model, "Make a better model." And it would, and the next one could make its replacement, and so on.
They're not there yet. Once they get there, that's literally the definition of Singularity.
But they are getting closer. Recursive Self-Improvement used to be a phrase people mocked LessWrong crowd for using and worrying about, now it's something both OpenAI and Anthropic already publicly admitted not only to pursue, but to already be benefiting from.
Sure, it's happening...but, is IT happening? By that, I mean, we can see that the models are able to iterate at a pace and scale that humans can't match, and that provides gains in model performance and efficiency. But, humans are still needed in the loop, and not just because it's necessary for safety/alignment reasons. I don't think any significant discovery has been made by models on their own, and I don't know that LLMs will ever have the capacity to invent. They can synthesize from known data amazingly well, and since they know everything "known data" is extremely broad. But, the leaps, so far, have all come from humans.
So far, I don't think the models are capable of running away on their own. Of course, it would be playing with fire to not at least consider the risks of such a runaway scenario and build in safeguards against it. But, there is no model that can build a better model on its own, thus far, to the best of my knowledge (which is far more limited than the models, so maybe I should ask them).
Recursive Self-Improvement isn't instant, it starts slow and accelerates.
It starts with what they already claim to be doing - increasingly relying on existing models in non-trivial work related to training, evaluating and optimizing the next, more capable generation of models. As long as the proportion of work keeps shifting towards agents doing more and more of it, and humans less and less, that's RSI at play.
It may be that it turns out LLMs lack some fundamental level of judgement and it plateaus, but frankly I find this notion absurd; LLMs already show better judgement than most people. The alternative is, at some point LLMs will show the ability to futz their way into improvement of the next generation of models even without humans in the loop - even if much less efficient at first, if generation N+1 is more capable than generation N, it'll either take off or burn out.
They absolutely do, unless you believe they are lying about the fact they're using current generation models extensively to develop the next generation of their models.
The letter S stands for "Self" and word "using" is for sure not a superset of the word "self". Basically LLMs are assisting someone who does improving of said LLM, while RSI is a carpal... ahem, RSI is "self" improvement, meaning no intemediary in a human form. PS: it's also not recursive but iterative improvement, even if it ever happens.
I just find this unlikely personally, think about the great research that's happening in the open source world, I'm sure inside anthropic + openai they've also made a bunch of discoveries and improvements (and I'd guess way more due to them attracting the best talent + the better internal models they have)
Maybe. Or maybe having the best AI model on the planet becomes like having the best super computer on the planet. Useful for some niche stuff, but not too useful in terms of people's daily lives or what is used in business.
Already businesses that have more compute and access to data seem to eat the world around them. If, and ya its and if, we can make something that self learns into RSI it's not looking like any business that came before this.
There have been external testers, so we know it exists.
But yes, the word "leapfrogged" is premature until it's available, as we know the other labs also have internal, unreleased models and don't seem to announce releases beforehand.
I'm not necessarily defending this obvious marketing speak but maybe the "starting point" was wider than assumed. So far, nobody has caught up to US and Chinese labs for example despite lots of funding in Europe. This is also despite abundant in-depth research papers being published alongside open source code and weights by some Chinese labs
The US companies still have trillion dollar valuations like there is a monopoly. There just isn't one. They are all within a few percent of each other on the benchmarks.
The slightly lower Chinese open models are good enough for almost everything, too, and much cheaper. Like with humans there is plenty of employment for people with below genius level IQ's.
I feel like the frontier labs are going to serve fast/lower intelligence models at a better per token cost than the open chinese models. You're telling me that in the long run, you're going to self-host your own ai infra for cheaper than google can serve it to you? I don't really buy it. I think the dedicated AI data centers are going to serve AI at a lower marginal cost than random businesses self-hosting, and then it's a question of how much of that margin they can capture.
Yes, it may come down to how well they get efficiencies from scale.
People always compare the inflated API prices, but subscription prices of American models are competitive for the intelligence. You get >20x the subscription cost in tokens.
Agree entirely but that's the point, if it's a margin knife-fight with marginal product differentiation/pricing power nobody is going to be making bank.
Is there a dividing line between good enough and best in class capabilities? It's blurry from where I stand. Will model makers cede ground or is there a market making moment up for grabs (singularity)?
> And Chinese labs openly publishing so much of their methodology destroyed any hope, which was inevitable
I think the secrecy doesn't make sense. People swap jobs between labs so I'd say the big players can' really keep secrets for long, and any secret sauce advantage gets incorporated by competitors in a major product cycle at most.
I have never understood the whole "this is a winner take all game" mentality - the sheer size of the pie is so great that from a purely rational standpoint companies should just be trying to productively get a slice of it and be profitable. winner-take-all is just greed/capitalism run amok, where it is not enough to be profitable, you have to own the entire market (and presumably extract rents)
It’s like the supposed first mover advantage OpenAI believed they had. In practice it’s almost always more like a first mover massive tax, and companies coming afterwards benefit from your discovery of a market, publicity, and everything else that has already been validated
The present leapfrogging is not a contraindication because companies are not necessarily releasing their best models; we know they have smarter internal models. Furthermore, humans are still involved in model creation. Human involvement is expected to decrease over time, and when model iteration is completely automated, progress will happen at the machine's pace, leading to runaway intelligence, barring any ceilings.
AI is a commodity. One that is showing to be more readily commoditized than most has anticipated. As of now, the only moats are the financing for the hardware to run it and the hardware vendors themselves - with the latter largely not yet a commodity because of ecosystem lock and a limited capacity of the most advanced fabs in the world.
> We’ll continue to gather feedback from early testers as we iterate on guardrails before making Argon available to developers, enterprises, and consumers as soon as possible.
Gemini not beating the "can't release a model" allegations
My Gemini app (updated today) and https://gemini.google.com/ has _3.6_ as the latest selectable model, as a paying Pro user in the US. How is that even possible? Gemini 3.7 was released in August, 3.8 early September. What is going on over there?
I would imagine you are experiencing a bug. I've been using 3.8 daily since its release (on a Pro plan in Canada). I believe this is true of many people.
What is your reason to believe this is not a bug specific to a small set of Pro users?
Would that change anything about the conclusion? Having a "bug" that changes available model options for some small set of Pro users 3 months after launch certainly qualifies as a wtf-are-they-even-doing level of bug in my book.
That's weird. I got 3.8 and 3.7 on the days they were released. https://imgur.com/a/Xu4wRLM (this is on gemini.google.com, but the same is true of the iOS app and the desktop app).
They will go through the usual transition of "can't release a model" to "won't load in a harness normal people can use for 3-4 weeks" to "it's smart as hell but completely inept at tool use and coding" like every Gemini release.
They're just following the current AI marketing playbook. "Our new model is simply too dangerous to release to the public right away" is now standard practice.
They even gave their model a random nonsensical name suffix simply because OpenAI is now doing it, too. Monkey see, monkey do.
I'm still at a loss as to what argon has to do with anything. Say what you will about Luna-Terra-Sol-Astra, or Haiku-Sonnet-Opus, they make sense. I don't see how Google can make sense of argon; it's in a fairly strange place in the periodic table...
Yeah, I heard the next one was Barium...or was it Boron?
I was going to say I don't know what they'd do for C, since Carbon and Calcium are already things. But knowing Google, they'll probably call it Chromium.
Opus 5.5 and Sol 6.1, literally state of the art (in their respective class), were just released without any prior announcement. This has pure and simple become a Google thing.
I don't read that as the same category: There was no announcement, no benchmarks, no limited release and no promises about what will happen with that model. It failed internal safety standards. Might be scrapped entirely due to a failed training run, for all we know.
Great, so they _finally_ decided to add a non-flash model and it's not available to regular subscribers for an indefinite period. What's the point of paying for the AI Ultra plan? Anthropic doing the same with Fable as far as I know, OpenAI at least allows Pro plan subscribers to use Astra. I subscribe to Gemini AI Ultra and ChatGPT Pro, and have enterprise access to Claude at work. To be fair, Gemini's flash models since at least 3.6 have been quite useful for non-complex work, but for any task where there is a bit of complexity involved, I've had to check and recheck the work multiple times myself or sometimes with another LLM to get it to follow plans accurately. It's disappointing to see yet another Gemini release ignore adding newer pro models.
Edit: seems I was wrong about Anthropic restricting Fable, I guess our enterprise plan doesn't include it. But, the block from Anthropic regarding Mythos for regular subscribers/enterprise-users is still true I think.
As we’ve seen from the leaked Anthropic prospectus, revenue from actual users is a pittance. What really matters is what you can get from investors, and that you have a model smart enough for self-improvement.
Fable is available for a couple of months and even got an update on 1st of September. It’s really good, but since Opus 5.5 was released, there is not much point in using Fable anymore
> Argon agents are working on migrating C/C++ codebases to Rust across Google
Man, I remember back in the days when the cppnext team was refusing to even consider Rust, instead looking at absurd stuff like Carbon and Swift (!), even though half of the engineering staff already knew where this was headed. I hope they got a few good promos out of the delays at least.
A RewriteInRustBench would be unironically useful at this point since all the main agents can write it reasonably well despite its relative scarcity in the input data.
Rust is the best language for LLMs b/c it gives by far the best debug messages. Just tons of verifiable reward signal for post-training. Even the most rudimentary LLMs can school me on idiomatic Rust
I've narrowed in on only using Go or Rust generated code (Go for APIs right now) and rust for some TUI or other thing. TS for web interfaces (w/React).
On the other hand, Rust's borrow checker is very picky, and even a frontier LLM still sometimes struggles to respond to roadblocks sensibly (refactoring so whatever it's trying to do can be done safely) rather than stupidly (introducing some horrible global arena thing so it can make the borrow checker go away). A lot depends on how good your instructions are, and how good the existing code is, since bad input begets bad output.
I've (more or less; I've read quite a bit of the code) vibecoded several houndred thousand lines of Rust and I've not seen this happen a single time. It sounds like something it'd do when you ask it to "write a linked list while satisfying the borrow checker". Are you sure you haven't (possibly unknowingly) been giving it instructions which ended up luring it into doing these things?
Not always. In my experience, if you're not working on a small, trivial codebase, LLMs will sometimes just create spaghetti unreadable, inefficient code to satisfy the constraints of the type system/borrow checker.
Carbon was clearly DOA the moment it was announced, IMHO. It looked cool but it served none but Google, and now with LLMs you have a massive incentive not to use a niche or new language due to how better LLMs get the bigger the corpus is
The only somewhat realistic proposal in this space is Herb Sutter's cpp2, which is arguably a massive improvement and I'm puzzled why nobody in the standard thought to give it a spin, there's just to much cruft they'll never be able to get rid of unless they make an alternate yet backward compatible syntax with C++ that changes the defaults from "random 80s nonsense" to something better
Cpp2 has been dead for a while afaik, while Carbon is still going.
I don't think Carbon is dead, it just all depends on how easy it actually is to rewrite "all of C++" in Rust. (The jury is still out on this one, but it's not looking good.)
Version 0.0.0.0 after 4 years. Their goal of "full interop with C++ while being a completely new language without any of the flaws of C++" is plain absurd.
It's DOA because Google doesn't have any idea of what Carbon should be, and to be completely honest, at least 80% of what they currently use C++ for should be rewritten Go, you know, that language developed specifically because of the issues with C++ by teams within Google.
There is 0 practicality in inventing an entirely new coding language that only one company uses, and you have to teach it to thousands of new engineers. Rust exists and fits the job totally fine and is used in more places and has actual support outside of a single entity (i.e you can actually hire people that feasibly know the language).
It was clearly done because some PL guys at google really wanted to make a new cool language and Google was the perfect place to incubate it without it getting axed. Probably got a couple of promos out of it too. This is clearly not the best use of time or money, but I guess if you're google you have so much of both it probably doesn't really make a dent, and you can keep a few very smart people happy with shiny new projects.
Also, LLMs being used for a large portion of coding nowadays sort of remove the need for these types of languages, IMO. They make less "silly" bugs (both logical and structural) that languages like this are meant to catch, and they are much better at languages that are better represented in the training corpus. This somewhat obviates the need for very niche "type/dummy-safe" languages like carbon (and even rust/zig, imo). So even if you did want to use Carbon, you'd likely have to bootstrap a decent amount of your own "good" carbon code to post train an LLM, and even then, it likely won't have that big of a gain vs just having an LLM write C++ or even Rust. If you are a company that still reviews code, you should just have an LLM code in a language most people can understand anyway to make verifiability tractable.
Rust is not the end of history. One of the difficulties with the language lies exactly with porting existing code written in an OOP style to idiomatic Rust, as those codebases weren't written with ownership in mind.
Such rewrites will contain judicious uses of Cell, RefCell, unwrap() etc. which make for ugly code that's not exactly simple to understand and might even have some landmines (crashes).
Getting rid of these requires a subtantial amount of engineering effort, which I'm not sure how well these LLM manage.
Given the nigh-universal experience of LLMs producing an ungodly mess when left to their own devices, I have my concerns.
Hmm, I don't disagree with you that LLM's remove the need for type-safe languages, but as the blog mentioned, Google is porting their C++/C code to rust. Does this mean the port is waste of time and that they should just rely on the LLM's to catch memory errors?
I mean Rust definitely has a better tradeoff than Carbon in this case, re readability/verifiability by a person (and sufficiently good internet training data).
I personally think that you _could_ use an LLM to catch these types of boundary case errors without having to port the _entire_ C++ codebase to Rust, but maybe pre-emptively porting to Rust now can catch some of these cases for cheaper than doing a full LLM sweep. Also more cynically, its a good benchmark lol.
I guess if you really believe in curve of LLM capabilities you should just use a language that has the best performance, safety, flexibility, and extensibility, since in the limit few/no people will actually read the code anyway. I think this ends up being Rust.
I'm not an expert on this. But isn't it the case that C++ code could have errors that span the entire codebase, like a setup in file A triggered by a bug in file B which is immensely far away on the import graph? A classic would be a use-after-free. To me that's the thing that Rust can help with, even if silly bugs aren't being written by AI.
The other thing is just that rewriting some old human-written codebase in Rust probably immediately catches many bugs. It would be hard to prompt the AI to properly scan for such bugs itself, they're lazy when working in that modality.
> There is 0 practicality in inventing an entirely new coding language that only one company uses, and you have to teach it to thousands of new engineers
They did that for Go and it seems to have worked out for them though.
I’m not entirely sure Google should have both Go and Carbon but when you have billions in server costs it makes sense to do extreme stuff for even basis points of performance. I’m still surprised at how much java there is.
I wonder, why Google don't make Gemini - open weights model?
Considering, Gemini 4 is in the same ballpark as SOTA models, just open source it and kill any competition from openAI and Anthropic, and be market leader.
This will be so good on so many dimensions - buying time for Google to iterate on next model, best for all folks like us, kill funding or destroy valuation of competitors and force them to be open up their model or force them to a create a much superior model than open source Gemini.
Only downside, is revenue loss from Gemini API, which I am not sure is really significant as compared to Google other revenue sources and a part of this can be captured by GCP, as you need to host the model somewhere.
