Posted by mfiguiere 16 hours ago
it's similar to Claude code ultracode.
there is no ultra effort level implemented on the backend. it's just alias in the codex to max effort setting and single line addition to prompt to use subagents proactively. that's all
as far as we know pro models work differently. for once those are backend implementations and they probably run multiple parallel reasonings for any chunk and use some judgement model to pick best version as persistent one. but that's what I believe is most popular guess, because this is openai secret sauce.
there is still no way to use pro models from codex, or at leat so far there is no trace of it anywhere.
This misses an important detail. In Claude Code [1], ultracode suggests the agent create a JavaScript code to deterministically orchestrate sub agents. This is different from just having the main agent launch sub agents and (non-deterministically) manage them.
The resulting workflow is called “dynamic” because CC creates this orchestration script dynamically, “on the fly”.
[1] https://claude.com/blog/introducing-dynamic-workflows-in-cla...
Another useful thing about dynamic workflows is you can ask Claude to make them durable as skills (or slash command) that can be invoked later.
I believe inside Google they have a similar concept called “deterministic workflows”.
I find ultracode extremely useful. Of course you have to watch how your 5 hour and weekly session usage percentages are getting used. So I had Claude make a status-line with 3 progress bars: for context window, 5h session, 7d session:
https://pchalasani.github.io/claude-code-tools/tools/statusl...
It’s sad to see folks like Karpathy make a big deal about looping, than to find that the loop command is broken and it’s crap vibe coded documentation isn’t even accurate on the Claude docs.
This whole dynamic workflow idea is on face bad. It’s all done as a massive cope for the fact that real determinism (I.e using structured outputs to enforce control flow of tools deterministically) is bad for alignment/safety so they can’t let you have access to those tools anymore…
https://github.com/openai/codex/blob/98d28aab54ed86714901b66...
https://github.com/openai/codex/blob/98d28aab54ed86714901b66...
https://github.com/openai/codex/blob/98d28aab54ed86714901b66...
The way I see it is that they try to normalize and ease the use of practices established by the community.
I haven't really used it yet.
2 months ago management was showing us scoreboards, praising leaders who used most tokens. Last few weeks, we're getting weekly emails, telling us that whenever we can - we should use cheaper models, and that we should watch the page which shows our tokens usage.
But of course to some asshole non-technical people it meant asking for their vibe coded bullshit to be merged into production without review and fighting about it.
People’s job is to partially reign in this volatile environment by creating processes with stable output.
That is to say I do agree that building reliable processes out of unreliable parts with feedback is the modus operandi. However AI cannot meaningfully handle feedback and learn. And that is a key unsolved problem.
If the Model makes repeated mistakes on the same subject matter, you can update your agent.md file, or you can add skills to deal with specific prompts, or you provide a better default harness.
The whole idea of coding agents is their harness makes a big difference vs a pure raw model.
> However AI cannot meaningfully handle feedback and learn
How do you think models are created? They are trained on feedback and learn.
Its not cheap but you can post train models. This is how custom models are mode, that deal with specific tasks more efficiently and accurately.
Example ... Composer? Its base Kimi v2.5 model that has been post-trained 2 weeks, to create Composer 2.5, what is a much better coding model.
Its literally trained to make less mistakes by feeding it correct data. Hell, a lot of the models you are using, are often the same base model, where v2.0 was the initial released model but the model keeps training, so when they release v2.1, its still the same model, but with more training time on feedback provided to v2.0.
LLM Models are not a cake you cook one time and they are done, and you start from zero again. If you have the money, and a powerful server setup, you can take a model like GLM 5.2 and post-train it, to reduce specific errors. Sure, you need a ton of money because its a large model.
But people have been doing this with 5M, 100M, 1B, 5B models for a long time already. To the point that some of the small models can do specific tasks, almost or better then some of the huge more general trained models.
That's all just prompting.
> How do you think models are created? They are trained on feedback and learn.
