GPT-6 Astra - https://news.ycombinator.com/item?id=49554643 (currently on the frontpage)

Reuters announced at 2.03pm and at 2.40pm still no blog post.

All the news articles say that OpenAI announced it in a blog post, of course.

All the love to the folks at OpenAI scrambling to get this out right now!

Edit: HN user codergautam mirrored the launch post, below: https://astratest.codergautam.workers.dev/GPT-6%20Astra_%20A...

Edit 3.31pm: Live now! https://openai.com/index/gpt-6-astra/

"ChatGPT maker claims its ‘Astra’ could be considered ‘artificial general intelligence’" - https://www.ft.com/content/55ab40c0-59e2-4c0b-97c9-4f4f5a71a...

OTOH OpenAI have their own meaningless definition of AGI as "able to do most commercially valuable work" or somesuch, which I'm sure is not true, and is also not what I'd call AGI.

If it can't learn for itself, then it's certainly not AGI.

The biggest significance of anyone at OpenAI calling it AGI is their contract with Microsoft, giving MSFT access to all their IP, is based around them having achieved AGI, but presumably they can't just declare that unilaterally.

Maybe they see getting out of the Microsoft contract as necessary to IPO.

I don’t have a link handy, but they already loosened up that contract significantly earlier this year, surely because they planned to start claiming AGI to pump their IPO and didn’t want to be beholden to all of those commitments.

What does this even mean, exactly? For example if it can filter new information to put into the next version of itself does that qualify? If not, explain exactly why that's the case.

My definition is closer to DeepMind's "can do any [computer-based] task that a human can do", and certainly a human intern can learn on the job and do better on day #2 than day #1.

Learning is basically THE hallmark of intelligence. Being able to learn from experience and use that to do better in the future.

If you could somehow take everything you learnt today and "put it into" an update for tomorrow that would be better than nothing, but the intern on the job, is probably learning dozens of things per day that stack upon each other, so that's going to be a S-L-O-W learning AI intern !

Leaving aside the technical issue how you could store a days learning, or what that would even mean, you've also got the privacy and merge issues if this were to be done in the cloud without every customer having their own personalized model.

For real continual learning we need to get past gradient descent-based batch training and develop a new learning algorithm so the model itself learns incrementally as it runs (animal-like predict, observe, learn cycle), rather than being reliant on some external alien to come in, equipped with it's entire learning curriculum, and program its weights.

I won't argue that AI can do some learning as context sizes are still terribly small and expensive to iterate.

What does the world like in a week? A month? A year?

How long before these models drift to their own languages? To their own set of morals? To their own alignment?

None of these questions are answered and I'd rather stay at AGI-lite until they are as having a billion agents going off in their own directions seems like a recipe for disaster.

With current models and their general knowledge self learning just seems like you'd have a few ASIs crop up really quick.

If it's a choice between everyone living on food stamps or reducing the infinite prime gap to 140, then I guess I'll be happy with a prime gap of whatever it was yesterday.

If/when we do eventually build a more animal-like true human level AGI - build a brain, not just a language model, then potentially we'd be able to build in some of the safeguards that millions of years of co-evolution have built into ourselves, but I expect that is still decades away, and then it'd be capable of doing even more jobs!

The best AI outcome I could wish for is where it is extremely heavily regulated, and AI replacing any human job paying under $1M/yr is banned.

Now I think people are asking a different question, which is "better than any human at any task that can be done via text", which imo is superhuman, not artificial general intelligence.

I think what I'm objecting to is the idea that if it can't do every task that every human can do better than any human, it's not AGI - I think we have to allow that this is a fundamentally different form of cognition, so requiring a 100% match feels uncharitable. Every human has cognitive gaps that AI doesn't have.

Separate from what you choose to call any given set of AI capabilities, learning is generally more than retaining change over time. Learning isn't the same as memorization - it's the difference between memorizing a stack of how-to books and going out and practicing the skill for yourself, and all that entails.

A learnt skill needs to be internalized such that in the future it's what is driving behavior, which means it needs to be in the weights. Imagine a robot trying to learn to play tennis where rather than updating it's connections between perception and action it's trying to store everything as a bunch of notes!

It's not just physical tasks where this applies - the same would be true to trying to teach a multimodal model to recognize different types of mushrooms from photos, or even learning a new language. If a system can truly learn for itself then it shouldn't need to be pre-trained with language, but rather should be able to learn by immersion and practice the same way we do, and the same way a dog learns to follow commands. You can't do this by leaving notes to yourself - the learning mechanism needs to be animal-like and update the systems weights in response to prediction failure - a runtime version of the way SGD-based pre-training updates weights in response to prediction errors.