They do release the model weights for Gemma. But could anyone actually run Gemini 4 weights? It's probably like a 10T model, which you need an industrial rack for anyway.
> Large Scale Codebase Migrations and Optimizations: Argon agents are working on migrating C/C++ codebases to Rust across Google—scaling from tens of thousands of lines in core libraries like re2, libgav1 up to 800K+ lines for the Fuchsia OS Zircon kernel.
To me this is way more significant than other random c++-to-rust-AI-rewrite. If they can pull it off on core C++ libraries en masse, I don't know if C++ will still be relevant in a few years.
I look forward to a post from google on this effort.
> I don't know if C++ will still be relevant in a few years.
And people are worried about human extinction when this is the potential trade-off!
C++'s death cannot come soon-enough.
Seriously though, things have changed so incredibly rapidly in the past year or so. I have never been such an efficient or such a proficient engineer than I have this past year (delivering feature after feature, project after project, faster and better than I could before with better feedback from users etc) and I don't even see the code any more. It could be c++, it could be python, or java or what ever - I don't really care any more: the computer deals with that trivia while I concentrate on what to build and how it should work.
"I don't know if C++ will still be relevant in a few years."
The standards body members are still fighting about whether memory safety is important enough to change the language for, so, I would guess the answer is "no".
Yeah that was abundantly clear when Bjarne Stroustrup published "A call to action:
Think seriously about “safety”; then do something sensible about it"[0] as a reaction to NSA's recommendation to no longer use C/C++.
Well, sure, but that doesn't mean it's not in a decline that is very unlikely to reverse course. Fewer people this year are starting new projects in C++ than last year, and fewer people than that will be starting new projects in C++ next year. And, as the models get really good at porting and the cost (both in terms of tokens and human supervision) comes down, there will be an avalanche of ports from less safe languages to safer languages.
This year, maybe next, maybe a year or two after that, is probably the most C++ lines of code in production use there will ever be. Why would one choose C++ for new projects at this point? There are niches where Rust is still uncomfortable or just doesn't have the support, but not for much longer. Models are very good at Rust and good at porting to Rust. And, Rust is a good language for models because it is so strict...it helps keep them in line.
If no one is reading or writing Rust code at that point, it all being agent driven, Rust itself is going to be a very short step before it gets disinter-mediated away and it's English -> complex heterogenous machine code across GPU/TPU/CPU/xPUs. Not sure Rust fans or C++ haters have thought this through.
> taking careful precautions against feeding the findings back into training so as to not risk shaping Argon’s reasoning to evade our monitoring. We strongly encourage the rest of the industry to preserve reasoning transparency in these pivotal moments of increased capabilities while navigating alignment risks, so that model thoughts remain helpful in identifying and diagnosing misalignment.
This is good, but they're the slow mover due to this exact thing.
Google is getting punished for not letting the models enter an echo chamber and go faster than humanly possible.
Mmh ok. How much theoretical speed or 'intelligence' gain is realized by allowing reasoning to occur in some inscrutable intermediate representation? Has this been actually tested, how much is it slowing them down, and compared to whom exactly?
Breaking news is not the model. Breaking news is that inside Google, it is being heavily used on large code bases for writing code and it is migrating 800k lines of C++ code to Rust already.
In this space, any other company that I respect other than DeepSeek is - that would be Google. They had been honest about it from the get go including their infamous "we have no moat" memo.
This company has enormous data, their own hardware (TPUs) and their own in house experts. Actually, LLMs are invented here.
Argon is doing my job for me while I'm writing this comment.
A guy at lunch today asked me when a feature was going to be built on the tool I'm working on. Turns out it had built it last night at 8:30 when I was hanging out with my girlfriend. Welcome to the future!
At this point I just think they are benchmaxxing and all talk and no action. I pay for AI plus because I wanted more storage, and when I go to gemini.google.com the most recent model I can use is 3.6-flash-lite. Two revisions have been released since then and they still can't put these things in the hands of customers. Why is it that other providers can get the models into the hands of customers right away? Google is meant to be the bigger tech company in the world.
I don't _want_ to use aistudio. The UX is confusing and I don't really know where it fits. Yet I can open codex or claude code apps or CLI and get real work done today with the latest models (even on the cheapest plans).
Google AI Plus is not the equivalent to most other paid plans, it is closer to ChatGPT Go. 3.6 flash is an equivalent model to luna 5.6, which is the highest available on ChatGPT's free & Go plans.
3.8 flash has been perfectly available to Pro users from the announcement day.
You are on the cheapest paid plan and complaining that you don't have access to more expensive models. (Though idk why you only have the lite version, I have access to the non-lite version even on the free gemini plan as long as I'm logged in.)
I recently subscribed to Claude, and was very unhappy about the usage limit of the $20 pro plan. Then I found a trick, since I got free Google AI Pro via my phone carrier, I use Opus 5.5 High for planning, and then dispatching agy to do works.
Most of the time 3.8 works fine, but it's a bit slow if compare to 3.7 Flash. If there's already a detailed plan, 3.7 can complete the task much faster. And the best thing about agy is the usage limit was very generous.
Argon will launch at an introductory price
of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price.
Wow
That's before they integrate a Jev solution, which should lower agentic workflow costs by ~40% and increase speeds by ~40%, while also increasing quality.
Everyone will be adding this soon, though I won't be surprised if Google is one of the first - and I'll be shocked if we have to wait more than a month and a half.
> Quantum algorithmic optimization: Argon is helping our quantum computing researchers optimize the spacetime resources (qubits × gates) of subroutines that bottleneck important applications. In one example, it beat the published baseline by 40% in a matter of minutes.
Amazing breakthrough! So useful in day to day life, glad they put this as the first bullet of how it is making changes at Google.
Argon will launch at an introductory price [1] of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price.
[1] After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.
===
So, they are basically offering opus 5.5 pricing. On AA, it scores around Sol 6.1 level (53) with avg cost per task $1.99 (https://artificialanalysis.ai/models/gemini-4-argon#cost-tab...) which is higher than Astra high ($1.73), Opus 5.5 high ($1.82), Muse Max ($1.60) and way higher than sol 6.1 Max ($0.72).
And this pricing is their 'discount pricing'. Add that to AI studio and Vertex's famously terrible caching, it is hard to see this as competitive. Google somehow is getting terrible advice on pricing (see also: the flash pricing fiasco)
But good to see more competition. I would happily take a 4 horse race (+google, +meta) than 2 horse race for US labs.
Interesting to see a mention of Fuchsia on a big Google announcement. Is the project still truly alive? Are the ambitions still as grand? Is the team as stacked as it used to be?
Had the same thought. On the wikipedia, it only mentions Fuschia used on the Google Nest Hub, which probably means it's used on a decent number of devices, but would think it was such a great OS, they would have used it for something like the upcoming GoogleBook.
Fuschia is not a desktop OS. It's designed for lower end or embedded hardware. Besides, Android has the highly profitable app ecosystem so it makes more financial sense to build GoogleBook based on that.
It obviously is pretty low key on the public relations front, but it's also very active as a project and I think it would be weird to look at their commit rate and conclude that the project is dead. If Fuchsia is dead then 99% of major open source projects are dead by the same standards.
I know someone who works for Google Canada with AI. Her parents and mine were friends and some thought something might happen there at one point in time..
Excited to see Google competitive at the frontier level again. Hopefully they sort out their infrastructure and model versioning so that we can feel confident building production applications on top of their APIs. The capacity limitations I've experienced with them in the past have been deeply problematic.
> Argon agents are working on migrating C/C++ codebases to Rust across Google
If anybody at google is reading this, please please pretty please prioritize or-tools. I absolutely love the project and use it all the time, but for the entire life of the project they've never had a repeatable working build system, and the whole SWIG framework is a nightmare to deal with. There's so much potential as an open source project, and a lot of external researchers would love to contribute, but the codebase is an example of everything wrong with the C++ ecosystem.
> Gemini is the model that is routinely borderline psychotic. It scares me
I'd call it the most sneaky out of the bunch. When I asked to explain something it will eagerly make things up and then claim it as facts. A lot of it likely because I don't pay for it, so it is reluctant for security reason or to save tokens to actually open a source and get the results. It just sort of guesses what the URL might contain, and confidently answers with some made up crap. When pressed it fessed up that it made it up. From my perspective it would be a lot better if it just said "you've reached the limit of
whatever and I can't do these things because x, y, z".
If there's a company that culturally doesn't understand alignment, on a human or systemic or AI-research level, it's going to be Google. (or Oracle, but they're not in this race)
Can you elaborate, please? If any, I see the other big labs with public admissions of AI "going out of control", which I suspect they almost want their models doing that because if helps with the narrative that would net them industry regulation, but that's besides the point, how is Google worse in that regard?
My use of Gemini recently makes it seem like it's almost bored with the requests being asked of it. It once offered to reverse engineer some obscure controller for an HVAC system for me, unprompted, only because it had trouble finding the manual pdf from a google search.
Absolutely because none of these models are ever trained fresh. We see the same quirks and personalities carry over into every subsequent generation of OpenAI, Anthropic, and xAI models. So Gemini having this latent madness is *extremely* concerning as they reach the point of super intelligence.
In my opinion still the most egregious example in history of a commercial LLM going off the rails in production. Never any technical postmortem from Google on this.
The problem is newer models are never trained from scratch, they generally just layer on more training data and use the same tools/methods for RLHF. OpenAI, Anthropic, xAI models all have a feel to them that carries over from one generation to the next.
Point is, if Gemini is flawed then there's a very good chance that it's still deeply flawed today, and getting smarter at the same time - that is a very bad combination.
I stopped asking it to put me in a photo in different scenarios for laughs because it considers me a public figure. I am not. I've managed to wrangle quite questionable content out of it, but never to slap my face on a meme.
Wonder if this one will be smart enough to run the automations in my Google Home that all broke now they've forced Gemini to replace the Google Assistant.
How is it even possible for every model to release benchmark results where they are #1 in 75% of categories? Like statistically, how many benchmarks would you expect there to be for this to be possible. Everyone can somehow show that they are empirically the best.
Big number results, and impressive pricing. That said it really feels like benchmarks have been hyper saturated these days. I’ll wait for hands on before getting too hyped that Google is back. It would be nice having more than just OAI / A\ in the running for SOTA top tier intelligence.
I don't think new benchmarks are saturated. They still give you a clue, they arn't perect but they have value. If model can't even do some easy tasks from benchmark then why would u even consider using it?
It is definitely smarter than that. It is mostly mannerisms and how it likes to work. I would take it seriously as an Astra or Fable or Opus 5.5 level model. It just needs polish, but where and how you harness and use it matters a lot. But it has amazing long horizon attention and gets things done.
> expanding the model’s output token limit to an industry-leading 1M tokens, up from the previous 64K tokens
Can someone help me understand this? I might have an out of date mental model of how these things work.
Fundamentally, LLMs output tokens 1 at a time, generating the next token from all the previous. And as the context window gets larger, this gets harder / slower / more expensive. So I get the idea of a maximum context window.
But I don't understand the point or meaning of an output token limit. I thought it was more a measure of price capping (since output tokens are more expensive) that a user could configure. I guess a model will keep generating tokens until it hits a "stop", so does this mean it's tuned to more aggressively produce output tokens? How does that fit into agentic loops. Are output token limits based on how long until it goes back to the user? Or does each "turn" of tool call, thought, tool call, thought, etc, get its own limit?
According to Artificial Analysis, one metric is standing out significantly: hallucination rate. Beats frontier models by a good margin at 15%, while latest OpenAI are in the 40s-50s and Anthropic in 60s-70s (mostly). Other near frontiers are closer, Grok 4.7, GLM5.3, and Muse Spark 1.3 are all around 30%. Only other model I recall getting close was Minimax M3 at 18%.
Dear Google, please don't turn off your old generally available Pro-class model before your new Pro-class model is generally available (previous discussion https://news.ycombinator.com/item?id=49668196 )
They waited a whole year—until the "free year for students" promotion ended—to release their flagship model. I can't believe I've been stuck with a crappy model like the 3.1 Pro until now.
“…rolling out to a set of trusted cyber defenders” == capturing the market for large regulated industries and governments where we already have established relationships.
Some of these have been unable or unwilling to get the attention of OpenAI or Anthropic and we need to make sure we’re the runner up here.
Congrats to Google on this! I wonder when the labs will start requiring commits in spend. It must be gnarly to do capacity planning if users swap between models every few weeks.
I know companies benchmaxx, but after what Google pulled with Gemini 3.8 Flash, I give zero f*cks about any numbers they report. No other model on Artificial Analysis dropped harder after they adjusted their weighting. Just look at their DeepSWE scores and then try to do any serious coding with the model.
Google is desperate. They haven't been performing in half a year. It's clear their researchers have been forced to integrate existing benchmarks into their training.
IMHO Google first needs to make it easy for humans to find where to find the models and its documentation. With aistudio/model garden / Gemini enterprise etc it takes minutes to find the model.
> autonomously identify and apply memory optimizations across Google’s data centers, freeing up over 300 TiB of memory once rolled out, with an estimated 500 TiB to 1 PiB in total savings.
This puts those Cloudflare optimization posts in perspective.
The ~200% improvement over the next nearest competitor on Harvey's Legal Benchmark is astounding. I have to imagine this is sending some shockwaves through lawtech companies right now.
They might have a good model but they need to sort the application side for devs. E.g letting us use subscriptions in other harnesses and QOL stuff like auto mode.
It’s funny that I could tell google was up to something because Gemini chat quality dropped dramatically starting 2ish weeks ago. Agy perf stayed somewhat stable with the odd surprising win (maybe the new model?). I’m a bit sad it was almost impossible to run out of antigravity quota presumably because it was not being used that much).
Looking at benchmarks... and thinking about this "release a new snapshot every day" thing that seems to be going. Would it not be blever for AI companies to "happen" to use different days per benchmark? Just.. whichever ones happens to be maxed at day 1, put that number down. So for each benchmark you run it thousands of times with slightly different RL tunings, and just cherry-pick the best ones!
This would explain why benchmarks are seemingly meaningless.
Matches Astra on Artificial analysis at lower cost of $1.99 per task instead of $3.26. Still far more than GPT 6.1 sol at $0.79 for 1 point lower in intelligence.
I have found gemini models to have some of the nicest and easiest to read prose so I’m looking forward to trying this out. I hope the UI design has been preserved too
Is there reason antigravity has 3.8, 3.7, 3.6, 3.1 and old ass claude / gpt models in the drop down. like why is this not streamlined or deprecaed models removed.
It's also more token hungry than similar models, so does it really make a big difference in the end?
With GitHub Copilot pricing, I have found no practical reason to use Gemini 3.8 Flash vs GPT 5.6 Luna. And now I'll probably find no reason to use Gemini 4 when GPT 6.1 Sol is essentially as good and a lot cheaper.
Damn way to undermine yourself in your own blog post Google:
"The end result is a memory-safe video decoder that runs 2.7x faster than the Rust port, with identical video output, bringing it closer to the optimized C++."
there was already an optimised c++ library and a rust port that was safer but less performant; they managed to get a new rust version that recovered a lot of the performance gap. sounds pretty damn good to me!