No one is post training models on a single mistake. At least I have not seen it. I also doubt it is effective. Post-training on a single failure will not meaningfully change the model. That even sidesteps the entire problem that you don't even have access to models if you use a provider like anthropic/openai
And telling someone not to repeat a mistake is… ?
I feel like this line of thinking is kind of an unfair comparison. I'm not saying LLMs are magical beings that can suddenly learn by themselves after getting something wrong, but your "person making mistake then being corrected" assumes you do tell the person about the mistake and tell them to avoid doing the same mistake in the future, but for the "LLM making mistake" example you then intentionally avoid letting the prompt being changed in response to the mistake, which would be the "then being corrected" part on the LLM side of the comparison.
Similarly, if you just let a person make a mistake and don't let them know about the mistake, they might keep making that same mistake over and over again.
If you update how you use the LLM as you discover what mistakes it does, just like you'd correct a person, then you can use an LLM and also the LLM can "be taught to not make that mistake again".
> How many p are in strawperry?
> There are 0 “p”s in strawperry.
And I can trigger the same mistake with various words even when adjusting the prompt many times. So I cannot teach chatgpt to correctly count characters.
Well this is the central bet of AI coding isn't it? We, the humans-in-the-loop, get better at knowing ahead of time which patterns AI will handle better than others, all the while the models actually get better.
please take a look at the error(s) made in the prior run. what could've been done better? create or modify an existing skill to emphasize this, or suggest additional language in AGENTS.md.
We’re not at the point where AI is capable of knowing what went wrong and self-aware enough to understand how it could reliably change its own behavior.
For months I’ve been trying to have the agents stop manually writing our auto-generated SQL migrations and run the command that generates them instead. SOTA models insist on occasionally getting it wrong.
The whole "soul.md" stuff today is a poor approximation to that. But I wonder whether it will grow into it, like chain of thought prompting grew into reasoning models.
This is what I was referring to:
https://www.cnbc.com/2025/08/19/sam-altman-on-gpt-6-people-w...
It's marketing speak, but the goal is clearly there, no idea how achievable.
These models are great tools, but putting them and people on the same level does a disservice to our species and also is simply incorrect to what we know these models to be and their capabilities/limitations.
[0] https://www.theguardian.com/technology/2026/jun/01/meta-ai-h...
At the same time, one should acknowledge that not all tasks are on the same level.
I am using gpt to populate JSON objects conforming to a list of natural language constraints for purposes of generating fake customers. I am finding that gpt5+ never fucks up. Not even a little bit. I've ran this test hundreds of times with 20+ constraints and it's been perfect every time.
Stable information yields stable control flow. Humans are much more likely to forget one of the many constraints during testing. This happy mistake may incidentally cover an edge but it also means we lose coverage elsewhere.
I think whether or not the LLM should be allowed to directly author deterministic control flow (code) is mostly the same thing. If you have a lot of constraints you want to satisfy all at the same time, this can give you a hit very close to the ideal target very quickly. Not knowing exactly what you want is when the LLM takes you for a ride.
1) Writing code the "old fashioned" way (i.e., a Python program that does X, Y, Z) allows you to arrive at a battle tested solution that will not change over time. From a risk assessment perspective, the behavior is essentially immutable, allowing a business to guarantee consistent behavior over long periods of time.
2) Just because something hasn't happened to you do, does not mean that it will not happen. LLM are opaque. If you stay on the "happy path", you may see consistent behavior for long periods of time, but there's always potential for an edge case where something goes catastrophically wrong. This is without even opening the can of worms regarding prompt injection and intentional sabotage of a working system.
3) There are plenty of real world examples of an LLM spontaneously deleting data from a DB (or the entire DB) or otherwise going completely off the rails. These might seem hyperbolic, but it happened at our company (to a test DB, not production). The severity of errors that occur can be existential to a business' survival without the proper guard rails.