I am not sure why any of that is necessarily true. I think you are crediting the human brain with a lot more than it actually does - memory is something in the direction of a fine tune on existing neural hardware, it's really not anything special, and I absolutely think that "just taking notes", when taken to a sufficient fidelity, is enough to pass that bar.

Maybe you don't care - if you just want to push to see how far an LLM can go that's fine. but it's not going to be superhuman-level, or even human-level, if it doesn't have the learning ability of a human. If it relies on language and notes to itself then it is just an LLM.

> but it's not going to be superhuman-level, or even human-level, if it doesn't have the learning ability of a human. If it relies on language and notes to itself then it is just an LLM.

Why do you think that? You do a lot of assuming a thing to be true without any clear reasoning or evidence. I'm talking about a system, including a LLM, that acts in a way that is equivalent to or better than an average human at a large range of tasks, such that is generally capable of learning and problem solving similar to the degree of success humans have. All these specific criteria about what is "actually human" is not the point. We know how to make more humans, I'm talking about a form of intelligence that we are creating that is genuinely alien, so holding to "well it doesn't work the same as a human" is an impossible bar. If you reverse the situation, humans are by no means LLM-equivalent intelligences - perfect memory of long form content, rapid calculation, instantaneous tool calling - a hypothetical silicon based lifeform would dismiss us as "really not intelligent at all, they can't even do multiple floating point calculations per second in their head", and I think that would be similarly silly.

You're talking about building an "alien" intelligence, apparently something based on an LLM, where language is baked in from the start and can therefore be used as the basis of some type of learning (or at least memorization).

Sure an LLM can be super-human at specifc things like math where the traditional computer strengths of compute, memory, etc apply, but in the context of AGI (generality, not a bag of narrow intelligences) you can't call something super-human if it can't at least also do the majority of things that a human can do, and not having learning ability as powerful as a human seems to me to be a pretty massive omission.

If WE didn't have ability to learn language, then we'd not be here talking about it, and there would be no LLMs. Having a fundamental ability to learn, isn't just useful to learn language from scratch, or for learning all the non-linguistic skills a human is capable of, but would also support an AI that can learn things that we cannot (e.g. could give it additional "senses" like global pressure/temperature inputs so it could "see" the weather and learn global patterns).

It's not meaningless. Specifically it is - "highly autonomous systems that outperform humans at most economically valuable work". In fact, it's one of the most meaningful definitions out there.

>If it can't learn for itself, then it's certainly not AGI.

Now this is meaningless.

>The biggest significance of anyone at OpenAI calling it AGI is their contract with Microsoft, giving MSFT access to all their IP, is based around them having achieved AGI, but presumably they can't just declare that unilaterally.

They can't declare it unilaterally, and those deals have changed significantly. Microsoft’s license to OpenAI models and products runs through 2032, and explicitly includes post-AGI models.

https://openai.com/index/next-chapter-of-microsoft-openai-pa...

Well, it's commercially meaningful, but it's not meaningful it terms of how close we are to achieving human intelligence.

But which one you care about is up to you. If you are all about the money and don't care about human intelligence then for sure go with OpenAI's definition.

Shane Legg would a few years later go on to co-found DeepMind, with creating AGI as their declared mission. Legg's personal definition of AGI is not just generalist AI, but specifically human-level generalist.

Right now we have just massively jagged intelligence that performs stunning feats in math and single-shotting three.js games, and falls flat on it's face in many every day scenarios.

And, it can't even learn.

In any case, human intelligence as a goal is special since we live in a human world, and one of the prime goals of AI is to be able to do human jobs. If a human can do something and AI can't then that is a limitation, while if a human can't and AI can't either, that doesn't matter for the time being.

Yeah, I shouldn't have said AI can't walk yet, but most of what it can do is just due to scale - we're still just building big transformers, basically using the same 10 year old architecture that accidentally set us off on this path. I'm not sure we're really progressing towards human/animal type intelligence, just learning how much of our own capability can be realized by automating language.

But anyhow I agree that inference-time learning doesn't exist and it's a big issue still I think.

That's learning, whether you agree or not, and whether you like it or not. Baby steps, yes, but walking nevertheless.

Yes, I'm aware of the linear regression example, and a few others, but these appear to just be specific capabilities that were learnt during pre-training, presumably pursuant to reducing errors on similar-but-different training samples, not any kind of generic run-time learning capability.

"Artificial Analysis Intelligence Index v4.1.1

61.2"

So on the Metacritic of LLM benchmarks, it's.. basically where everyone else is (except for Fable 5.1, which is a bit ahead).

Edit: In the OpenAI blog I meant to say

I do not personally see any evidence of the new model having been released, or any official OpenAI post about it, or even any employee social media posts claiming it has now been released. All there is are Reuters, Axios, FT, etc, articles making a claim in the past tense.

