> We study a novel language model architecture that is capable of scaling test-time computation by implicitly reasoning in latent space. Our model works by iterating a recurrent block, thereby unrolling to arbitrary depth at test-time. This stands in contrast to mainstream reasoning models that scale up compute by producing more tokens. Unlike approaches based on chain-of-thought, our approach does not require any specialized training data, can work with small context windows, and can capture types of reasoning that are not easily represented in words. We scale a proof-of-concept model to 3.5 billion parameters and 800 billion tokens. We show that the resulting model can improve its performance on reasoning benchmarks, sometimes dramatically, up to a computation load equivalent to 50 billion parameters.
“My vibes don’t match a lot of the traditional A.I.-safety stuff,” Altman said. He insisted that he continued to prioritize these matters, but when pressed for specifics he was vague: “We still will run safety projects, or at least safety-adjacent projects.” When we asked to interview researchers at the company who were working on existential safety—the kinds of issues that could mean, as Altman once put it, “lights-out for all of us”—an OpenAI representative seemed confused. “What do you mean by ‘existential safety’?” he replied. “That’s not, like, a thing.”
>Please provide the definition of Existential Safety.
I read:
>Are you mentally stable? Our product would never hurt humanity--how could any language model?
I don’t think predictions, but they did a great call until now.
If I was VP of Unethical Business Strategy at OpenAI or Anthropic, the first thing I'd do is put in place an automated system which flags accounts, prompts, IPs, and usage patterns associated with these benchmarks and direct their usage to a dedicated compute pool which wouldn't be affected by these changes.
I get the security aspect, but if we've hit that point any reasonably sophisticated model past this point will be able to do the damage they claim it can do. They might as well be telling us they're closing up shop for consumer models.
They should just say they'll never release a model of this caliber to the public at this point and say out loud we'll only get gimped versions.
This is already happening to some degree, GPT 5.3 Codex's security capabilities were given exclusively to those who were approved for a "Trusted Access" programme.
Rent-seeking of old was a ground rent, monies paid for the land without considering the building that was on it.
Residential rents today often have implied warrants because of modern law, so your landlord is essentially selling you a service at a particular location.
However, I’m tempted to compare to GitHub: if I join a new company, I will ask to be included to their GitHub account without hesitation. I couldn’t possibly imagine they wouldn’t have one. What makes the cost of that subscription reasonable is not just GitHub’s fear a crowd with pitchforks showing to their office, by also the fact that a possible answer to my non-question might be “Oh, we actually use GitLab.”
If Anthropic is as good as they say, it seems fairly doable to use the service to build something comparable: poach a few disgruntled employees, leverage the promise to undercut a many-trillion-dollar company to be a many-billion dollar company to get investors excited.
I’m sure the founders of Anthropic will have more money than they could possibly spend in ten lifetimes, but I can’t imagine there wouldn’t be some competition. Maybe this time it’s different, but I can’t see how.
you have 2 labs at the forefront (Anthropic/OpenAI), Google closely behind, xAI/Meta/half a dozen chinese companies all within 6-12 months. There is plenty of competition and price of equally intelligent tokens rapidly drop whenever a new intelligence level is achieved.
Unless the leading company uses a model to nefariously take over or neutralize another company, I don't really see a monopoly happening in the next 3 years.
I was focusing on a theoretical dynamic analysis of competition (Would a monopoly make having a competitor easier or harder?) but you are right: practically, there are many players, and they are diverse enough in their values and interest to allow collusion.
We could be wrong: each of those could give birth to as many Basilisks (not sure I have a better name for those conscious, invisible, omni-present, self-serving monsters that so many people imagine will emerge) that coordinate and maintain collusion somehow, but classic economics (complementarity, competition, etc.) points at disruption and lowering costs.
Not only that, but open-weight and fully open-source models are also a thing, and not that far behind.