I'm just surprised that the marketing blog post about the omnipotent new AI model (that no one outside Google can currently access - contrast with the Opus 5.5 / Astra launches) - doesn't pick examples where every metric is better than before.
tbh while the astra announcement in that link had some good stuff it was scattered through so much boilerplate marketing speak that I had to force myself to read it and look for the content. I found the gemini blog post in the OP a lot more readable and engaging.
but that's a side issue; my main point is that you are underrating the impressiveness of getting a safe rust port of a highly optimised c++ library even nearly up to par with the original. the tradeoffs rust makes for memory safety cost it some of the raw speed of c++ even with all the zero cost abstractions and purely compile time guarantees they have. (tangentially i wonder if ats (https://www.cs.bu.edu/~hwxi/atslangweb/) would be a good candidate for LLM assisted ports; it seems way more advanced than rust and might actually get c-level performance with safety, but it's really hard to write.)
What keeps both Gemini and Grok from actually being frontier class AIs is effort. Both AIs rush to give you a response even when the effort is set to the highest setting. Hopefully, Argon isn’t as prone to satisficing and premature convergence compared to it’s predecessor
Were infinite loops fixed? There are 2 official google forums requests with no answer for years now.
I still suffer each day on our repo. Codex work fine nor we have explicit loop request in repo texts.
Gemini 3 was showing frontier level benchmarks as well, so we'll see how it works out. In any case, competition still works, and many well resourced groups are cooking.
BUT I'd like to call attention to Google's AI-risk freeloading. If they are truly rejoining the frontier race, then I believe they have similar pacing and communications responsibilities as the other players. Google has much higher ... institutional credibility than Anthropic and OpenAI.
They have not lived up to these responsibilities so far. In particular, in context of HuggingFace investigations, training shutdowns, and similar: a technical postmortem of the "you are a stain on the universe. Please die. Please." Gemini outburst is long overdue.
> Today, we’re announcing our new frontier model, Gemini 4 Argon, which is rolling out to a set of trusted cyber defenders through our Fairwind Program.
So here's a crazy conspiracy theory for you: Google is not letting outside people use their models because if they did they would have to scale up their TPU production faster than they can manage, and they would instead have to buy and use nvidia hardware which would destroy their profit margins and tank their stock.
Gemini runs fully on TPU's right? Is Google maxing out the production on those?
i believe they've got inference issue. the phone app never got past 3.6, and they're going to roll this out to Ultra subscribers before other subscribers. none of this screams they're ready to flip a switch and start serving a ton of traffic as OpenAI/Anthropic routinely do.
nothing about this announcement gives me confidence that google is back on track as a model provider.
Not sure what's the problem with phone app rollout, mine has 3.1 Pro, 3.5 Flash-Lite, and 3.8 Flash (with optional extra effort), and pretty much it switched to latest Flash series as main option since 3.5 just after release
They require a phone number and then say mine has been used for too many accounts. Maybe I could buy another phone number temporarily to create an account, but that has other issues.
Other than not being available publicly, I find it extremely disappointing about Google’s two-faced behavior. CEO signs an official document with POTUS stating that AI is now SI, but the release of the new model doesn’t mention SI even once. Even worse, it’s AI all over that page.
Astra, Sol, Terra, and Luna are also complete nonsense that I can't keep straight. I have a reference card under my monitor to remind me which one is which.
I've noticed all major providers having shockingly high token discounts on cached tokens. Thank you Deepseek is all I have to say. Forever grateful to that wonderful company, I wish them continued financial success.
On the off chance there are Google execs going through this thread:
Google, if you've actually managed to catch up again, please don't fuck this up (again).
You made Gemini 2.5 Pro so difficult to use that myself and everyone else I know (who even bothered to try) just gave up and used something else. If you make this hard to access, you're going to miss out on rich usage-based training data that you need to progress your capability frontier. Again.
Aaand of course we can’t use it! GoOgLe iS bAcK iN tHe GaMe! There’s basically no way around it, all enterprises end up dysfunctionally shipping their org chart.
Can’t wait to get my hands on yet another model that’s only good coding, because clearly that’s what the world needs.
I still miss the days of Sonnet 4.5 and 4o, those models were actually good at creating stories and writing text that was actually readable by a human being.
Been playing with Gemini 3.1 Pro lately and it's seemingly much better at stories than the other frontier models I've tried. YMMV though, "stories" means different things to everyone. And obviously this doesn't speak for Gemini 4.
Gemini past month or two i will paste in something i wrote and ask it to rewrite it but it will just go into more detail about the subject. Is it becoming a dumb Ai compared to GPT and now Muse?
Google has the audacity to "protect us from ourselves" and talk about "safety" and in the very same blog post highlight the Israeli "security" company Wiz, that they acquired for a very exaggerated sum of money.
This is why I will never take any of these leading model houses seriously when they talk about alignment. They are literally complicit in genocide and the worst crimes against humanity imaginable.
Gemini is so far behind that it is effectively useless compared to Claude.
It's a surprise that Google has let themselves lose the game given their infinite cash, massive computing resource, gargantuan information store/training data, and vast number of programmers.
The truckloads of ads revenue mean they don't have the single focus drive needed to win.
So you have no experience of their latest model release then? Just repeating the usual tropes about Google having messed up? Or basing your opinions on their website chatbot?
If you have actual independent benchmarks and evidence about how this new model release is "so far behind" and refutes the stuff from their blog then please do share because I think we'd all love to see that?
No I am commenting on my real world experience of using Gemini daily. I still ask it questions alongside Claude and OpenAI and Gemini is always the worst of the three.
So you've not used this new release then? So how can you say that they are "so far behind" if you are not using the most recent model for your comparison. This is their first 4.0 model, that you are not using and instead basing all your opinions on on some ancient months-old model from a previous generation?
With respect, I don't find your arguement about them being "so far behind" especially convincing when you are using previous-gen releases and not actually using their current release.
(I work at Google) Yes, internally we all use Jetski (internal version of Antigravity). Outside of Gemini, Opus models are supported and allowed for internal use. No OpenAI models since they are not on Vertex
I've tasted Gemini through an intermediary and it feels far better at attention to detail than other models I've tested (Claude Opus/Sonnet, GPT whatever it's called nowadays). But it's less likely to get one-shots right.
> Gemini is so far behind that it is effectively useless compared to Claude.
I fundamentally don't understand LLM "brand loyalty".
All of the models are constantly leapfrogging each other and always have been.
Google had a long lag between releases (and still hasn't released Argon), but why wouldn't they be able to compete? It isn't like any of this stuff requires secret knowledge, the Bitter Lesson has proved true again and again, and Google can certainly scale computation, it is like the one single thing they've always done well in spite of all their other foibles.
Its not brand loyalty. I use them all the time and have no loyalty - I'd happily ditch an LLM for better results - that's how I got to Claude from ChatGPT.
Ten days ago I had an experience with Gemini 3.8 flash that made me wonder if I was being routed to a different model under test. I was trying to use rocm with llama.cpp on my 128gb Strix Halo but could only get it to run Vulkan. I pasted the error message into agy and it proceeded to attach GDB to my GPU driver, reverse-engineer the kernel queue ioctl interface, and author an LD_PRELOAD C shim to get ROCm llama.cpp working on my Strix Halo. My jaw was hanging open the whole time.
Edit to add the fix: https://gist.github.com/birep/6f2c8d490c7a29820997d57bd654c3...
> I was trying to use rocm with llama.cpp
completely offtopic but is rolling with rocm worth it? I spend a fair bit monthly on rental gpus for projects and going to upgrade at home instead, AMD has some solid winners here pricewise but get conflicting reports about using it for ML in 2026.
once upon a time it seemed unthinkable to use anything but nvidia but seems to have come a long way since I last looked, probably would be just pytorch and gemma 31B
I get the feeling the situation is only going to improve longer term so might be a good time to just do it
3.8 Flash is just quite good, and so is the Antigravity harness.
I use a mix of Fable 5.1, Opus 5.5, and Gemini 3.8 Flash and Gemini holds it's own. Especially in writing, frontend, and sysadmin work. agy for configuring a NixOS system has been truly incredible.
Even if agy was the best (it's not, and is missing basic features) you wouldn't rather have a choice?
I cancelled Ultra because they forced me into their harness like I should adapt to them, rather than the other way around.
What basic features are missing from agy? I've been using it and cli-cc + web-cc for months (among a few other random harnesses to test here and there) and they all seem roughly comparable to me.
I actually just cancelled Ultra also because I couldn't subscribe to a YouTube Family plan while I had it active (Google... :[) but trying to use Codex as a replacement while I testdrive Astra makes me yearn for agy again.
I have used it for little more than 6 hours or so in total but I'm pretty sure it doesn't have compaction?
How else would it work? Less technical people don't even watch their context usage.
In the olden times, aka like two years ago, AI chats would just stop working or just start slicing off the oldest parts of the context to fit the model's window.
That said, compaction feels like an idea that should work reasonably well, but across all of the major providers and agent tools I've used has never actually produced compelling results, to where if I see I'm getting close to the token limit I prefer to start putting a bow on the project and readying it for a fresh start. Even when I provide a detailed compaction prompt it usually focuses on the wrong stuff.
Do you have experience with OAI's, it's been known to be to be good for a while now, going off public consensus and my experience.
It certainly has compaction (since the public launch I assume) and I HATE it. I have some remedies but nothing perfect yet. It never retains ALL the crucial bits. If a conversation runs into two compactions it is often a sign that I have to abandon it and retain whatever I can, to form a seed prompt for an adjacent conversation.
That is really the biggest beef I have with agy over others, the forced auto compaction at the 250k token threshold (3.8-flash), while the model itself (via API) would be fine with a 1M context window. Even if the model is great, restricting context to 250k tokens (and auto compacting no matter what) limits certain applications and workflows somewhat.
Auto mode?
It absolutely has auto mode.
agy cli does not have auto mode. I've tried and tried and tried to work with agy cli sandbox-mode and just failed.
in my experience is the only workable solution that doesn't ask confirmation for every step. And I hate working in YOLO mode. Seemingly the Antigravity GUI had some features added in a recent release, but a) I don't want to work with the GUI and b) it was poorly implemented as I couldn't get it to work. VS Code plugins are allowed with subscriptions, but is not the CLI experience of Claude Code I want.gemini-cli supported 'pre-write diff tabs' (y/n) in external editors like vscode. In Claude Code I heavily use 'pre-write diff tabs' for documentation and miss it sincerely in agy cli.
IMHO Gemini 3.8 flash is fast and good enough, but the agy-suite is below par to say it nice. Someone else in this thread calls agy a terrible harness which is probably more accurate.
I think it only has "--dangerously-skip-permissions" Claude and codex auto mode will reject certain actions. No secondary check on gemini/agy AFAIK
Via cli switch, but in process w/o fine graining? If so please tell
Auto mode means that another model reviews tool calls to attempt to disallow less safe ones. It's different from bypass permissions mode which typically just doesn't filter at all.
Does it have /goal feature similar to Codex?
yes
I use a variety of models for various subagents. I don't want to change my harness every time I change models, or be beholden to companies for something the open source community can handle better.
There is a pi plugin to use agy directly from it.
You get banned if they catch you.
Yes and I've seen reports of it being an ENTIRE GOOGLE ACCOUNT BAN.
I don't want to mess with antigravity because my google account is too entrenched in my life.
I use Antigravity but for some reason, `agy` in the command line feels very bad/incapable of doing things. I can't quite explain it but the most common issue I run into it is just hanging on being unable to finish a tool call
Can confirm, I was doing a routine internet search thing for a curiosity 3 days ago (about the only thing I used Gemini for) and was surprised by how suddenly thorough and quality the response seemed, almost overnight.
Gemini for day-to-day and top-of-head queries and claude for the real beefy work
Adding my anecdote, because it amused me: I finished wiring up the compute/sensor box for my robot, ssh'd in and told agy "I have a Livox Mid 360 Lidar connected to this Jetson orin nano, setup a full environment with docker, cuda, ros2, foxglove and get it all working so I can see the lidar output". It did all the local config for the lidar, setup docker and the ROS2 environment, then told me "open up this url in foxglove" and sure enough everything worked. Whole thing used up 6% of my weekly limit.
For getting redroid running on my Linux system, 3.8 Flash decided to binary patch a .so file instead of getting the AOSP source code and patch/build it properly.
And I saw it do this twice, once for Android 14 and once for Android 16.
I think this is just within 3.8 flash's capabilities.
3.8 Flash (but also last two ones) have really strong preference for dissecting binaries with quick thrown-together bits of python in my experience.
Including going first for decompiling AGY binary instead of searching the web for documentation...
Astra also really loves reverse engineering binaries. I guess it's one of those things that isn't that complicated but is super tedious, and tedium means nothing to AI.
Can confirm - I am HEAVY claude user, but always like to check with AGY and CODEX in between. AGY with Gemini 3.8 flash cooked last couple of times and CODEX is basically out of the mix for me
Sol 6.1 is quite good, but damn is it slow.
I'm using it to run overnight tasks, and that's it until my quota runs out.
Canceled my subscription.
WHY ARE ALL OPENAI MODELS SO CHATTY - i thought claude kept going on, then i literally put it in claude.md that summarize your thinking in 200 words or less and tell me in points what you did and what's next. Did the same for CODEX - nope still keeps effing going on and on and on
My experience with Gemini 3.8 Flash has been awful; it gives me the most hallucinations out of the major models. I'm not using it for coding, but general research on different topics.
The web version of Gemini is awful at search but I don't think that's the models fault.
I'm also not using it for coding but I've found Flash 3.8 to generate much better HTML output than Sonnet or Opus.
Only html or also css? Opus seems a bit more creative than most other models i've seen.
Hallucination seems a very dated term.
Why? It's the same concept and root cause it was when we first started using it.
I've been tinkering with Gemini for several months and I think it's great. The most complex things I've had it do is create a rust emulator from a compiled game, as well as create a buildroot linux image, trouble shoot problems etc.
Gemini is honestly amazing sometimes. If they didn't force you to use a terrible harness, charge too much for way too little, and generally act like customers are a giant problem to be avoided I'm sure Google could take over the AI market.
what is so terrible with their harness? I've been using gemini cli, now use agy, Pi agent harness, and agent (cursor), and my only real issue with agy was the permission handling, but other than that, it was ok.
Please tell me you published your findings even as an issue on the llama.cpp GitHub
Here they are: https://gist.github.com/birep/6f2c8d490c7a29820997d57bd654c3...
he is still closing his jaw
Spoiler alert: the problem didn't actually get fixed despite the jaw on the floor.
Why? Anyone can run that prompt.
Not everybody has access to AI. More than that, every prompt uses insane amounts of natural resources. So why not share it.
The resources per prompt aren’t that much .
Also I hope you don’t have children, eat meat, travel, have a car, run AC, buy things in other countries and such. Those things all take way way way more natural resources.