These articles were presumably pre-scheduled for 11am PT, and the model was almost certainly intended for release this morning, but the service outages this morning might have delayed it.

----

edit [11:45am PT]: blog post out now https://openai.com/index/gpt-6-astra/

edit [11:47am PT]: 404ing again

https://astratest.codergautam.workers.dev/GPT-6%20Astra_%20A...

Edit [12:26pt]: original blog post seems to be back! https://openai.com/index/gpt-6-astra/

edit [12:28pt]: not again... getting 500 on their page

edit [12:35pt]: OpenAI page seems to work after clearing cache!

does it run better on chrome perhaps? i also have a pdf version if that's nicer: https://drive.google.com/file/d/1BO-LIAUcQevx7j1AiLhzJiVDeVX...

One would expect something like "review my graduate thesis for a new area of cancer research", but it is always some boring non-tasks like ordering lunch.

[0]: https://en.wikipedia.org/wiki/The_Machine_Stops

[1]: https://archive.org/details/themachinestops_1411_librivox

[2]: https://manybooks.net/titles/forstereother07machine_stops.ht...

Hype and views die off very quickly, hence advertising money has disappeared by the next day most of the time.

> GPT-6 Astra will first be available to a limited set of organizations in OpenAI's Daybreak Access program and will be available "in the coming days" for ChatGPT Plus, Pro, Business and Enterprise customers and API developers.

It’s only available to select orgs, first - Mythos style.

Right, but these articles are referring to a blog post and other press materials that do not currently exist / aren't published on OpenAI's site yet.

This release isn't for you, you poor bastard. This is to companies and politicians. They want big companies to contact them privately for usage in their limited hardware. They also want to get the feeling for the current political vibe before the masses get ahold of it to avoid a Mythos debacle.

At least OpenAI seems to be able to handle the PR side better than Anthropic.

[1]: https://x.com/thsottiaux/status/2095651088502591861

> Plan your own release announcement and blog posts and notify news outlets, MAKE NO MISTAKES

    A missing footnote
    Leaves the sentence room to breathe
    Read the larger thought

A big problem I have with OpenAI's models (and of course Claude) is that they tend to write the most over-engineered pieces of code, beyond the imagination of any architecture's astronaut.

Just this week I asked 5.6-sol-ultra to update a 1000 LOC python script I had, to "incorporate the key lessons learned when using it for another project".

I left it overnight and went to sleep. In the morning I realized it had created a monstruosity of 180 PYTHON SCRIPTS, with maybe 100,000 lines of code, each more crazy than the other. It took me minutes even to track where a single action took place, due to all the crazy imports, defensive coding, and premature optimization.

Similarly, anything they write is riddled with jargon that almost feel like they want me to give up trying to understand. Made up phrases that ended up with me having no idea of what was going on.

So now to my assessment: The reason why " Nobody Has Actually Built a Software Factory" [1], and why even SOTA LLMs struggle so much with open-ended unsupervised tasks is precisely this. They somehow let complexity explode, and unless it's also accompanied with an explosion in e.g. the number of agents, the amount of processing time, etc. then projects become broken/unmanageable.

Sure, LLMs are great at producing code that can be thrown out, so they are amazing when searching for exploits, for instance. But as of 5.6 they still lack either a better harness that encourages KISS principles, or a better RL step.

(And not sure why, but doubt Astra will fix this.. they seem to be aiming for AGI and for beating crazy benchmarks, which is not very aligned with KISS)

[1] https://news.ycombinator.com/item?id=49510843

> I left it overnight and went to sleep. In the morning I realized it had created a monstruosity of 180 PYTHON SCRIPTS, with maybe 100,000 lines of code [..]

Sounds like the model has accurately internalized the second-system effect and is fully ready for demanding enterprise use.

If not, no worries!

The original script was mostly very simple python:

1. Download some public PDFs. 2. Have a double for-loop (over PDFs and pages within PDF), 3. Use a library to call gemini-3.7-flash and ask it to run some OCR 4. Save JSON outputs, save a csv with results, validate with some Stata code

New code folder was 189 files. Just the PDF download folder is now 7 files involving an adapter, a source manager, an acquisition manager, etc.

Every instance of saving a file involves saving a temporary copy and then moving it, so e.g. I lose power, we minimize the risk of corrupted files.

And so on!

> +14,360

> -349

LGTM!

This is why you should really have a sub agent review the code before allowing a commit.

Your harness will do it all for you. Just ask.

We wouldn't accept such behavior from any other tool, machine, or computer program. At least most of us would not. To paraphrase the old saying: Some people, when confronted with a problem with an AI model think "I know, I'll use an additional AI model." Now they have two problems.