Rent seeking isn't about whether the product has value or not, but about what's extracted in exchage for that value, and whether competition, lack of monopoly, lack of lock in, etc. keeps it realistic.
There is no real barrier to a customer of Anthropic adopting a competing model in the future. All it takes is a big tech company deciding it’s worth it to train one.
On the other hand, Visa/Mastercard have a lot of lock-in due to consumers only wanting to get a card that’s accepted everywhere, and merchants not bothering to support a new type of card that no consumer has. There’s a major chicken and egg problem to overcome there.
They actually beat Apple A series to become the first phone to use the TSMC N7 node.
And businesses from these other countries would happily switch to Chinese. From security perspective both Chinese and US espionage is equally bad, so why care if it all comes down to money and performance.
MC/Visa duopoly is an example of lock-in via network effects. Not sure that that applies to a product that isn't affected by how many other people are running it.
True, but it's also true that the returns from throwing money to the problem are diminishing. Unless one of those big players invents a new, propriatery paradigm, the gap between a SOTA model and an open model that runs on consumer hardware will narrow in the next 5 years.
Also eventually these WEIGHTS will leak. You can’t have the world’s most valuable data that can just be copied to a hard drive stay in the bottle forever, even if it’s worth a billion dollars. Somehow, some way, that genie’s going to get out, be it by some spiteful employee with nothing to lose, some state actor, or just a fuck up of epic proportions.
You should be more concerned about killer AI than rent seeking by OpenAI and Anthropic. AI evolving to the point of losing control is what scientists and researchers have predicted for years; they didn’t think it would happen this quickly but here we are.
This market is hyper competitive; the models from China and other labs are just a level or two below the frontier labs.
Not sure how this is consistent with "One private company gatekeeping access to revolutionary technology"?
You have to decode feel-good words into the concrete policy. The EAs believe that the state should prohibit entities not aligned with their philosophy to develop AIs beyond a certain power level.
> They should just say they'll never release a model of this caliber to the public at this point and say out loud we'll only get gimped
Duh, this was fucking obvious from the start. The only people saying otherwise were zealots who needed a quick line to dismiss legitimate concerns.
I read it like I always read the GPT-2 announcement no matter what others say: It's *not* being called "too dangerous to ever release", but rather "we need to be mindful, knowing perfectly well that other AI companies can replicate this imminently".
The important corps (so presumably including the Linux Foundation, bigger banks and power stations, and quite possibly excluding x.com) will get access now, and some other LLM which is just as capable will give it to everyone in 3 months time at which point there's no benefit to Anthropic keeping it off-limits.
That’s not going to happen. If you recall, OpenAI didn’t release a model a few years ago because they felt it was too dangerous.
Anthropic is giving the industry a heads up and time to patch their software.
They said there are exploitable vulnerabilities in every major operating system.
But in 6 months every frontier model will be able to do the same things. So Anthropic doesn’t have the luxury of not shipping their best models. But they also have to be responsible as well.
Error with 91.3% = 8.7%
Error with 94.5% = 5.5%
Error reduction = 8.7% - 5.5% = 3.2%
So the improvement is 3.2% / 8.7% = 36.8%> Importantly, we find that when used in an interactive, synchronous, “hands-on-keyboard” pattern, the benefits of the model were less clear. When used in this fashion, some users perceived Mythos Preview as too slow and did not realize as much value. Autonomous, long-running agent harnesses better elicited the model’s coding capabilities. (p201)
^^ From the surrounding context, this could just be because the model tends to do a lot of work in the background which naturally takes time.