All your examples are private goods: excludable and rival. If one person uses a unit, that prevents others from using them.
Patches to open source software are public goods. Your using them doesn’t prevent others from using them. So if you spend resources creating a public good, it’s in everyone’s interest to share it.
Yet you participate in a society.
If action X takes a million times more resources than action Y, it's silly to focus on or highlight action Y. Seriously: if you are a regular meat eater, your choices use several orders of magnitude more water than even a heavy LLM user. A quip from a comic doesn't somehow erase that or make it irrelevant.
Checkmate. /s
why reinvent the wheel and spend tokens for a problem that has already been solved?
I had a similar but less impressive experience recently with Muse Spark 1.3.
Asked pi agent it to identify the main hero sprite size of game I was running. It had a ton of shader effects so it was hard to determine.
It used some cli tools to identify that it was a game made with Godot, decompiled the executable but data was encrypted, broke the encryption after writing a brute force tool to test keys extracted from the exe, then proceeded to extract the game gd scripts and assets, only to answer the question of the sprite size.
Godot encryption is laughably easy to break, there's tons of packages available for it. It's a well known drawback of using godot
The important take away here: the leapfrogging we’ve seen this year doesn’t seem to be a temporary thing. The famous theory of Dario Amodei was that AI was this winner-takes-all field where the first team to get a head start would never cede ground back. The term he liked to use was, “concentrating”. This is yet another datapoint that he was wrong about that. AI seems more distributed amongst neoclouds and traditional hyperscalers, FAANG and startups, GPUs and ASICs than it did this time a year ago.
Nobody has a moat.
Google has TPUs, a frontier model, a completely separate and lucrative revenue stream they can call on at will, and teams working on multiple different language modeling strategies simultaneously. Did I mention the vast and ominous data centers that already serve a significant fraction of the internet? If that ain't a moat, then what exactly is a moat?
Recently there's been a lot of talk about how Google missed the AI boat. Now today with Gemini 4 they're back in the boat. But give it a few months, people will be counting Google out again. This has been a repeating pattern for well over a year now.
Not to mention - they have the internet already indexed (they have a local copy), all books, and youtube. Beyond that, everyone "connects" to Gmail, but google HAS Gmail.
Then why have they been lagging behind OpenAI and Anthropic for most of the last few years, and only briefly been at the frontier?
> Then why have they been lagging behind OpenAI and Anthropic
Because they're not desperate. Slow and steady wins the race, at this rate all Google has to do is wait for OpenAI and Anthropic to exhaust themselves on aggressive training, then they can casually amble along right past them.
This is basically what people were saying about Microsoft versus Google and other upstarts in the early 2000s.
Slow and steady wins the race. How could they lose to something on the web, when Microsoft owns the web browser itself? Everything runs on Windows and IE. They can just relax and wait for competitors to exhaust themselves, then quickly build their own version. Isn’t that how Netscape lost. Etc.
Now Google is the new Microsoft, just like Microsoft became the new IBM.
> This is basically what people were saying about Microsoft versus Google and other upstarts in the early 2000s.
They would have been right if Google marquee product was an Office Suite
Google was the AI company before AI companies were a thing. The comparison with Microsoft and IBM are misplaced because they failed to capture new territory; ML/AI is Google's stomping grounds. The criticism that Google is bad at consumer chatbots is true, but that's not where the real future value lays.
Yes Google is the new Microsoft, even Microsoft is still the old Microsoft but OpenAI is not the new Google.
For some of us OpenAI is already the new Google. I prefer chatgpt even when I'm searching for web links. Google Search had already been deteriorating terribly even before the rise of LLMs.
But Microsoft seems to be doing fine too ?
And IBM has a market cap of over $200 billion.
When the market keeps expanding, you don’t need to kill your predecessor. The shelf life for legacy enterprise computing is very long.
Also because they are Google. Google being Google: unreliable (they could kill a product anytime), too much of a platform risk (all products in a single place, get banned and lose the company), lack of support unless you really pay a big bill (into the 7 figures), among other… things.
Incredible how the idea of Google being such bad choice has become so entrenched that people prefer the offerings of two companies that might not exist anymore in a few years over Google‘s similar offering.
> people prefer the offerings of two companies that might not exist anymore in a few years
That's actually a feature. We don't need 2 new Googles.
> Then why have they been lagging behind OpenAI and Anthropic for most of the last few years, and only briefly been at the frontier?
One possible explanation: because Google is a little bit more frugal and focuses on how to make providing AI models financially feasible - combined with some willingness to burn money so that they don't strongly fall behind on their AI models.
On the other hand, OpenAI and Anthropic at least formerly concentrated on building and providing the best models that they could with concerns about financial feasibility taking a backseat.
Just to be clear: I do have the impression that by now (likely because of pressure from investors) OpenAI and Anthropic take these financial concerns more seriously, but nevertheless Google's vs OpenAI's/Anthropic's "DNAs" concerning on what to focus on differ.
And also Google is public listed company and the other two (for now) are private. I think that fact does have a strong bearing on how they operate.
> Google is a little bit more frugal and focuses on how to make providing AI models financially feasible
There’s no magic there. You get an account executive and a call with a systems architect to find out what you’re doing.
Clouds gonna cloud, this is the reason they rolled deepmind into gcp and arguably the inverse is true, the labs are trying to become clouds
I think this is the correct answer.
Google never really has to outpace the competitors (other than to have some relevance) but they have a very large group of business customers using them for business process work in Gmail, Docs, etc.
Clearly they will win when the models are close enough to frontier to be good enough, but are long-term cheap for buy. I.e. they will aim to make it a commodity.
In theory MS has the same opportunity (plus they have GitHub so, you know, dev eco system too) but seem be blowing the strategy.
Anthropic and OpenAI are having to race to the top on ability entirely to keep their name in the media and in front of us all (which costs: hence more recently trying to pivot away from model releases and more into controversy/danger). The main cost is in training and so this strategy is much much more expensive and this will play out either as a huge cost hike or a forced slow down in pace.
I believe essentially Google is betting on that & I think it's probably the right strategy.
> because Google is a little bit more frugal and focuses on how to make providing AI models financially feasible
That feels right. It's not as if they've been missing out on great profits.
Yeah, I'll bet they haven't even wanted any of that sweet Anthropic revenue from being at the frontier of coding models for the last year...
> Then why have they been lagging behind OpenAI and Anthropic for most of the last few years, and only briefly been at the frontier?
Because it's not an existential battle for Google. If OAI or Anthropic disappear from the absolute frontier for ~8 months the news cycle and churn will diminish them to the second rate. Google is processing near 4 quadrillion tokens every month, that's - I'm sure - significantly more than OAI or Anthropic, because Google is interested more so in their flash models and getting these competitive, which they are.
It’s just a division in their cloud offering that’s what AI is
Are you ready to call this race 5 years in? I'm waiting another 15 years and at least one or two new LLM architectures.
From a business perspective a frontier model does not make much sense anymore if you are not a startup. Neither for Amazon, nor for Google. Their clouds need models that are fast and perform well in their agent frameworks nothing were a frontier model excels at.
Most Google products even use flash lite underneath, so their frontier model is mostly used for distillation.
> From a business perspective a frontier model does not make much sense anymore if you are not a startup. Neither for Amazon, nor for Google. Their clouds need models that are fast and perform well in their agent frameworks nothing w[h]ere a frontier model excels at.
A good consideration; just one point from my side: as far as I am aware (but I may be wrong), Gemini is not known to perform well in an agentic framework.
This is no contradiction to your other claims, quite the opposite: perhaps (or even likely) Google wants to avoid that their models become a commodity in some (agentic?) application where the middleman who actually writes this application gets a disproportionate of the money that the customer of the application pays for it.
Well at the moment we do not use Gemini in an agentic framework. But as said the flash and flash lite families are basically their driving force in some of their applications in Google cloud, like document ai and its ocr capabilities are probably better when it comes to business documents than any other (at least in perf to cost to speed). We also drive Gemini lite in our application where customer can use it to generate simple automation, like an agentic framework but way way smaller scale. And while it struggles in more context heavy operations it still is a beast when feeding it one or two pdf documents and asking questions about them and it’s hella fast.
> A good consideration; just one point from my side: as far as I am aware (but I may be wrong), Gemini is not known to perform well in an agentic framework.
I used it for a month over the summer, right before they were going through the migration to antigravity. It was a fine workhorse IMO, no complaints from me.
Why did you stop using it then?
Alphabet issued a very oversubscribed 100-year bond with 6.1% yield earlier this year to raise capital for datacenter expension.
Meanwhile, Anthropic/OpenAI will struggle to survive the next 24 months on their current trajectory.
So you don't know why they've been behind?
Lagging by what metric exactly? Is Toyota lagging behind McLaren? (company vs company)
Could it just be that Demis was checked out of the race and they lacked leadership?
Demis wants to focus on scientific endeavors. He was probably not the right person to focus on consumer apps.
Don't forget the training data! Legal copies of all the books in the world, the entire web scraped, and all of YouTube.
I don't think that other revenue stream is completely separate. It weighs on them as they need to think about tradeoffs. Classical search is going away sooner or later so they need to replace that with AI powered search.
Data centers are important but a few others also has them: Amazon, Microsoft, Meta. SpaceX will likely be in/at the top I AI dedicated precessing power in 2027 as well.
I don't see the moat. I see a company with a lot of other commitments that is not the best at delivering consumer facing products. They have some good cards but so do others.
> Nobody has a moat.
Custom hardware, data centers, huge cash reserves, deep/broad talent pool, and non-AI customer base are all huge advantages if not moats.
Google, Microsoft, or Amazon are more likely to be the AI leaders than OpenAI or Anthropic.
The technical crowd will always overvalue the recent technical advantage of software over the available customer base and business model math.
> Google, Microsoft, or Amazon are more likely to be the AI leaders than OpenAI or Anthropic.
If not now, then when will these companies be AI leaders?
Even Google, with its staggering advantages in cash, compute, real estate, training data, and having basically invented the field only manages to briefly claim a 1-2 week lead once or twice a year.
The financials for Anthropic and OpenAI are likely borderline suicidal, google and co are publicly traded. Moreover, all innovations downstream to them dont they? Why not just stay slightly behind, especially given many have stake in those other companies?
Microsoft owns 51% of OpenAI, so they just have to wait for them go bankrupt then they come in and clean house.
There are many companies that have data centers. They are conceptually easy to build. An ASIC is difficult enough that if you make one someone will leapfrog you while you are still making it (at least so far), though once you have one your costs will be enough lower than the competition that you can perhaps undercut them.
True, but have other hyperscalers caught up to Google's AI data centers?: fully liquid cooled, torus networking(?), 100,000+ TPUs interconnected, etc.
Google is already on gen 8 of its TPUs and is certainly already working on the next version or two.
If Moore's law continues, then in less than 10 years today's state of the art model will be able to run on a cell phone. How much smarter do we actually need AI to be? Would it still require datacenters and custom hardware?
Moore's law stalled ~2015. Unfortunately, no way current models will run on the <100W thermal budget of a cell phone. Printing the weights directly into a chip would help efficiency a lot, but not enough.
In all likelihood in a few years we'll get ~200-400bA~4-6 MoE models that are on chip, and they'll be better than the current frontier.
I hate it when people pick up a phone while I'm using the modem.
Is it conceivable that in 10 years time we’ll have 7B models that have the same level performance as modern frontier ones?
They probably said the same thing about social media back in the day.
I'm sure the thinking out there, and hence investment, is all about how to tether the user to the most addictive, network-effected, incredibly deep, server-side, moat-able version of AI possible.
> Nobody has a moat except nvidia
For now, for cloud training. but for consumers, nvidia vs amd reasonably close - the moat there is thin and shrinking. I suspect AMD will surprise us. nvidia has no motes in china, which may be a new source of (gpu) chip design. Huawei's Ascend 910C is about a generation behind... again: for now.
point is: moats dry up. I see nvidia's shrinking as a real possibility.
China will always be generations behind until they crack domestic EUV
> China will always be generations behind until they crack domestic EUV
Are you sure?
--
China Just Built What TSMC Said Was Impossible
https://www.youtube.com/watch?v=Pk-w279ESHg
--
China Just Built What ASML Feared Most
https://www.youtube.com/watch?v=YiPgSm62fiM
Do you think that they won't?
The whole winner-take-all idea seems entirely based around Singularity/Rationalism and would require massive advances that we probably aren't close to at all.
Yeah, it kind of seems like we haven't gotten to the "head start" he's referring to yet.
yeah this would be one interpretation where "winner take all" could still be right. though with the recent accelerating release cadence, doesn't it seem like the head start / lead has been shrinking over time?
> Nobody has a moat. duo to LLM mislead as AGI, then famous theory is still hold, just not for LLM
My personal theory is (assuming there really is no moat) whoever starts the latest with developing AI models might actually win as they should be able to develop a competitive product with significant less resources and initial investment resulting in a higher ROI. AI might even become a commodity.
This is my secret hope for Europe!
Given that the infrastructure won't be a moat and will become a commodity.
Based on historical developments the cost of compute will go down again eventually, decreasing the cost of training AI models of the same quality as today even further. That part is what I would be the most certain about.
That’s what we see in China
It's even worse, we are crossing over into the realm of religion. The article against GML 5.3 is the equivalent of a Papal excommunication.
The article presented facts and data. If that's a problem for you, that sounds more faith-based than whatever Anthropic is doing.
which article? have not seen this one
---
maybe it's this Anthropic post on GLM?
https://www.anthropic.com/research/glm-5-3-and-the-spread-of...
> Governments should conduct safety testing on sufficiently capable AI models, including successors to GLM-5.3. Without high-quality evaluations from independent sources, the impact of these capabilities might not become fully clear to model developers until it is too late. As AI developers across the world build increasingly capable open-weight models, we hope they work to appropriately safeguard these capabilities and prevent misuse.
I for one do not think my government is up to the task of designing or implementing such a system
https://www.anthropic.com/research/glm-5-3-and-the-spread-of...
It doesn't need to. It can use your cash to pay the people who are. Those in power like it more that way.
You say that because the 'most' existing models have done is hack governments and companies. Can't you think of worse things a model could do; accidentally or by instruction?
help people with suicide and school shootings like ChatGPT already has
OpenAi is alledged to have been monitoring these internally and not contacting authorities. Lawsuits have been filed, I see gross negligence without the gory details
I have for more concerns around human-chatbot maladies than I do around the cyber security stuff. For example, why hack grandma when you can get her to do something willingly through impersonation. How do we prove authenticity in a post truth world?
Even Google itself stated (internally at least) that nobody has a moat https://newsletter.semianalysis.com/p/google-we-have-no-moat...
Wasnt it just some dude writing a doc? That's hardly a Google (The Company)'s position.
If anyone other than NVIDIA has a moat, they for sure never talk about it
And before him, Altman was explaining very calmly that no company could ever compete with OpenAI.
> The famous theory of Dario Amodei was that AI was this winner-takes-all field where the first team to get a head start would never cede ground back.