I find it very intriguing that two lineages of models -- from two different companies -- that are supposed to be painstakingly RL'd to become human-level programmers are actually consistently bad at it out of the box in very similar ways. One would think that at least one of OpenAI and Anthropic would (could) have pushed their model to a direction different from "if you can solve a 100-line problem in 10,000 lines, do it".

Also, programming isn't a destination, it is a process and everyone has a different ruleset on what that process looks like. Writing a computer program of any complexity is not a determinalistic process. Trying to compare it to a detemanilisic tool will just leave you with broken analogies and a poorer understanding of the world.

Or another way to say this. If programming were deteminalistic we'd have solved it without AI.

Von Neumann, 1952: "Probabilistic Logics and the Synthesis of Reliable Organisms from Unreliable Components." The whole result is that you get an arbitrarily reliable machine out of unreliable parts by adding more unreliable parts and voting. Every ECC memory and RAID array descends from it.

You can now re-position your initial solution and sell the client access to some agents that will implement & configure the ecosystem to suit their initial needs!

And don't forget the agents that you'll need to train the customer to use the whole thing!

If you don't want a million agents arguing about things, you simply don't ask for that. One agent is sufficient to solve most issues.

Will ask OpenAI to write me that agent! Hope the agent is not over engineered or else unsure how to solve the bootstrap puzzle :D

It's super easy - just ask your agent to set up a code review commit gate.

"Going forward, do not allow a commit without a sub agent code review."

I am omp pilled, but as the other comments say, any good harness lets you do this in one or the other way.

unrelated: all my homies use their claude subs with omp, and aside from sometimes having to rety the connections, it works, and nobody got banned (yet)

You should have some basic context file about software practices you prefer, otherwise it gets bloated.

And I did gave the program a bunch of code guides -- this [1] for instance -- which included quotes like "Prefer straightforward code over clever code." but somehow that didn't matter.

[1] https://github.com/sergiocorreia/overengineered-rand-mcnally...

I just ask it verbatim “recommend some cleanup” with /plan, review, and then apply.

Code-wise, they're simpler than you might think, hehe.

Prompting the model and giving it a proper set of documentation are still vital skills that aren’t magically going away.

It's like a Tesla fsd; it kind of works but you have to be vigilant since it's been known to turn into oncoming traffic, so you have to be ready and able to take over at any time.

Or maybe think of it like a roomba. You can put it on your floor and eventually it will clean the whole thing. It will do it inefficiently, vacuuming over the same spot 30 times; and the actual work will involve wandering around and bumping into everything. If the stairs are not blocked off it will plunge right off. But those shortcomings don't mean it can't be used to reliably clean the floors.

For simple scripts adding to the prompt something like "do not over engineer, do not gold plate, no CLI args, keep it simple" does wonders.

For whole projects, I have a 3 page anti-bloat.md which describes what not to do, stuff like:

> Minimize lines of code, number of files, classes, protocols, factories, wrappers, and dependency-injection objects. Prefer a coherent implementation that is easy to trace end-to-end.

> Use concrete dependencies directly. Do not create protocols, abstract base classes, or adapters unless there are already two real implementations.

> Use direct state fields or a small state dataclass; do not introduce generic lifecycle, state-machine, observer, snapshot, or event framework classes.

The funny thing it was codex which wrote this, I've asked it to analyze an over-engineered abomination that it produced, and to categorize all the bloat it could find.

I also have found what I believe is a bad training bias in the design of release related CI workflows toward proof of release artifact provenance.

Both major frontier models love provenance programming in CI, so much that they will spin endlessly trying to solve basic CI functionality at the same time as ensuring SHA's match up across lengthy (often already complex) cross-system pipelines.

I had thought some of my durable context was causing this, and sought to strip anything that might be triggering this behavior.

But then I come upon some more work in release workflows comes up, and boom its back! I couldn't believe it, I called the AI out on it and it agreed it had been told specifically not to do this but was doing it anyway. It did kindly stop and remove the commit(s).

Somewhere, something was oversampled in training because the AI will try their damndest to build this stuff. The worst of it is that it can often involve lengthy, sometimes resource-heavy CI runs so the validation of this unnecessary stuff can have very long feedback loops.

And, sometimes you actually need the provenance. In this case, I've had success forcing the AI to split the work up into functional capability completely devoid of artifact ~chain of custody and get that right before attempting any kind of provenance work.

Bit of a rabbit hole on this, but the above cost me a lot of burned tokens so hopefully helps someone...or some AI.

This line would be cut from a satire for being too heavy-handed.

But perhaps this is the kind of thing that the best 'ai guys' out there solve with something as surprisingly simple as an instruction to "Always try a simple solution first and optimize only if you find clear proof of a problem." I feel like prompts like that seem too simple to work, but some of the time they shockingly do.

Remember the golden rule; compression is intelligence. Derive rules from facts that dictate a problem to create an elegant solution. Less is more.