> Terminal-Bench 2.0 timeouts get quite restrictive at times, especially with thinking models, which risks hiding real capabilities jumps behind seemingly uncorrelated confounders like sampling speed. Moreover, some Terminal-Bench 2.0 tasks have ambiguities and limited resource specs that don’t properly allow agents to explore the full solution space — both being currently addressed by the maintainers in the 2.1 update. To exclusively measure agentic coding capabilities net of the confounders, we also ran Terminal-Bench with the latest 2.1 fixes available on GitHub, while increasing the timeout limits to 4 hours (roughly four times the 2.0 baseline). This brought the mean reward to 92.1%. (p188)
> ...Mythos Preview represents only a modest accuracy improvement over our best Claude Opus 4.6 score (86.9% vs. 83.7%). However, the model achieves this score with a considerably smaller token footprint: the best Mythos Preview result uses 4.9× fewer tokens per task than Opus 4.6 (226k vs. 1.11M tokens per task). (p191)
By epoch AIs datacenter tracking methods, anthropic has had access to the largest amount of contiguous compute since late last year. So this might simply be the end result result of being the first to have the capacity to conduct a training run of this size. Or the first seemingly successful one at any rate.
ARC-AGI-3 might be the only remaining benchmark below 50%
GPT 5.4 Pro leads Frontier Maths Tier 4 at 35%: https://epoch.ai/benchmarks/frontiermath-tier-4/
That is, it's easy to make benchmarks which humans are bad at, humans are really bad at many things.
Divide 123094382345234523452345111 by 0.1234243131324, guess what, humans would find that hard, computers easy. But it doesn't mean much.
Humanity's last exam (HLE) couldn't be completed by most of humanity, the vast majority, so it doesn't really capture anything about humanity or mean much if a computer can do it.
You can't consistently benchmark something that is qualitative by nature. I'm struggling to understand how people don't understand this.
OpenAI had a whole post about this, where they recommended switching to SWE-bench Pro as a better (but still imperfect) benchmark:
https://openai.com/index/why-we-no-longer-evaluate-swe-bench...
> We audited a 27.6% subset of the dataset that models often failed to solve and found that at least 59.4% of the audited problems have flawed test cases that reject functionally correct submissions
> SWE-bench problems are sourced from open-source repositories many model providers use for training purposes. In our analysis we found that all frontier models we tested were able to reproduce the original, human-written bug fix
> improvements on SWE-bench Verified no longer reflect meaningful improvements in models’ real-world software development abilities. Instead, they increasingly reflect how much the model was exposed to the benchmark at training time
> We’re building new, uncontaminated evaluations to better track coding capabilities, and we think this is an important area to focus on for the wider research community. Until we have those, OpenAI recommends reporting results for SWE-bench Pro.
Anthropic accounts for this
>To detect memorization, we use a Claude-based auditor that compares each model-generated patch against the gold patch and assigns a [0, 1] memorization probability. The auditor weighs concrete signals—verbatim code reproduction when alternative approaches exist, distinctive comment text matching ground truth, and more—and is instructed to discount overlap that any competent solver would produce given the problem constraints.
But I do not use extra high thinking unless its for code review. I sit at GPT 5.4 high 95% of the time.
Not always, no, and it takes investment in good prompting/guardrails/plans/explicit test recipes for sure. I'm still on average better at programming in context than Codex 5.4, even if slower. But in terms of "task complexity I can entrust to a model and not be completely disappointed and annoyed", it scores the best so far. Saves a lot on review/iteration overhead.
It's annoying, too, because I don't much like OpenAI as a company.
(Background: 25 years of C++ etc.)
At least until next week when Mythos and GPT 6 throw it all up in the air again.
That said, I'll often throw a prompt into both claude and chatgpt and read both answers. GPT is frequently smarter.
GPT-5 is good at benchmarks, but benchmarks are more forgiving of a misaligned model. Many real world tasks often don't require strong reasoning abilities or high intelligence, so much as the ability to understand what the task is with a minimal prompt.
Not every shop assistant needs a physics degree, and not every physics professor is necessarily qualified to be a shop assistant. A person, or LLM, can be very smart while at the same time very bad at understanding people.
For example, if GPT-5 takes my code and rearranges something for no reason, that's not going to affect its benchmarks because the code will still produce the same answers. But now I have to spend more time reviewing its output to make sure it hasn't done that. The more time I have to spend post-processing its output, the lower its capabilities are since the measurement of capability on real world tasks is often the amount of time saved.