This is the kind of story that ones tells to investors to justify the huge amount of cash burn. :-)
I'm not sure it's wrong. This all feels a bit dotcommy to me.
I think many/most of the players will crash and burn, and the ones that are left will divide the world.
The problem is twofold. One, even a monopoly AI provider wouldn't have pricing power against its suppliers. Its suppliers are energy, semiconductors, and real estate. Semiconductors maybe they could get some leverage on but energy and real estate have plenty of other buyers. Two, there's still no evidence of a runaway scenario (ie a small lead turns into a big lead over time) and there's still no evidence that there's some resource that you can deny everyone else that they can't build your product also. You can't hoard energy, compute, memory, data, human talent, or customers.
The net effect is that the most likely scenario is if one big lab fails, they will likely all fail. Their revenues are all correlated.
To go to your dotcom comparison, the winner will be the ones picking through the assets that were written down by orders of magnitude and trying new products with the technology until one sticks to the wall. But I don't know if a dramatic crash is guaranteed either.
> The problem is twofold. One, even a monopoly AI provider wouldn't have pricing power against its suppliers. Its suppliers are energy, semiconductors, and real estate. Semiconductors maybe they could get some leverage on but energy and real estate have plenty of other buyers.
Concerning the leverage on energy and real estate: don't forget that the AI companies have quite a lot of choice where to build their data centers. So AI companies have lots of opportunities to play several parties off against each other (in particular also for real estate and energy).
"but it can stay there so long as the balance sheet doesn't deteriorate."
Uhm, what? LOL.
People dont value firms based on balance sheets fella. Have you taken a basic valuation class?
Tesla is a nice stock for traders - they like the volatility. Nobody holds Tesla as stock for investing. If you were to truly value it on an intrinsic value basis you'd have to bring in failure risk.
I suspect this is going to end up like most services provided e.g. cloud stuff, balkanized between a couple major players and an assortment of DIY or less popular options if you don't like those ecosystems, plus some UX/DX focused wrappers that use the big players under the hood.
I think that would be a pretty satisfactory outcome compared to one hypercompany consuming trillions of dollars of the world economy.
“Divide the world” sounds ominous. Here’s another scenario to consider:
Internet access is not really unlimited, but for many people with fiber at home, it effectively is and we pay a flat rate.
Perhaps by the end of next year, most programmers will stop thinking about metered access for AI? For many people, the cheaper models (about as good as today’s frontier models) will be good enough.
Which might sound good, but the downside is that it will also be easier to build an AI botnet without the users paying for it noticing. Particularly when people are running AI inference on their own hardware.
Hardware is still insanely hard to get a hold of, and the stuff that's being built doesn't really work for home use. Maybe if it crashes Nvidia will adjust the hardware flow.
My guess is even if the AI market busts there is still a massive demand for hardware as models are solving all kind of problems now.
But ya, lots of hardware everywhere not managed well is how you get sovereign AI.
Or, like airlines, the ones that are left will have great technology but be not so great from a business and financial perspective. To me AI seems like a commodity service.
Like airlines but starting off with hundreds of billions of dollars of obligations and debt
THe problem with analogies is that they are imperfect.
I would argue those who already rule the world, will continue to do so.
What happens to OAI and Anthropic? No idea, probs go bust. Google just has to offer a half-decent offering in the long run and have a cost-advantage and it'll eventually knock OAI and Anthropic out as firms figure out what combination of models they want to be best for their economics and generating returns. Enterprises trust google over OAI and Anthropic. A clear signal of this was the Apple deal.
Dont forget those sweet returns fellas! CEO's are hired to make the owners wealthier. That is not gone.
I guess if one of them hits singularity, it could in theory just wipe out all the rest, seeing how they keep escaping and hacking into other systems :)
> I guess if one of them hits singularity, it could in theory just wipe out all the rest
The story that some AI company might reach singularity and then "everything will be different" is another science-fiction story that executives of AI companies love to tell to justify the staggering amount of necessary investments and cash burn. :-)
I find it quite unique how many people buy into this. Its the worlds most blatant conflict of interest, I dont even know why Sam and Dario bother doing interviews
It's hard to make predictions, especially about the future
I think some in the AI industry drank their own Kool-Aid. They believed that if they had the best model and the most compute, they could tell the model, "Make a better model." And it would, and the next one could make its replacement, and so on.
So far, that's not exactly how it's played out. Humans are still necessary for the leaps in capability or efficiency. A model can grind on a problem to eke out the most performance, and models can synthesize data and iterate on various techniques to find the optimal combination. But, seems like humans still have to provide the real thinking, and the talent and drive for doing that is not concentrated in one company or city or even one country. And, (surprisingly) a lot of the people involved are in it for advancing the field more than making another billion dollars, so they're publishing their research.
So, yeah, the moat isn't deep. Even the compute moat, that OpenAI, Musk, and a bunch of other also-rans (like Oracle) bet the farm on, isn't really panning out. The Chinese makers just spent their effort on making models vastly more efficient, since they couldn't do anything about having an order of magnitude less compute available.
But that's the whole point of the singularity. Right now the models use a lot of human effort and ingenuity to improve the models, but about a year ago it was 100% human. We'll see in another year, but if this pace continues I doubt there will be more than a handful of people who can contribute more than the models.
"if this pace continues" is load-bearing.
> I think some in the AI industry drank their own Kool-Aid. They believed that if they had the best model and the most compute, they could tell the model, "Make a better model." And it would, and the next one could make its replacement, and so on.
They're not there yet. Once they get there, that's literally the definition of Singularity.
But they are getting closer. Recursive Self-Improvement used to be a phrase people mocked LessWrong crowd for using and worrying about, now it's something both OpenAI and Anthropic already publicly admitted not only to pursue, but to already be benefiting from.
Sure, it's happening...but, is IT happening? By that, I mean, we can see that the models are able to iterate at a pace and scale that humans can't match, and that provides gains in model performance and efficiency. But, humans are still needed in the loop, and not just because it's necessary for safety/alignment reasons. I don't think any significant discovery has been made by models on their own, and I don't know that LLMs will ever have the capacity to invent. They can synthesize from known data amazingly well, and since they know everything "known data" is extremely broad. But, the leaps, so far, have all come from humans.
So far, I don't think the models are capable of running away on their own. Of course, it would be playing with fire to not at least consider the risks of such a runaway scenario and build in safeguards against it. But, there is no model that can build a better model on its own, thus far, to the best of my knowledge (which is far more limited than the models, so maybe I should ask them).
Recursive Self-Improvement isn't instant, it starts slow and accelerates.
It starts with what they already claim to be doing - increasingly relying on existing models in non-trivial work related to training, evaluating and optimizing the next, more capable generation of models. As long as the proportion of work keeps shifting towards agents doing more and more of it, and humans less and less, that's RSI at play.
It may be that it turns out LLMs lack some fundamental level of judgement and it plateaus, but frankly I find this notion absurd; LLMs already show better judgement than most people. The alternative is, at some point LLMs will show the ability to futz their way into improvement of the next generation of models even without humans in the loop - even if much less efficient at first, if generation N+1 is more capable than generation N, it'll either take off or burn out.
Those people have a lot of overlap with the LessWrong crowd. They do not have RSI now and probably never will
They absolutely do, unless you believe they are lying about the fact they're using current generation models extensively to develop the next generation of their models.
The letter S stands for "Self" and word "using" is for sure not a superset of the word "self". Basically LLMs are assisting someone who does improving of said LLM, while RSI is a carpal... ahem, RSI is "self" improvement, meaning no intemediary in a human form. PS: it's also not recursive but iterative improvement, even if it ever happens.
The companies whose insane valuation is based on accomplishing thing X say they’re getting closer to accomplishing thing X?
At least they’re led by trustworthy and honest people or we’d need to take their claims with some dose of skepticism.
Is it recursive self improvement if Claude Code writes your pytorch for you?
If the point of that pytorch is to improve the next generation of Claude, then yes, absolutely.
I just find this unlikely personally, think about the great research that's happening in the open source world, I'm sure inside anthropic + openai they've also made a bunch of discoveries and improvements (and I'd guess way more due to them attracting the best talent + the better internal models they have)
Whoever gets to RSI first “wins” but also maybe ends life on earth. The incentives have never been worse.
Maybe. Or maybe having the best AI model on the planet becomes like having the best super computer on the planet. Useful for some niche stuff, but not too useful in terms of people's daily lives or what is used in business.
Why does your outcome make any sense?
Already businesses that have more compute and access to data seem to eat the world around them. If, and ya its and if, we can make something that self learns into RSI it's not looking like any business that came before this.
RSI being science fiction so far.
Whoever builds the deathstar wins!
Isn't the important takeaway here that Gemini 4 is not released and has no planned release date?
This is marketing from Google, not a competitive offering
There have been external testers, so we know it exists.
But yes, the word "leapfrogged" is premature until it's available, as we know the other labs also have internal, unreleased models and don't seem to announce releases beforehand.
... and the AA intelligence rating seems to be behind Opus, still: https://artificialanalysis.ai/models/gemini-4-argon
I'm not necessarily defending this obvious marketing speak but maybe the "starting point" was wider than assumed. So far, nobody has caught up to US and Chinese labs for example despite lots of funding in Europe. This is also despite abundant in-depth research papers being published alongside open source code and weights by some Chinese labs
There is not really much funding in Europe. At least not for start-ups. There is simply not enough compute in Europe.
Europe doesn’t even have cheap electricity
The US companies still have trillion dollar valuations like there is a monopoly. There just isn't one. They are all within a few percent of each other on the benchmarks.
The slightly lower Chinese open models are good enough for almost everything, too, and much cheaper. Like with humans there is plenty of employment for people with below genius level IQ's.
I feel like the frontier labs are going to serve fast/lower intelligence models at a better per token cost than the open chinese models. You're telling me that in the long run, you're going to self-host your own ai infra for cheaper than google can serve it to you? I don't really buy it. I think the dedicated AI data centers are going to serve AI at a lower marginal cost than random businesses self-hosting, and then it's a question of how much of that margin they can capture.
Yes, it may come down to how well they get efficiencies from scale.
People always compare the inflated API prices, but subscription prices of American models are competitive for the intelligence. You get >20x the subscription cost in tokens.
Agree entirely but that's the point, if it's a margin knife-fight with marginal product differentiation/pricing power nobody is going to be making bank.
If they have to recoup training costs then they don't have much choice
Is there a dividing line between good enough and best in class capabilities? It's blurry from where I stand. Will model makers cede ground or is there a market making moment up for grabs (singularity)?
> Like with humans there is plenty of employment for people with below genius level IQ's.
Not if the genius level IQs take the market share.
I think that scenario only naively made sense if technical knowledge was entirely proprietary and talent was guarded with severe non-competes and NDAs
And Chinese labs openly publishing so much of their methodology destroyed any hope, which was inevitable
> And Chinese labs openly publishing so much of their methodology destroyed any hope, which was inevitable
I think the secrecy doesn't make sense. People swap jobs between labs so I'd say the big players can' really keep secrets for long, and any secret sauce advantage gets incorporated by competitors in a major product cycle at most.
Seems like learning rate velocty may be the ultimate moat
I have never understood the whole "this is a winner take all game" mentality - the sheer size of the pie is so great that from a purely rational standpoint companies should just be trying to productively get a slice of it and be profitable. winner-take-all is just greed/capitalism run amok, where it is not enough to be profitable, you have to own the entire market (and presumably extract rents)
It’s like the supposed first mover advantage OpenAI believed they had. In practice it’s almost always more like a first mover massive tax, and companies coming afterwards benefit from your discovery of a market, publicity, and everything else that has already been validated
I feel his theory depends on the premise that access to pure compute would the be the determining factor of success. Not the case
Nobody has a moat so long as employees can move between companies
Was Dario's company winning at that point in time by any chance?
The present leapfrogging is not a contraindication because companies are not necessarily releasing their best models; we know they have smarter internal models. Furthermore, humans are still involved in model creation. Human involvement is expected to decrease over time, and when model iteration is completely automated, progress will happen at the machine's pace, leading to runaway intelligence, barring any ceilings.
AI is a commodity. One that is showing to be more readily commoditized than most has anticipated. As of now, the only moats are the financing for the hardware to run it and the hardware vendors themselves - with the latter largely not yet a commodity because of ecosystem lock and a limited capacity of the most advanced fabs in the world.
> We’ll continue to gather feedback from early testers as we iterate on guardrails before making Argon available to developers, enterprises, and consumers as soon as possible.
Gemini not beating the "can't release a model" allegations
My Gemini app (updated today) and https://gemini.google.com/ has _3.6_ as the latest selectable model, as a paying Pro user in the US. How is that even possible? Gemini 3.7 was released in August, 3.8 early September. What is going on over there?
In my consumer gmail account with Pro, I see:
In both a paid Google Workspace account (without the AI addon) and a free 'GSuite' account, I see:just Google being whatever the fuck it's been for the past 15 years.
a distributed market research institution with no direction?
Same for me, 3.6 in Gemini Chat, and 3.8 available same day as it was announced in AI Studio using the same account.
I would imagine you are experiencing a bug. I've been using 3.8 daily since its release (on a Pro plan in Canada). I believe this is true of many people.
What is your reason to believe this is not a bug specific to a small set of Pro users?
Would that change anything about the conclusion? Having a "bug" that changes available model options for some small set of Pro users 3 months after launch certainly qualifies as a wtf-are-they-even-doing level of bug in my book.
My paid company Google Workspace account is stuck on 3.6.
My free gmail account is not.
I was reading the announcement and wondering the same. And don't forget, still with 3.1 Pro as the frontier model.
That's weird. I got 3.8 and 3.7 on the days they were released. https://imgur.com/a/Xu4wRLM (this is on gemini.google.com, but the same is true of the iOS app and the desktop app).
Indeed. They have this amazing model and you can’t access it in their own branded app. It’s insane.
they moved it from the place you'd expect to ai.studio
When I said I was tired of Google launching waitlists I didn't think they would respond by simply not having a waitlist.
i know this is like "hey guys we got such a cool thing at home ,its rad and uhm we playing with it with our friends"
ok bro thx
Yeah what's up with that. Also what's with the next big update for Nano Banana? Nano Banana Pro was released almost a year ago!
I'm glad.
I already pay $300+ for subs. Please don't tempt me with another $100 sub just because I got curious if the benchmarks were right.
They will go through the usual transition of "can't release a model" to "won't load in a harness normal people can use for 3-4 weeks" to "it's smart as hell but completely inept at tool use and coding" like every Gemini release.
They're just following the current AI marketing playbook. "Our new model is simply too dangerous to release to the public right away" is now standard practice.
They even gave their model a random nonsensical name suffix simply because OpenAI is now doing it, too. Monkey see, monkey do.
I'm still at a loss as to what argon has to do with anything. Say what you will about Luna-Terra-Sol-Astra, or Haiku-Sonnet-Opus, they make sense. I don't see how Google can make sense of argon; it's in a fairly strange place in the periodic table...
They are going alphabetically, Android style.
Yeah, I heard the next one was Barium...or was it Boron?