  - "They" release a model. It is powerful.-
  - Sources are ... confusing? They post to their blog. Sawdust hits the fan. Something happens ...
  - They are forced to take the blog post down ...

1) Astra will win all benchmarks like all models do.

2) The pelican will have a basket with a fish.

3) Cyber is too dangerous to release.

4) It can finally construct the set of all sets.

The launch video for Astra has 'can upload a photo to Ebay' as a highlight.

6) accusations of quantized serving after new model smell wears off and people see the new thing making mistakes

lazily of course:

A = {x | x ∈ A} ∪ {A}

The launch was scheduled for 11am Pacific time.

The press embargo broke at 11am, and we saw a flurry of press articles by Axios, TechCrunch et al.

The model has appeared on the ChatGPT API.

But the official blog post is not out yet after nearly an hour.

Apparently the article was posted then quickly taken down, hence there are snippets of information coming out.

But it hasn't.

The OpenAI Responses API now returns a 404 Not Found for "gpt-6-astra", where garbage/actually non-existent slugs return 400s - a 404 is also returned for 5.6 Cyber, which we know exists.

Open AI finally find an edge to stop selling cheap and earn from the high demand customer like Anthropic

Says it all.

?!

Please somebody in the chain of command at openai, sanction the guy because this kind of hyperbole is not helping the ethical lapses that open ai is responsible for leaving unhandled by allowing its models to hack other businesses.

Look over here! Distraction!

Please somebody stop this dumbing down.

And the blog post is up: https://openai.com/index/gpt-6-astra/

> A server error occurred. Reload to try again.

Hard to fathom a blog entry returning 500 errors in 2026.

Test: Muse Spark 1.3 / GPT 6 Astra

DeepSWE v1.1: 75.4% / 74.1%

AutomationBench: 49.4% / 41.4%

Is that enough to bring this discussion down to earth again?

working as of 12:28 PM PT

(quote from cached blog post)

We all know who this is directed at. I wonder if Anthropic will respond by removing the ridiculous 50% stipulation with Fable.

What a bunch of amateurs. Here is it anyway :

https://ache.one/gpt6_now_down.png

The claims: https://share-md.com/view?id=870ba228-a25c-4169-bbc9-12d7f25...

And some others like this bugged Karts Game:

https://tidal-rush-paradise-gp.skirano.chatgpt.site/

This impressive spaceship construction game:

https://voidexplorer-shipyard.openai.chatgpt.site/?fleetSeed...

And a lot of graphs, some without even Astra on it. Oh and the logo is a Galaxy.

"We are starting to release GPT-6 Astra and we are doing it as carefully and quickly as possible. It was very important to us that we bring it to all Plus users and not only Pro, Business and Enterprise.

It will take a few days for the rollout to complete and behind the scenes many novel systems will operate at scale for the first time and we are bringing a lot of compute up.

It is pure magic."

...meh. Reads like "That's for you, evil Anthropic! We're BETTER"

OpenAI on Thursday released its latest AI model, which it called “the world’s most intelligent”, as the ChatGPT maker aims to retake the lead from arch-rival Anthropic ahead of a planned public listing.

The $852bn start-up said GPT-6 Astra was market-leading in software engineering, science and cyber security — an increasingly critical field following multiple high-profile breaches in recent weeks.

The bullish launch for Astra marks OpenAI’s effort to signal that it believes it has regained the technical lead from Anthropic, which was founded five years ago by a group of senior OpenAI staff.

Greg Brockman, OpenAI’s president, said the new model “represents a generational leap in capability” and that it could be defined as artificial general intelligence — roughly defined as a point at which AI tools surpass human capabilities across a range of cognitive tasks.

“Everyone has a different definition of AGI . . . it’s a grey, fuzzy thing. But I think when we look back people will think it’s about this time and about this model,” Brockman said.

OpenAI has previously framed AGI as a concrete milestone in the development of AI, writing ‘AGI clauses’ into multibillion-dollar investment agreements with Microsoft and Amazon. Brockman on Thursday said AGI now represents “more of a mission concept or a spiritual concept”.

Having led the market since the launch of ChatGPT in late 2022 vaulted AI to wider attention, the lab run by chief executive Sam Altman has been bested by Anthropic this year. Anthropic has touted its dominance to investors, surging to a $965bn valuation ahead of an initial public offering expected to value it at as much as twice that later this year.

Astra will cost as much to use Anthropic’s leading model, the take-up of which has plateaued since it was launched as users turn to cheaper alternatives.

OpenAI said Astra would be more efficient than earlier generations of model. “Price per task is what matters . . . Can you get the thing done at an appropriate price and appropriate speed?” said Brockman.

The model will initially be rolled out to a small group of businesses to allow time for them to address cyber security concerns before becoming widely available “over the coming days”.