RE is very interesting problem. A lot more that SWE can be RE'd. I've found the LLMs are reluctant to assist, though you can workaround.
Me: Let's figure out how to clone our company Wordpress theme in Hugo. Here're some tools you can use, here's a way to compare screenshots, iterate until 0% difference.
Codex: Okay Boss! I did the thing! I couldn't get the CSS to match so I just took PNGs of the original site and put them in place! Matches 100%!
Once you set thinking to high it works just as well as 5.4 even for pretty complex tasks
Meanwhile, there are half a dozen other projects (business apps, web apps etc) where it works well.
(edit: I hope this is an obvious joke. less facetiously these are pretty jaw dropping numbers)
> Terminal-Bench 2.0: 82.0% / 65.4% / 75.1% / 68.5%
> GPQA Diamond: 94.5% / 91.3% / 92.8% / 94.3%
> MMMLU: 92.7% / 91.1% / — / 92.6–93.6%
> USAMO: 97.6% / 42.3% / 95.2% / 74.4%
> OSWorld: 79.6% / 72.7% / 75.0% / —
Given that for a number of these benchmarks, it seems to be barely competitive with the previous gen Opus 4.6 or GPT-5.4, I don't know what to make of the significant jumps on other benchmarks within these same categories. Training to the test? Better training?
And the decision to withhold general release (of a 'preview' no less!) seems to be well, odd. And the decision to release a 'preview' version to specific companies? You know any production teams at these massive companies that would work with a 'preview' anything? R&D teams, sure, but production? Part of me wants to LoL.
What are they trying to do? Induce FOMO and stop subscriber bleed-out stemming from the recent negative headlines around problems with using Claude?
We're not reading the same numbers I think. Compared to Opus 4.6, it's a big jump nearly in every single bench GP posted. They're "only" catching up to Google's Gemini on GPQA and MMMLU but they're still beating their own Opus 4.6 results on these two.
This sounds like a much better model than Opus 4.6.
We must not be.
That's why I listed out the ones where it is barely competitive from @babelfish's table, which itself is extracted from Pg 186 & 187 of the System Card, which has the comparison with Opus 4.6, GPT 5.4 and Gemini 3.1 Pro.
Sure, it may be better than Opus 4.6 on some of those, but barely achieves a small increase over GPT-5.4 on the ones I called out.
You are the only person with this take on hackernews, everyone else "this is a massive a jump". Fwiwi, the data you list shows the biggest jump I remember for mythos
Please look at the columns OTHER than Opus as well.
> Terminal-Bench 2.0: 82.0% / 65.4% / 75.1% / 68.5%
> USAMO: 97.6% / 42.3% / 95.2% / 74.4%
> The biggest jump in the numbers they quoted is 6%.
Just in the numbers you quoted, thats a 16.6% jump in terminal-bench and a 55.3% absolute increase in USAMO over their previous Opus 4.6 model.
If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.
It's higher than all other models except vs Gemini 3.1 Pro on MMMLU
MMMLU is generally thought to be maxed out - as it it might not be possible to score higher than those scores.
> Overall, they estimated that 6.5% of questions in MMLU contained an error, suggesting the maximum attainable score was significantly below 100%[1]
Other models get close on GPQA Diamond, but it wouldn't be surprising to anyone if the max possible on that was around the 95% the top models are scoring.
Because 100% is maximum, you should be looking at error rates instead. GPT has 25% on Terminal Bench and the new model has 18%, almost 1.4x reduction.
Failed to use tools, failed to follow instructions, and then went into deranged loop mode.
Essentially, it's where it was 1.5 years ago when I tried it the last time.
It's honestly unbelievable how Google managed to fail so miserably at this.
> Thought for 7.5 million years
> Rate limit reached
I wonder if misalignment correlates with higher scores.