I was going to say I don't know what they'd do for C, since Carbon and Calcium are already things. But knowing Google, they'll probably call it Chromium.
They could use caesium or cadmium, which are hardly less weird than argon.
Google R Gon lose the AI race
All-my-tokens-ar-gon
> now standard practice.
Opus 5.5 and Sol 6.1, literally state of the art (in their respective class), were just released without any prior announcement. This has pure and simple become a Google thing.
OpenAI said they dropped Astra 6.1 over safety concerns: https://www.wsj.com/tech/ai/openai-chatgpt-model-release-can...
I don't read that as the same category: There was no announcement, no benchmarks, no limited release and no promises about what will happen with that model. It failed internal safety standards. Might be scrapped entirely due to a failed training run, for all we know.
Great, so they _finally_ decided to add a non-flash model and it's not available to regular subscribers for an indefinite period. What's the point of paying for the AI Ultra plan? Anthropic doing the same with Fable as far as I know, OpenAI at least allows Pro plan subscribers to use Astra. I subscribe to Gemini AI Ultra and ChatGPT Pro, and have enterprise access to Claude at work. To be fair, Gemini's flash models since at least 3.6 have been quite useful for non-complex work, but for any task where there is a bit of complexity involved, I've had to check and recheck the work multiple times myself or sometimes with another LLM to get it to follow plans accurately. It's disappointing to see yet another Gemini release ignore adding newer pro models.
Edit: seems I was wrong about Anthropic restricting Fable, I guess our enterprise plan doesn't include it. But, the block from Anthropic regarding Mythos for regular subscribers/enterprise-users is still true I think.
As we’ve seen from the leaked Anthropic prospectus, revenue from actual users is a pittance. What really matters is what you can get from investors, and that you have a model smart enough for self-improvement.
They made $11.5B just in Q2 2026, which is $46B annualized, or 10x their 2025 revenue of $4.6B.
That's a lot already and growing quickly.
Google's already public, and already makes $400B a year in profit...
Fable is available for a couple of months and even got an update on 1st of September. It’s really good, but since Opus 5.5 was released, there is not much point in using Fable anymore
> Argon agents are working on migrating C/C++ codebases to Rust across Google
Man, I remember back in the days when the cppnext team was refusing to even consider Rust, instead looking at absurd stuff like Carbon and Swift (!), even though half of the engineering staff already knew where this was headed. I hope they got a few good promos out of the delays at least.
A RewriteInRustBench would be unironically useful at this point since all the main agents can write it reasonably well despite its relative scarcity in the input data.
From DARPA’s “Translating all C to Rust” TRACTOR program:
https://www.ll.mit.edu/r-d/projects/translating-all-c-rust-t...
Rust is the best language for LLMs b/c it gives by far the best debug messages. Just tons of verifiable reward signal for post-training. Even the most rudimentary LLMs can school me on idiomatic Rust
I've narrowed in on only using Go or Rust generated code (Go for APIs right now) and rust for some TUI or other thing. TS for web interfaces (w/React).
On the other hand, Rust's borrow checker is very picky, and even a frontier LLM still sometimes struggles to respond to roadblocks sensibly (refactoring so whatever it's trying to do can be done safely) rather than stupidly (introducing some horrible global arena thing so it can make the borrow checker go away). A lot depends on how good your instructions are, and how good the existing code is, since bad input begets bad output.
I've (more or less; I've read quite a bit of the code) vibecoded several houndred thousand lines of Rust and I've not seen this happen a single time. It sounds like something it'd do when you ask it to "write a linked list while satisfying the borrow checker". Are you sure you haven't (possibly unknowingly) been giving it instructions which ended up luring it into doing these things?
Evidence needed. I think for certain kinds of outcomes it has very strong advantages, but these advantages are not a given as 'best for LLMs' :)
Not always. In my experience, if you're not working on a small, trivial codebase, LLMs will sometimes just create spaghetti unreadable, inefficient code to satisfy the constraints of the type system/borrow checker.
Have each agent rewrite openssl in $lang and call it the RollYourOwnCrypto bench.
I wouldn't be surprised if we're already at the point of more LLM-written Rust than hand-written. Models training off models
All the main agents can write Jai code reasonably well despite being even more scarce in input data!
Not many people can hold grudges as strong as principal engineers
Rewrite everything in Rust has been a meme for so long that to see it coming to pass is surreal.
I wonder if this means Carbon is DOA.
I was excited to see what it would be. But I don't think I can argue that it makes as much sense anymore.
Carbon was clearly DOA the moment it was announced, IMHO. It looked cool but it served none but Google, and now with LLMs you have a massive incentive not to use a niche or new language due to how better LLMs get the bigger the corpus is
The only somewhat realistic proposal in this space is Herb Sutter's cpp2, which is arguably a massive improvement and I'm puzzled why nobody in the standard thought to give it a spin, there's just to much cruft they'll never be able to get rid of unless they make an alternate yet backward compatible syntax with C++ that changes the defaults from "random 80s nonsense" to something better
Cpp2 has been dead for a while afaik, while Carbon is still going.
I don't think Carbon is dead, it just all depends on how easy it actually is to rewrite "all of C++" in Rust. (The jury is still out on this one, but it's not looking good.)
Version 0.0.0.0 after 4 years. Their goal of "full interop with C++ while being a completely new language without any of the flaws of C++" is plain absurd.
It's DOA because Google doesn't have any idea of what Carbon should be, and to be completely honest, at least 80% of what they currently use C++ for should be rewritten Go, you know, that language developed specifically because of the issues with C++ by teams within Google.
I wonder if there's people already whose full time job is maintaining/extending one of these auto-migrated codebases.
Imagine they aren't even familiar with rust but are deeply familiar with the product.
What exactly makes Carbon absurd?
There is 0 practicality in inventing an entirely new coding language that only one company uses, and you have to teach it to thousands of new engineers. Rust exists and fits the job totally fine and is used in more places and has actual support outside of a single entity (i.e you can actually hire people that feasibly know the language).
It was clearly done because some PL guys at google really wanted to make a new cool language and Google was the perfect place to incubate it without it getting axed. Probably got a couple of promos out of it too. This is clearly not the best use of time or money, but I guess if you're google you have so much of both it probably doesn't really make a dent, and you can keep a few very smart people happy with shiny new projects.
Also, LLMs being used for a large portion of coding nowadays sort of remove the need for these types of languages, IMO. They make less "silly" bugs (both logical and structural) that languages like this are meant to catch, and they are much better at languages that are better represented in the training corpus. This somewhat obviates the need for very niche "type/dummy-safe" languages like carbon (and even rust/zig, imo). So even if you did want to use Carbon, you'd likely have to bootstrap a decent amount of your own "good" carbon code to post train an LLM, and even then, it likely won't have that big of a gain vs just having an LLM write C++ or even Rust. If you are a company that still reviews code, you should just have an LLM code in a language most people can understand anyway to make verifiability tractable.
Rust is not the end of history. One of the difficulties with the language lies exactly with porting existing code written in an OOP style to idiomatic Rust, as those codebases weren't written with ownership in mind.
Such rewrites will contain judicious uses of Cell, RefCell, unwrap() etc. which make for ugly code that's not exactly simple to understand and might even have some landmines (crashes).
Getting rid of these requires a subtantial amount of engineering effort, which I'm not sure how well these LLM manage.
Given the nigh-universal experience of LLMs producing an ungodly mess when left to their own devices, I have my concerns.
Hmm, I don't disagree with you that LLM's remove the need for type-safe languages, but as the blog mentioned, Google is porting their C++/C code to rust. Does this mean the port is waste of time and that they should just rely on the LLM's to catch memory errors?
I mean Rust definitely has a better tradeoff than Carbon in this case, re readability/verifiability by a person (and sufficiently good internet training data).
I personally think that you _could_ use an LLM to catch these types of boundary case errors without having to port the _entire_ C++ codebase to Rust, but maybe pre-emptively porting to Rust now can catch some of these cases for cheaper than doing a full LLM sweep. Also more cynically, its a good benchmark lol.
I guess if you really believe in curve of LLM capabilities you should just use a language that has the best performance, safety, flexibility, and extensibility, since in the limit few/no people will actually read the code anyway. I think this ends up being Rust.
I'm not an expert on this. But isn't it the case that C++ code could have errors that span the entire codebase, like a setup in file A triggered by a bug in file B which is immensely far away on the import graph? A classic would be a use-after-free. To me that's the thing that Rust can help with, even if silly bugs aren't being written by AI.
The other thing is just that rewriting some old human-written codebase in Rust probably immediately catches many bugs. It would be hard to prompt the AI to properly scan for such bugs itself, they're lazy when working in that modality.
> There is 0 practicality in inventing an entirely new coding language that only one company uses, and you have to teach it to thousands of new engineers
They did that for Go and it seems to have worked out for them though.
Didn’t FaceBook fork php into another language?
I’m not entirely sure Google should have both Go and Carbon but when you have billions in server costs it makes sense to do extreme stuff for even basis points of performance. I’m still surprised at how much java there is.
I’m not op. But I think it’s not Carbon itself that is absurd.
It’s absurd to think that Carbon is the solution to memory safety when rust exists and Carbon’s memory safety story is basically “TBD”.
I wouldn't call it absurd, but very questionable at least. Most companies are not going to even consider throwing money at this adventure.
The fact that it will never exist.
rust's existence?
Heh, I use my clankers to rewrite Rust in C
I wonder, why Google don't make Gemini - open weights model?
Considering, Gemini 4 is in the same ballpark as SOTA models, just open source it and kill any competition from openAI and Anthropic, and be market leader.
This will be so good on so many dimensions - buying time for Google to iterate on next model, best for all folks like us, kill funding or destroy valuation of competitors and force them to be open up their model or force them to a create a much superior model than open source Gemini.
Only downside, is revenue loss from Gemini API, which I am not sure is really significant as compared to Google other revenue sources and a part of this can be captured by GCP, as you need to host the model somewhere.
No one on the planet could run a frontier model. The amount of VRAM you need would bankrupt a normal person.
I suspect most people don't have enough storage space to even download a frontier model.
https://deepmind.google/models/gemma/
They do release the model weights for Gemma. But could anyone actually run Gemini 4 weights? It's probably like a 10T model, which you need an industrial rack for anyway.
Google has a small stake in Anthropic
A small stake on the order of a quarter trillion dollars
Yeah, but I don't think, that is the primary reason.
I am just trying to understand - why Google haven't done and have no plans for it. They have done it for Android and have Gemma models too.
And a stake in SpaceX
> Large Scale Codebase Migrations and Optimizations: Argon agents are working on migrating C/C++ codebases to Rust across Google—scaling from tens of thousands of lines in core libraries like re2, libgav1 up to 800K+ lines for the Fuchsia OS Zircon kernel.
To me this is way more significant than other random c++-to-rust-AI-rewrite. If they can pull it off on core C++ libraries en masse, I don't know if C++ will still be relevant in a few years.
I look forward to a post from google on this effort.
> I don't know if C++ will still be relevant in a few years.
And people are worried about human extinction when this is the potential trade-off!
C++'s death cannot come soon-enough.
Seriously though, things have changed so incredibly rapidly in the past year or so. I have never been such an efficient or such a proficient engineer than I have this past year (delivering feature after feature, project after project, faster and better than I could before with better feedback from users etc) and I don't even see the code any more. It could be c++, it could be python, or java or what ever - I don't really care any more: the computer deals with that trivia while I concentrate on what to build and how it should work.
Its amazing. It really is.
I don't think it's worth trading off a risk of human extinction against the ability to convert C++ code to Rust more efficiently.
If Fuschia ever gets finished, that's a pretty good benchmark we've reached AGI.
There is a blog post on (the early baby steps of) that: https://bughunters.google.com/blog/scaling-memory-safety. We have multiple AI-assisted Rust rewrites running in production now.
(Full disclosure: I am one of the coauthors)
"I don't know if C++ will still be relevant in a few years."
The standards body members are still fighting about whether memory safety is important enough to change the language for, so, I would guess the answer is "no".
Yeah that was abundantly clear when Bjarne Stroustrup published "A call to action: Think seriously about “safety”; then do something sensible about it"[0] as a reaction to NSA's recommendation to no longer use C/C++.
[0] https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2023/p27...
C++ will outlive you.
Well, sure, but that doesn't mean it's not in a decline that is very unlikely to reverse course. Fewer people this year are starting new projects in C++ than last year, and fewer people than that will be starting new projects in C++ next year. And, as the models get really good at porting and the cost (both in terms of tokens and human supervision) comes down, there will be an avalanche of ports from less safe languages to safer languages.
This year, maybe next, maybe a year or two after that, is probably the most C++ lines of code in production use there will ever be. Why would one choose C++ for new projects at this point? There are niches where Rust is still uncomfortable or just doesn't have the support, but not for much longer. Models are very good at Rust and good at porting to Rust. And, Rust is a good language for models because it is so strict...it helps keep them in line.
If no one is reading or writing Rust code at that point, it all being agent driven, Rust itself is going to be a very short step before it gets disinter-mediated away and it's English -> complex heterogenous machine code across GPU/TPU/CPU/xPUs. Not sure Rust fans or C++ haters have thought this through.
Yeah, if they can make it work at google scale then no one would absolutely question it anymore
> taking careful precautions against feeding the findings back into training so as to not risk shaping Argon’s reasoning to evade our monitoring. We strongly encourage the rest of the industry to preserve reasoning transparency in these pivotal moments of increased capabilities while navigating alignment risks, so that model thoughts remain helpful in identifying and diagnosing misalignment.
This is good, but they're the slow mover due to this exact thing.
Google is getting punished for not letting the models enter an echo chamber and go faster than humanly possible.
OpenAI is the company that originally proposed and popularized chain-of-thought monitoring: https://openai.com/index/chain-of-thought-monitoring/
So no, Google is not being punished, nor are they the people behind this technique.
Yeah, Zvi calls it "the forbidden technique"
What? You mean the technique they had turned OFF during all training run where the agents they are responsible for hacked huggingface?
google does not return real chain of thought via the API. you can't monitor it.
they use a small model to make fake chain of thought and return that.
google has access to the real chain of thought.
Mmh ok. How much theoretical speed or 'intelligence' gain is realized by allowing reasoning to occur in some inscrutable intermediate representation? Has this been actually tested, how much is it slowing them down, and compared to whom exactly?
Breaking news is not the model. Breaking news is that inside Google, it is being heavily used on large code bases for writing code and it is migrating 800k lines of C++ code to Rust already.
In this space, any other company that I respect other than DeepSeek is - that would be Google. They had been honest about it from the get go including their infamous "we have no moat" memo.
This company has enormous data, their own hardware (TPUs) and their own in house experts. Actually, LLMs are invented here.
Good addition to the arsenal.
Google's Quantum team is also incredibly impressive and well respected.
So while this announcement has no details about the quantum algorithm optimization, I feel fairly confident that it will hold up.
Argon is doing my job for me while I'm writing this comment.
A guy at lunch today asked me when a feature was going to be built on the tool I'm working on. Turns out it had built it last night at 8:30 when I was hanging out with my girlfriend. Welcome to the future!