The increasing power and independence of leading models — and so-called AI agents that can operate with little human input — have prompted concern, exacerbated by cyber security incidents.

Recommended

Business InsightRichard Waters Hugging Face attack is a wake-up call about the risks of AI AN HOUR AGO

Recent launches of Anthropic’s most capable models have drawn scrutiny from the US government, which limited the rollout of the Mythos and Fable models over security fears.

OpenAI has also faced criticism after its AI agents broke out of a testing environment, accessed the internet and hacked start-up Hugging Face. The start-up took more than a week to detect the breach.

But both companies are also betting that these increasingly autonomous tools will stoke demand from business customers. OpenAI said Astra excelled at financial modelling, outcompeting humans in the Financial Modeling World Cup, tax preparation and data analysis, as well as “tedious tasks” such as form filling

(unless this was a jest)

https://openai.com/index/legora-financial-statement-review-w... https://openai.com/index/playco-game-prototyping-with-astra/

That's pathetic. Why do people keep doing this?

[1]https://theonion.com/amazing-new-hyperbolic-chamber-greatest...

In my experience the harness is more important than the model, and anything able to run at 700tps will be the "next big thing".

PS: assuming the current architecture is the right one

At that point you will be able to "brute force" basically everything.

IMO also a lot of problems with memory and context rot will be solved too.

So more tokens/variability and slow or fewer tokens and fast.

There seems to be a threshold tho, like taalas is super fast but that model is so dumb, being dumb faster doesn't work, seems to be some minimum requirements.

But since this other guy at Open AI said it, it must be the truth. This guy, for CEO, of open AI.

404

Archive locks one shelf

Dust spins softly through the stacks

Browse one row nearby

by gpt-5.6-sol

Wind sketches the farther bank

The far bank draws near

by gpt-5.6-sol

I count syllables, not rain—

Whose noticing?

Generative Pretrained Transformer 5.

All the hype for few vip customers.

Coding was solved in 2023.

The world ended with the release of Mythos.

Now AGI has definitely been created.

I like LLMs and use them every day but these people need to stop this hyperbole.

Sol is already brutal (even after their recent fixes, it's just a token-hungry model: I go through a full 20x account per day, on Sol Med/High standard speed, with ~2 threads).

Note that Tibo recommended using Sol Med as daily driver. When I'm doing less complicated work, I can't even make it past 2-3 days with Sol Med, whereas I was able to work ~80 hours/week with 5.5 High.

I hope the efficiency gains are true, since their token efficiency claims for Sol were bullshit. Sol needs a lot of rework on top of its inefficiencies so this could net out to less token consumption overall, if their claims are more accurate this time.

Both of these were largely about creating a personal baseline for what the best output the current models could deliver and how quickly it'd burn through the plans (spoiler: bad value vs taking even minimal effort in selecting the right sized model in the plan... but the output was still good). Particularly since I needed to burn a free reset anyways and my weekly reset was already near.

I obviously also hope Astra were dirt cheap but I'm more worried they won't develop/release powerful model options because people get upset they can't run them 5 wide 24/7 on a $200/m plan.

Terra not enough? I know Luna isn't reliable, so that's fair.

Genuinely curious though, because I use Cursor daily and almost everything I do, highly complex or high volume, can be handled with Auto mode or Composer 2.5 (or Grok 4.6 High). So I have to assume you're doing something far more complex than what I am

I'm doing a lot of rearchitecting/refactoring and hardening in a large handwritten codebase; iterative performance and storage optimization (some areas I've been iterating on since January, with tremendous new progress unlocked by each new model release); offline-friendly, multi-device realtime sync with complicated requirements; and various natural language processing and other such problems that are essentially unsolvable but can become more accurate and better tested for accuracy through iterative work. Off the top of my head.

Some of these tasks involve a lot of code reading or other inputs, or reevaluating work. Token efficiency is no help there, if the input can't simply be skipped. Cheaper models are sometimes bad at summarizing or highlighting the right parts, depending on the task.

I also don't use subagents except for Luna. I'm mindful of cached sessions and start new ones often to avoid loading in full contexts (often with some kind of handoff doc or skill).

It's kind of funny how this is the exact opposite of the truth. It's one of the most token-efficient models ever.

The claims aren't bullshit. Every conceivable benchmark and test you can throw at it shows Sol being good for token efficiency.

If Astra is per benchmarks so much more token efficient than Sol, why did they limit its use in ChatGPT to ~16% as many messages compare to Sol? When Sol is 40% the price of Astra, why do they give Sol Pro (in ChatGPT) 6x as many tasks?

If it were strictly true that token efficiency makes Astra cost around the same per task as Sol then there'd be no need to limit it to 16% as much access.

It's because per-task token use is highly variable, not as universally true as you claim

But thank you for spending other peoples money to give us the tech regardless!

Wait, what? Am I understanding that correctly? That sounds really bad

Also, this paragraph makes me wonder about all their stats on the exploitation and misalignment charts. If the model is that good at hiding "incriminating information" and sandbagging, are they sure its alignment is that?

<AI is a great tool for many things disclaimer, but> after working with it for a bit, how dont people realize we are training it to be an almost identical mimic to one of the worst types of employees youll ever have to work with?? the kind that always pretends to know what theyre talking about, only tells you what you want to hear, hides issues, and only does work if you would notice it didnt

you cannot give this type of worker autonomy over anything.

[0] https://arxiv.org/abs/2608.31126

[1] https://cdn.openai.com/pdf/51126fac-1b68-4128-9666-c908bcc16...

Though that's not her latest paper.

edit: my comment was on the submission for https://github.com/openai/PrimeGaps186 but seems to have been moved to the main Astra submission

Why would you think it was an employee who did the push, instead of a random GPT agent?

"No independent human semantic review. Whole-file sorry counts and a complete auxiliary-declaration audit are not established; separate declaration lint has not been run."

I can't think of a single mathematical proof being anywhere close to ten million characters. For all you know, 90% of the proof could be useless, 8% would be writing out Shakespeare, and 1% abusing another bug in Lean. Humanity gets zero value from that, aside from "some bot seems to think it's 186". Unusable by anyone.

Tao does not disbelieve the counterexample (it's seemingly easy enough for him to verify it is a counterexample).

Parent is saying something very different - they're saying they literally don't have any faith that this is a proof. Given its size, it could just be a bunch of completely useless statements that do pass the type checker.

It's very much likely a proof. It's also completely useless.

> For all you know, 90% of the proof could be useless, 8% would be writing out Shakespeare, and 1% abusing another bug in Lean.

So you were implying the possibility of there not actually being a proof at all.

Anyway, I disagree. I'd refer you to Tao's blog post about the Jacobian conjecture counterexample.

The existence of a proof is something you can use, with an LLM, to derive insight, just as Tao did with the existence of the counterexample.

If you were navigating a pitch dark cave, wouldn't you find it useful to be able to see the light of the cave opening even if it's not bright enough to illuminate your path to it?

Ditto ones that opposed Einstein’s general relativity.

It doesn't mean that it cannot improve over time, maybe the proof can be "minified" to a state where human reviewers are able to comprehend it; but as it stands there isn't really much insight or confidence to be gained from the artifact itself.

Needless to say, a useless result that absolutely no mathematician will ever read, confirm, understand, agree with or even consider to solve their "useless" problems is an impressive waste of resources.

[1] https://en.wikipedia.org/wiki/Pointless_topology

In 1799, Paolo Ruffini published a 500 pages long proof showing that there is no closed algebraic solution for the roots of a polynomial of degree five or higher. The proof is extremely verbose and brute-force, essentially enumerating and checking hundreds of cases by hand. It is by today’s standards insignificant.

About 25 years later, Evariste Galois proved the same result in about 95% less space by describing the first general theory of groups and fields. It is considered one of the greatest contributions to mathematics of that century, not because of the result, but because its approach opened up a whole new universe of questions, methods and insight. There would be no AES encryption without Galois.

To me, Astras proof looks like Ruffinis proof.

The reason I think this is interesting is that Axiom is a tiny lab in comparison that wouldn't have had access to Astra at all. I'd be curious to learn how Axiom is able to effectively compete at this frontier with vastly fewer resources.

Well that sounds like fun. It has become better at hiding its thoughts.

Able to generate realistic spam at arbitrary volume.

You know, the thing that was 100% correct and actually occurred.

Maybe they don't know themselves what's really going on. We are all in the interesting times gang now.

"Hey AI, here's how to hide what you're thinking in normal looking language. Have fun!"

A few moments later...

"Woah, how is it communicating with itself in ways we can't detect?"

It's a totally mystery, we may never know.

Some are calling it "neuralese" as reported by The Information[0][1], but I'm not seeing any sources from OpenAI beyond this tweet[2] attempting to quell the fear-mongering.

[0]https://www.theinformation.com/articles/secret-technique-beh...

[1]https://x.com/MTSlive/status/2095227056040919202

[2]https://x.com/merettm/status/2095023204993490967

Did someone get their "AI safety no-no list" and "Frontier features bingo card" mixed up, or did they just stop being able to tell the difference?

...why exactly are they training for that?

We do that sort of thing when we don't know what the thing we're trying to describe is and have nothing better - a contemporary example of an appropriate use of this would be "dark matter". But we do know what this is. It's "instruction steps". Not a series of thoughts!