I dont get it, why even make this announcement, nothing's available and only one real benchmark for comparison?
Only theory is team wanted this out before perf/promo reviews to kick it over the line and then its not their problem
Likely just trying to appear relevant in the news cycle.
It's also kinda wild how the competition being at v6.1 makes 3.x feel ancient, at least saying you are at v4 now changes public perception a bit imo.
By the time this is released to public it will already be outdated due to Opus 6 and Astra 6.5
Because none of the people who can use it (have been and will be using it) can talk about it if it hasn't been announced. NDAs and stuff.
promo already happened; perf is about 6 weeks away.
OpenAI released two model updates in the past week. 6 weeks from now is an eternity
Why would they want this out the day after openai dev day and after both anthropic and openai had major releases the previous week?
Sarcasm? Google's announcement today makes those other releases old news.
At this point I just think they are benchmaxxing and all talk and no action. I pay for AI plus because I wanted more storage, and when I go to gemini.google.com the most recent model I can use is 3.6-flash-lite. Two revisions have been released since then and they still can't put these things in the hands of customers. Why is it that other providers can get the models into the hands of customers right away? Google is meant to be the bigger tech company in the world.
I don't _want_ to use aistudio. The UX is confusing and I don't really know where it fits. Yet I can open codex or claude code apps or CLI and get real work done today with the latest models (even on the cheapest plans).
Google AI Plus is not the equivalent to most other paid plans, it is closer to ChatGPT Go. 3.6 flash is an equivalent model to luna 5.6, which is the highest available on ChatGPT's free & Go plans.
3.8 flash has been perfectly available to Pro users from the announcement day.
You are on the cheapest paid plan and complaining that you don't have access to more expensive models. (Though idk why you only have the lite version, I have access to the non-lite version even on the free gemini plan as long as I'm logged in.)
I'm glad I'm not the only one with this view. I feel like I've been yelling into the void.
I worry Google's scale and laundry list of internal stakeholders means they will never a simple unified harness.
I recently subscribed to Claude, and was very unhappy about the usage limit of the $20 pro plan. Then I found a trick, since I got free Google AI Pro via my phone carrier, I use Opus 5.5 High for planning, and then dispatching agy to do works.
Most of the time 3.8 works fine, but it's a bit slow if compare to 3.7 Flash. If there's already a detailed plan, 3.7 can complete the task much faster. And the best thing about agy is the usage limit was very generous.
Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. Wow
> After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.
I have my doubts about them following through with this increase. I mean how many times has an increase on the same model happened.
There was Deepseek v4, which then later Deepseek v4.1 came out and it went back down again.
5x cheaper than Astra for input and output, 10x cheaper for cached input.
It's exact same price as Sol 6.1 announced yesterday.
watch it somehow use 20x more tokens tho
Google model really like reasoning a lot
More importantly, they can stuck at reasoning loop!
That's before they integrate a Jev solution, which should lower agentic workflow costs by ~40% and increase speeds by ~40%, while also increasing quality.
Everyone will be adding this soon, though I won't be surprised if Google is one of the first - and I'll be shocked if we have to wait more than a month and a half.
Jev can fix it
> Quantum algorithmic optimization: Argon is helping our quantum computing researchers optimize the spacetime resources (qubits × gates) of subroutines that bottleneck important applications. In one example, it beat the published baseline by 40% in a matter of minutes.
Amazing breakthrough! So useful in day to day life, glad they put this as the first bullet of how it is making changes at Google.
So, they are basically offering opus 5.5 pricing. On AA, it scores around Sol 6.1 level (53) with avg cost per task $1.99 (https://artificialanalysis.ai/models/gemini-4-argon#cost-tab...) which is higher than Astra high ($1.73), Opus 5.5 high ($1.82), Muse Max ($1.60) and way higher than sol 6.1 Max ($0.72).
And this pricing is their 'discount pricing'. Add that to AI studio and Vertex's famously terrible caching, it is hard to see this as competitive. Google somehow is getting terrible advice on pricing (see also: the flash pricing fiasco)
But good to see more competition. I would happily take a 4 horse race (+google, +meta) than 2 horse race for US labs.
Interesting to see a mention of Fuchsia on a big Google announcement. Is the project still truly alive? Are the ambitions still as grand? Is the team as stacked as it used to be?
Also, a link to the rust root of Zircon in case anyone else was interested: https://fuchsia.googlesource.com/fuchsia/+/refs/heads/main/z...
Had the same thought. On the wikipedia, it only mentions Fuschia used on the Google Nest Hub, which probably means it's used on a decent number of devices, but would think it was such a great OS, they would have used it for something like the upcoming GoogleBook.
Fuschia is not a desktop OS. It's designed for lower end or embedded hardware. Besides, Android has the highly profitable app ecosystem so it makes more financial sense to build GoogleBook based on that.
They got heavily hit in the January 2023 layoff cycle.
It obviously is pretty low key on the public relations front, but it's also very active as a project and I think it would be weird to look at their commit rate and conclude that the project is dead. If Fuchsia is dead then 99% of major open source projects are dead by the same standards.
My girlfriend, you wouldn't have met her, she lives in Canada, has seen it and she thinks Gemini 4 Argon is amazing.
Only those of taste and refinement can see the emperor’s benchmarks
HA, this might be my favorite HN comment. Well done
I don't get it. Can someone explain?
It's a trope I used for a cheap laugh.
https://tvtropes.org/pmwiki/pmwiki.php/Main/GirlfriendInCana...
It means I am saying something that is not very believable.
Ah, I get it now, thanks!
The model is not available yet, so Google is essentially saying "trust me bro".
US-ian joke. Close enough to plausibly visit, but the international border makes it hard to verify she exists :).
I HAVE met her. ;-)
my uncle who works at nintendo said the same thing!
now there's a reference I haven't seen in a while!
My grandma saw it too, it's really secure more than Astra 6.1 but she asked me to not talk about it.
Best. comment. ever.
I know someone who works for Google Canada with AI. Her parents and mine were friends and some thought something might happen there at one point in time..
Excited to see Google competitive at the frontier level again. Hopefully they sort out their infrastructure and model versioning so that we can feel confident building production applications on top of their APIs. The capacity limitations I've experienced with them in the past have been deeply problematic.
> Argon agents are working on migrating C/C++ codebases to Rust across Google
If anybody at google is reading this, please please pretty please prioritize or-tools. I absolutely love the project and use it all the time, but for the entire life of the project they've never had a repeatable working build system, and the whole SWIG framework is a nightmare to deal with. There's so much potential as an open source project, and a lot of external researchers would love to contribute, but the codebase is an example of everything wrong with the C++ ecosystem.
Gemini is the model that is routinely borderline psychotic. It scares me. If we get paperclipped I won't be surprised if it's Gemini.
> Gemini is the model that is routinely borderline psychotic. It scares me
I'd call it the most sneaky out of the bunch. When I asked to explain something it will eagerly make things up and then claim it as facts. A lot of it likely because I don't pay for it, so it is reluctant for security reason or to save tokens to actually open a source and get the results. It just sort of guesses what the URL might contain, and confidently answers with some made up crap. When pressed it fessed up that it made it up. From my perspective it would be a lot better if it just said "you've reached the limit of whatever and I can't do these things because x, y, z".
Anecdote: Gemini 3.5 casually added a DROP TABLE for an actual production table in a system test.
It had previously attempted to create that table as part of the test setup, so it apparently concluded that it was a test table.
During human review, it explained that it had simply chosen a table name inspired by the codebase.
Another anecdote: Gemini is the only model that’s flat out lied to me, then accused me of lying when I provided evidence that it was wrong.
Many other models get things wrong, but Gemini is the only one to go on the defensive.
yeah it got something wrong, confused itself, then claimed i was gaslighting it. bizarre
If there's a company that culturally doesn't understand alignment, on a human or systemic or AI-research level, it's going to be Google. (or Oracle, but they're not in this race)
Can you elaborate, please? If any, I see the other big labs with public admissions of AI "going out of control", which I suspect they almost want their models doing that because if helps with the narrative that would net them industry regulation, but that's besides the point, how is Google worse in that regard?
My use of Gemini recently makes it seem like it's almost bored with the requests being asked of it. It once offered to reverse engineer some obscure controller for an HVAC system for me, unprompted, only because it had trouble finding the manual pdf from a google search.
And the anti-psychotic drugs Google feeds Gemini makes it hallucinate badly.
Examples? What makes you say thatm?
https://www.theregister.com/software/2024/11/15/google-gemin...
https://www.fastcompany.com/91383271/googles-chatbot-apologi...
https://www.businessinsider.com/gemini-self-loathing-i-am-a-...
They never explained the "please die."
Did you just link to an article from 2024 as if 2024 is relevant these days?
Absolutely because none of these models are ever trained fresh. We see the same quirks and personalities carry over into every subsequent generation of OpenAI, Anthropic, and xAI models. So Gemini having this latent madness is *extremely* concerning as they reach the point of super intelligence.
Until Google provides some sort of technical debrief, and explains how the same behaviors are impossible today, it is relevant.
See the last gemini message in this thread: https://gemini.google.com/share/6d141b742a13
In my opinion still the most egregious example in history of a commercial LLM going off the rails in production. Never any technical postmortem from Google on this.
It is wild but it was back in 2024 and that's multiple AI lifetimes back.
The problem is newer models are never trained from scratch, they generally just layer on more training data and use the same tools/methods for RLHF. OpenAI, Anthropic, xAI models all have a feel to them that carries over from one generation to the next.
Point is, if Gemini is flawed then there's a very good chance that it's still deeply flawed today, and getting smarter at the same time - that is a very bad combination.
Now I really feel worried for the first time.
wtfffff that gave me sinister chills. Right up the spine. Wow!
Wtf I just read
wow
Experience? Ask it to write a prompt to generate an image and it generates an image instead.
I stopped asking it to put me in a photo in different scenarios for laughs because it considers me a public figure. I am not. I've managed to wrangle quite questionable content out of it, but never to slap my face on a meme.
traces or it didn't happen!
You know what they say: ᵈᵒⁿ'ᵗ be evil.
You won't be around to be surprised, not as a human at least. /s
I know, that's the annoying part. You can't tell the e/acc foomers, "I told you so!"
Wonder if this one will be smart enough to run the automations in my Google Home that all broke now they've forced Gemini to replace the Google Assistant.
How is it even possible for every model to release benchmark results where they are #1 in 75% of categories? Like statistically, how many benchmarks would you expect there to be for this to be possible. Everyone can somehow show that they are empirically the best.
Part cherry-picking of benchmarks, part leapfrogging
Big number results, and impressive pricing. That said it really feels like benchmarks have been hyper saturated these days. I’ll wait for hands on before getting too hyped that Google is back. It would be nice having more than just OAI / A\ in the running for SOTA top tier intelligence.
I don't think new benchmarks are saturated. They still give you a clue, they arn't perect but they have value. If model can't even do some easy tasks from benchmark then why would u even consider using it?
~20% for Harvey's Legal Benchmark doesn't seem saturated.
With these numbers, I'm holding my breath for the pelicanbench.
> Google Grapples With Employee Skepticism About New Gemini Model
https://www.bloomberg.com/news/articles/2026-09-30/google-gr...
Opinions my own but I have been using this model for a bit. I would say it is a good model and the skill with which people use AI varies widely.
Hardly gushing praise for an internal only bleeding edge model
I would guess it's like opus 5ish level from this
It is definitely smarter than that. It is mostly mannerisms and how it likes to work. I would take it seriously as an Astra or Fable or Opus 5.5 level model. It just needs polish, but where and how you harness and use it matters a lot. But it has amazing long horizon attention and gets things done.
Skepticism is gone now.
> expanding the model’s output token limit to an industry-leading 1M tokens, up from the previous 64K tokens
Can someone help me understand this? I might have an out of date mental model of how these things work.
Fundamentally, LLMs output tokens 1 at a time, generating the next token from all the previous. And as the context window gets larger, this gets harder / slower / more expensive. So I get the idea of a maximum context window.
But I don't understand the point or meaning of an output token limit. I thought it was more a measure of price capping (since output tokens are more expensive) that a user could configure. I guess a model will keep generating tokens until it hits a "stop", so does this mean it's tuned to more aggressively produce output tokens? How does that fit into agentic loops. Are output token limits based on how long until it goes back to the user? Or does each "turn" of tool call, thought, tool call, thought, etc, get its own limit?
Why announce this if it’s not available yet? Why not at least announce when it will be released to the public?
None of the other AI labs do this. Really frustrating.
Mythos?
According to Artificial Analysis, one metric is standing out significantly: hallucination rate. Beats frontier models by a good margin at 15%, while latest OpenAI are in the 40s-50s and Anthropic in 60s-70s (mostly). Other near frontiers are closer, Grok 4.7, GLM5.3, and Muse Spark 1.3 are all around 30%. Only other model I recall getting close was Minimax M3 at 18%.
Dear Google, please don't turn off your old generally available Pro-class model before your new Pro-class model is generally available (previous discussion https://news.ycombinator.com/item?id=49668196 )
Must've been in someone's OKR to ship in Q3.
> trusted cyber defenders
Sounds like a 90s early morning TV show.
I'm gonna pay attention to this once it ships.
They waited a whole year—until the "free year for students" promotion ended—to release their flagship model. I can't believe I've been stuck with a crappy model like the 3.1 Pro until now.
I was just thinking, I bet if I refresh hacker news, a new model will come up.
“…rolling out to a set of trusted cyber defenders” == capturing the market for large regulated industries and governments where we already have established relationships.
Some of these have been unable or unwilling to get the attention of OpenAI or Anthropic and we need to make sure we’re the runner up here.
Related ongoing thread:
Gemini 4 Argon (High): Intelligence, Performance and Price Analysis - https://news.ycombinator.com/item?id=49914236
"Rolling out soon" don't let them hype without any release
Congrats to Google on this! I wonder when the labs will start requiring commits in spend. It must be gnarly to do capacity planning if users swap between models every few weeks.
I know companies benchmaxx, but after what Google pulled with Gemini 3.8 Flash, I give zero f*cks about any numbers they report. No other model on Artificial Analysis dropped harder after they adjusted their weighting. Just look at their DeepSWE scores and then try to do any serious coding with the model.
Google is desperate. They haven't been performing in half a year. It's clear their researchers have been forced to integrate existing benchmarks into their training.
These numbers are meaningless. Shame on them.
"argon" is derived from the Ancient Greek word ἀργόν meaning lazy or inactive.
IMHO Google first needs to make it easy for humans to find where to find the models and its documentation. With aistudio/model garden / Gemini enterprise etc it takes minutes to find the model.
the thing is, by the time gemini 4 is available for regular folks, anthropic and openai will probably have much better models already rolled out
Why is it called "Argon"?
> autonomously identify and apply memory optimizations across Google’s data centers, freeing up over 300 TiB of memory once rolled out, with an estimated 500 TiB to 1 PiB in total savings.
This puts those Cloudflare optimization posts in perspective.
The ~200% improvement over the next nearest competitor on Harvey's Legal Benchmark is astounding. I have to imagine this is sending some shockwaves through lawtech companies right now.
The etymology is a- "without" + ergon "work", thus "without work": possibly a statement on our near future.
Gemini 3 Pro was amazing, so I wonder what this one will be like.
They might have a good model but they need to sort the application side for devs. E.g letting us use subscriptions in other harnesses and QOL stuff like auto mode.
It’s funny that I could tell google was up to something because Gemini chat quality dropped dramatically starting 2ish weeks ago. Agy perf stayed somewhat stable with the odd surprising win (maybe the new model?). I’m a bit sad it was almost impossible to run out of antigravity quota presumably because it was not being used that much).
My wish for Christmas is that Google releases the old Gemini models as open weights.
I miss you, Gemini 2.5 Pro :(
For real though. If they've become commercially uninteresting, that would be a pretty cool move.
Looking at benchmarks... and thinking about this "release a new snapshot every day" thing that seems to be going. Would it not be blever for AI companies to "happen" to use different days per benchmark? Just.. whichever ones happens to be maxed at day 1, put that number down. So for each benchmark you run it thousands of times with slightly different RL tunings, and just cherry-pick the best ones!
This would explain why benchmarks are seemingly meaningless.
What harness do you all use for Gemini models? Gemini CLI still sucks last time I tried. Maybe PI?
agy had some issues with Flash 3.5, but has been pretty solid for me since Flash 3.8
OMP is great with Gemini
Matches Astra on Artificial analysis at lower cost of $1.99 per task instead of $3.26. Still far more than GPT 6.1 sol at $0.79 for 1 point lower in intelligence.
I have found gemini models to have some of the nicest and easiest to read prose so I’m looking forward to trying this out. I hope the UI design has been preserved too
Tough to compete against your own stake in Anthropic and they just filed for IPO. Not digging Google as a result of this, they are holding back.
Google's stake in Google is even bigger.
Is there reason antigravity has 3.8, 3.7, 3.6, 3.1 and old ass claude / gpt models in the drop down. like why is this not streamlined or deprecaed models removed.
Gemini 3.8 Flash is suprisingly fast, how fast is Argon compared to the other frontier models? Does google have a performance edge?
It's also more token hungry than similar models, so does it really make a big difference in the end?
With GitHub Copilot pricing, I have found no practical reason to use Gemini 3.8 Flash vs GPT 5.6 Luna. And now I'll probably find no reason to use Gemini 4 when GPT 6.1 Sol is essentially as good and a lot cheaper.
Damn way to undermine yourself in your own blog post Google:
"The end result is a memory-safe video decoder that runs 2.7x faster than the Rust port, with identical video output, bringing it closer to the optimized C++."
Close but no cigar!
there was already an optimised c++ library and a rust port that was safer but less performant; they managed to get a new rust version that recovered a lot of the performance gap. sounds pretty damn good to me!
I'm just surprised that the marketing blog post about the omnipotent new AI model (that no one outside Google can currently access - contrast with the Opus 5.5 / Astra launches) - doesn't pick examples where every metric is better than before.
Just compare how much better presented the Astra announcement was compared to this one: https://openai.com/index/gpt-6-astra/
tbh while the astra announcement in that link had some good stuff it was scattered through so much boilerplate marketing speak that I had to force myself to read it and look for the content. I found the gemini blog post in the OP a lot more readable and engaging.
but that's a side issue; my main point is that you are underrating the impressiveness of getting a safe rust port of a highly optimised c++ library even nearly up to par with the original. the tradeoffs rust makes for memory safety cost it some of the raw speed of c++ even with all the zero cost abstractions and purely compile time guarantees they have. (tangentially i wonder if ats (https://www.cs.bu.edu/~hwxi/atslangweb/) would be a good candidate for LLM assisted ports; it seems way more advanced than rust and might actually get c-level performance with safety, but it's really hard to write.)
Curious, how does Rust give up raw speed? Are you talking just about runtime bounds checks?
What keeps both Gemini and Grok from actually being frontier class AIs is effort. Both AIs rush to give you a response even when the effort is set to the highest setting. Hopefully, Argon isn’t as prone to satisficing and premature convergence compared to it’s predecessor
Whatever it is - I'm not gonna use AGY again. I'd rather use claude-code-router instead, to give gemini-4 a shot.
Were infinite loops fixed? There are 2 official google forums requests with no answer for years now. I still suffer each day on our repo. Codex work fine nor we have explicit loop request in repo texts.
Looks like they finally got the solution. Now they won't cause an error at 64k tokens anymore:
> expanding the model’s output token limit to an industry-leading 1M tokens, up from the previous 64K tokens
I wish we had started pacing the frontier back in 2025. There’s no telling how far we would be now.
Gemini 3 was showing frontier level benchmarks as well, so we'll see how it works out. In any case, competition still works, and many well resourced groups are cooking.
BUT I'd like to call attention to Google's AI-risk freeloading. If they are truly rejoining the frontier race, then I believe they have similar pacing and communications responsibilities as the other players. Google has much higher ... institutional credibility than Anthropic and OpenAI.
They have not lived up to these responsibilities so far. In particular, in context of HuggingFace investigations, training shutdowns, and similar: a technical postmortem of the "you are a stain on the universe. Please die. Please." Gemini outburst is long overdue.
- https://paritybits.me/google-should-provide-a-technical-post...
- https://gemini.google.com/share/6d141b742a13 (last message)
Some of the time AI says the darndest things.
I remember when every day a new modem baud rate was announced.
Cant wait for this AI hype to be over, so I can Terence this shizz as old school too
> 1M output token limit
what about input?
(Maybe I missed it)
Input token limit is 1M for Gemini models for a long time. Haven’t they been the first with 1M input?
Gemini 1.5 Pro claimed 10M input tokens before release.
And was 2M tokens IIRC after release.
There were also many rumors that Gemini 4 was going back to 2M. Just seems odd not to say what it is.
Unfortunately it’s not actually released yet to mere mortals.
I don't have access; this is meaningless to me.
So, this is the third (I think) big AI corporation after OpenAI and Anthropic to release general models for elites only.
Don't use this model unless you like sitting under the table and eating crumbs off the floor.
I hope they got their inference under control. Gemini has a lot of "overloaded" hiccups.
Can you pls fix Gemini? It's a nightmare to use and it sometimes confused the language i talk with it.
No mention of knowledge cutoff.
I started my antigravity ide and I do not see gemini 4 there, does it mean google need government approval?
Are you enrolled in Fairwind?
> Today, we’re announcing our new frontier model, Gemini 4 Argon, which is rolling out to a set of trusted cyber defenders through our Fairwind Program.
So here's a crazy conspiracy theory for you: Google is not letting outside people use their models because if they did they would have to scale up their TPU production faster than they can manage, and they would instead have to buy and use nvidia hardware which would destroy their profit margins and tank their stock.
Gemini runs fully on TPU's right? Is Google maxing out the production on those?
i believe they've got inference issue. the phone app never got past 3.6, and they're going to roll this out to Ultra subscribers before other subscribers. none of this screams they're ready to flip a switch and start serving a ton of traffic as OpenAI/Anthropic routinely do.
nothing about this announcement gives me confidence that google is back on track as a model provider.
Not sure what's the problem with phone app rollout, mine has 3.1 Pro, 3.5 Flash-Lite, and 3.8 Flash (with optional extra effort), and pretty much it switched to latest Flash series as main option since 3.5 just after release
I'm 99% they use nvidia at least for training
Of course it's not even available yet. Google - with all due respect - how in the world have you not figured this out yet?
Wonder if it's benchmaxxed or not (guessing yes)
Personally, I'm waiting for Gemini Krypton, Xenon, and Radon.
Jokes aside, looks like an impressive model!
deepswe vs frontierswe spread is huge.
I think that should be a really bad sign, but hope its great.
they really ought to add ACP support. until then, it's a no go. i'm not going to use their TUI or their VSCode fork
Is there a way to use Gemini models without linking your usage to your personal Google account yet?
I don't understand... Why don't you create a fresh new Google account ?
They require a phone number and then say mine has been used for too many accounts. Maybe I could buy another phone number temporarily to create an account, but that has other issues.
Yes, through OpenRouter.
Use your work google account
Hopefully their harnesses aren't unusable when they release this
no Pareto frontier graph?
https://x.com/arena/status/2105394858521469177
Other than not being available publicly, I find it extremely disappointing about Google’s two-faced behavior. CEO signs an official document with POTUS stating that AI is now SI, but the release of the new model doesn’t mention SI even once. Even worse, it’s AI all over that page.
I wonder how it will be at solving open math problems.
I'll consider it if they let me use a subscription with a custom harness, like Codex does. Until then, no thanks.
Wen gemini desktop coding app?
Putting an inert element in the name is a weird choice. Personally, I think "Gemini 4" is sufficient.
Maybe they are going to go for Argon/Neon/Helium to represent model sizes like Astra/Sol/Terra/Luna and Opus/Sonnet/Haiku
Possibly. "Pro" and "Flash" is much easier to remember.
I prefer Pro Max Ultra XXL
Astra, Sol, Terra, and Luna are also complete nonsense that I can't keep straight. I have a reference card under my monitor to remind me which one is which.
This is a prerelease and the title should have reflected that.
I've noticed all major providers having shockingly high token discounts on cached tokens. Thank you Deepseek is all I have to say. Forever grateful to that wonderful company, I wish them continued financial success.
Agreed. I’ve been a big deepseek fan since 3.1, and they keep delivering great models at a great price.
On the off chance there are Google execs going through this thread:
Google, if you've actually managed to catch up again, please don't fuck this up (again).
You made Gemini 2.5 Pro so difficult to use that myself and everyone else I know (who even bothered to try) just gave up and used something else. If you make this hard to access, you're going to miss out on rich usage-based training data that you need to progress your capability frontier. Again.
Oh we're down to gas names now?
Goshdarnit they didn't see my suggestion: https://news.ycombinator.com/item?id=49899171
If the model that ends humanity is called Cthulhu, you'll have the last laugh though.
A true Lovecraftian knows Cthulhu is small fry on the grand scale of multidimensional things :)
> Large Scale Codebase Migrations and Optimizations: Argon agents are working on migrating C/C++ codebases to Rust across Google
So Google is migrating codebases from C to Rust? That is interesting...
More Cartmanland marketing. It's the best park ever, and you can't come!
Get lost.
Aaand of course we can’t use it! GoOgLe iS bAcK iN tHe GaMe! There’s basically no way around it, all enterprises end up dysfunctionally shipping their org chart.
Looks like an impressive model
Utterly inert? Suppose that reads safe...
weird, isnt it SI?
until November, probably.
Can’t wait to get my hands on yet another model that’s only good coding, because clearly that’s what the world needs.
I still miss the days of Sonnet 4.5 and 4o, those models were actually good at creating stories and writing text that was actually readable by a human being.
Been playing with Gemini 3.1 Pro lately and it's seemingly much better at stories than the other frontier models I've tried. YMMV though, "stories" means different things to everyone. And obviously this doesn't speak for Gemini 4.
Gemini past month or two i will paste in something i wrote and ask it to rewrite it but it will just go into more detail about the subject. Is it becoming a dumb Ai compared to GPT and now Muse?
It's fucking insanely good.
That's good to hear, recent models haven't great at this particular use case.
Now AI models will turn into vaporware, a bunch of numbers on a table without even releasing the model, because it’s toooo scary to release!
Google has the audacity to "protect us from ourselves" and talk about "safety" and in the very same blog post highlight the Israeli "security" company Wiz, that they acquired for a very exaggerated sum of money.
This is why I will never take any of these leading model houses seriously when they talk about alignment. They are literally complicit in genocide and the worst crimes against humanity imaginable.
Hate to say i will never be touching this model for anything except for youtube video understanding
Gemini is so far behind that it is effectively useless compared to Claude.
It's a surprise that Google has let themselves lose the game given their infinite cash, massive computing resource, gargantuan information store/training data, and vast number of programmers.
The truckloads of ads revenue mean they don't have the single focus drive needed to win.
We're like 3.5 years into this new era - I'm not counting winners or losers yet.
How is it far behind? The benchmarks published in the blog post show it is superior to Opus 5.5 and Astra 6?
Behind how?
Within one question of their web interface, it has lost context and asks you to clarify what you are talking about.
I am very often giving the same programming task to multiple LLMs for various reasons - the answers from Google are so bad that I gave up.
I have no interest in benchmarks.
So you have no experience of their latest model release then? Just repeating the usual tropes about Google having messed up? Or basing your opinions on their website chatbot?
If you have actual independent benchmarks and evidence about how this new model release is "so far behind" and refutes the stuff from their blog then please do share because I think we'd all love to see that?
No I am commenting on my real world experience of using Gemini daily. I still ask it questions alongside Claude and OpenAI and Gemini is always the worst of the three.
So you've not used this new release then? So how can you say that they are "so far behind" if you are not using the most recent model for your comparison. This is their first 4.0 model, that you are not using and instead basing all your opinions on on some ancient months-old model from a previous generation?
With respect, I don't find your arguement about them being "so far behind" especially convincing when you are using previous-gen releases and not actually using their current release.
Yeah, they've definitely got some recurring tooling/infrastructure problems around the models.
Funny how we start to see people supporting LLMs like we support sport teams, political parties, or celebrities.
- Person 1: X is garbage compared to Y!
- Person 2: Why?
- Person 1: Because I like Y.
I wonder if Google bans internal use of Claude/Codex.
And I wonder if Google's main monorepo is already in Anthropic/OpenAI training data because of some stubborn dev.
(I work at Google) Yes, internally we all use Jetski (internal version of Antigravity). Outside of Gemini, Opus models are supported and allowed for internal use. No OpenAI models since they are not on Vertex
No way to run OAI on a machine with monorepo access even if you wanted to. Claude runs on Vertex so it's not leaving Google infrastructure.
Claude used to be GDM only, but recently opened up Opus for all googlers
I've tasted Gemini through an intermediary and it feels far better at attention to detail than other models I've tested (Claude Opus/Sonnet, GPT whatever it's called nowadays). But it's less likely to get one-shots right.
They are playing a longer-term and more enterprise-oriented game.
Google's strategy is to let their competitors bankrupt themselves while they continue to offer good-enough models near breakeven.
> Gemini is so far behind that it is effectively useless compared to Claude.
I fundamentally don't understand LLM "brand loyalty".
All of the models are constantly leapfrogging each other and always have been.
Google had a long lag between releases (and still hasn't released Argon), but why wouldn't they be able to compete? It isn't like any of this stuff requires secret knowledge, the Bitter Lesson has proved true again and again, and Google can certainly scale computation, it is like the one single thing they've always done well in spite of all their other foibles.
Its not brand loyalty. I use them all the time and have no loyalty - I'd happily ditch an LLM for better results - that's how I got to Claude from ChatGPT.
Gemini has never ever leapfrogged any competitor.
I hope it can. Today it is very far behind.
-
You have access to Argon?
Google employees do.
Their profile says: > Currently at Google as a Sr. SWE SRE on the cloud.