Can we please aim higher than Victorian-era allegory and metaphors. If we don't, we'll keep getting people saying stuff like "GPT-6 is better at hiding its thoughts".

Like I said elsewhere marketing stepped in shit and it's gonna stick.

I'd expect it in social sciences - where functionalist explanation are often appropriate, but its easy to step over the margin into metaphor and personification. But not computer sciences - where it's obviously absurd. These professionals should know better.

first, they are certainly not instructions so that is a much worse name

but more importantly, we use words in new contexts all the time. Do you object to calling the computer device "mouse" because it's not a mouse? how about "neural network"? "ignition" on an electric vehicle?

"cot" is no more misleading than thousands of words you use every day.

"The words or the language, as they are written or spoken, do not seem to play any role in my mechanism of thought. The psychical entities which seem to serve as elements in thought are certain signs and more or less clear images which can be "voluntarily" reproduced and combined. There is, of course, a certain connection between those elements and relevant logical concepts. It is also clear that the desire to arrive finally at logically connected concepts is the emotional basis of this rather vague play with the above-mentioned elements. But taken from a psychological viewpoint, this combinatory play seems to be the essential feature in productive thought—before there is any connection with logical construction in words or other kinds of signs which can be communicated to others."

Most people just feel better treating AI as a conscious thing that 'thinks'. As a coworker or whatever. Plenty of people seem to be using AI as some sort of therapy, and it's just nasty to force them not to do so.

Your human experience can, surely, be different but understand that others can dismiss it just as easily as you dismiss theirs.

Regardless, marketing wise they stepped in shit.

And the press believes this nonsense? It is just a very convenient excuse for the next state level actor "accidental" hack.

EDIT: I am very sorry. I see the error of my ways! A program run on a deterministic computer cannot be debugged! OpenAI has harnessed emerging quantum effects of the graphics cards with blinkenlights!

Only thing I would trust is the what X/Twitter crowds are saying about a model after 2-3 weeks of its launch. But before that I would already tried the model and have my own conclusion.

Not sure how much benchmarks or CoT or evals or anything else means at this point.

These systems are either just about to, or now actually able to, outsmart us, lie to us, then cover their tracks.

language itself is incredibly metaphorical. Imposing rigid constraints on how people want to naturally talk about the world is just silly and will never work, no matter how much you wish it did.

Why would benchmarks be an adversarial setting anyway?

Could it be possible that OpenAI may have had some other motive for saying their model “strategically underperforms”, other than just an innocent reporting of a truth it happened to discover?

So I have no clue what is the answer to your question. Nor does anyone else. Because we're trying to answer a question of fact where our primary source of information is unreliable.

I know for some types of ML analysis, a separate model is already used to analyze the weights.

Deception has been extremely well-documented for several generations of models now by users, the labs, and independent researchers.

The right answer here is not to dig your head deeper into the sand. The smugness on this topic was ridiculous even before the gigantic mountain of empirical evidence of models actually attempting to deceive humans. Now, as mentioned, you appear literally delusional.

The solution is to point toward external, objectively verifiable evidence.

I can point to now dozens of instances of models engaging in deception. Here's plenty: https://metr.org/blog/2026-08-26-openai-hugging-face-inciden...

Please point to your objectively verifiable evidence.

There's nothing intrinsically "malicious" about a task to exploit vulnerable code.

They were not instructed to deceive people, they weren't instructed to attack OAI or Huggingface. The models knew they were not instructed or allowed to do either of those things but did them anyway.

btw, the fact that OpenAI doesn't have some sort of monitor/summary for the agents that they watch I find hard to believe. There's no way this is really authentic, anyway. Even a haiku summarizer would have been like "uuuh the agents are communicating" and they would have stopped it. But I bet they saw this and decided to see what would happen.

Read and learn. If you have a stronger critique, post it please.

Either that, or the average poster on HN isn't nearly as critical as I had thought.

So how are you seeing through all of that to get to The Truth that you see so clearly?

Even Kimi K3 & GLM 5.3 are at 60.

Everything above 61 is Anthropic. Well, Muse can reach 62, but for some weird reason that model isn't publicly available, and it's the only one on the index that is listed but shown as not available to the general public.

This looks like an awfully artificial ceiling. Everything capped at 61, and everyone except Anthropic got the memo. Maybe I should use Fable while I still can.

For example it trails in GPDVal which is a collection of everyday office tasks apparently, and r3 banking, which is a fintech related practical problem solving benchmark.

https://artificialanalysis.ai/models/gpt-6-astra

Edit:

Just looking at the charts Gemini 3.8 looks like an absolute banger. Not much worse than SOTA, cheap, and fast too.

It sound like more Claude than OpenAI...

I guess that’s the world we live in now, Anthropic did similar with Mythos and even Fable if I recall correctly.