4) There's no concrete way to truly confirm understanding between an LLM and a human. It can tell you that it completely understands what you want, and then it can do exactly the opposite. Followed by, "my bad" (Claude's new favorite catch phrase). Code can be audited and even proven to be correct given the appropriate level of time and energy.
My best results have been gleaned in using LLM to produce deterministic systems. I recognize everyone has different use cases and needs, but this seems to be the best use of the technology in my experience.
LLMs are probabilistic by design so running the same prompt multiple times will give you different results.
Otherwise, we wouldn’t needed LLMs and could replace it all with Postgres
Reasonably deterministic is the phrase. If I can be sure the LLM is giving me back the same result 99% of the time I need it, that's reasonable for me. Maybe this is not reasonable for others.
ie How often will an LLM get 2+2 wrong? Now expand until you're uncomfortable.
[1] https://homepages.cwi.nl/~storm/teaching/reader/Dijkstra68.p...
We're no different to AI. The code we write to solve a particular problem can (and probably does) change from day to day, depending on your "mood", what you had for breakfast, if you've been fighting with your significant other, other problems/human emotions.
You're there for blame, not much else. The systems are still going to churn garbage, but because it's a business, that business will rather pay less for garbage they can sell than pay you a living wage, eventuially.
But your point stands: for critical business processes that need predictability, we indeed need determinism.
I shot down similar arguments in favor of outsourcing overseas for years. Outsourcing any critical logic to an LLM is even worse.
Outsourcing to another continent of humans and supplementing workflows with LLMs are entirely different operational universes. I think it is fair to put them on the same spectrum, but they're really far apart.
I'd argue outsourcing is a far more aggressive abdication of ownership of the technology than bringing an LLM agent in house and having it light a few fires under a few asses.
[edit] as the sibling points out, decayed system knowledge leads to relying on the LLM to fix the bugs the LLM introduced, which causes further decay in the institutional ability to reason with the business logic in code.
So the "knowledge base" an LLM generates is not useful as code, nor is it useful to humans. It may be useful to other LLMs as a lossy compression scheme for the original intention of the business logic and code.
That's not even in the same universe as institutional knowledge. Handing any serious business to something like that is malpractice, and sloppy beyond belief.
I'm not surprised that a lot of people think this way, because they never really grasped the benefit of holistic business knowledge united with code to begin with. Those people always outsourced, always got shitty code, and never really unified their systems. Cheap people and cheap companies take cheap exits. That's fine. But yes, many fragmented humans who all understand their portion in depth is much better than a bunch of markup files at retaining the knowledge of why things are done, procedurally, the way they are.
Once every six months the CEO calls me and asks me to remind him why our software does something like, idk, create a reverse payment instead of voiding a charge in some situation. Or some other thing he has asked a dozen times before. And I know the answer to why, or I know where to look, because I was in those meetings 5 years ago or because I wrote the code myself and asked the question when I wrote it.
Institutional knowledge is a very large context window, if you need to think of it that way, and LLMs are a shitty compression method for that. They can tell you how, but they don't reason well with why.
I think I know the reason to this one. Maybe. Because the money is already gone! There are no rollbacks in the banking system only "counter"-transactions.
I've, for better or for worse, jumped full-on in with LLMs, and I can already feel my abilities radically declining over time. That's fine by me, but if you're a company looking to protect your institutional advantage - just handing over everything to an LLM suddenly means anybody can do exactly what you're doing, literally. At best and this is an absolute best case scenario, you're transferring your advantage to the stewardship of another company.
The "if AI code breaks, who will fix it? -> "AI will fix it" exchange has (anecdotally) been very common among executives, which is much closer to outsourcing than programming.
Something completely different (but with the same logic): do you outsource legal, hire your own team of business lawyers or will you let customer services use AI for legal problems (and only hire a lawyer for a day in court)? I think all three solutions are currently active in different firms. From a risk perspective I would always want a lawyer on my team. Insource those learnings. But perspectives vary.
Moreover, any of us can run essential questions by each other, the marketing director, the lawyer, the CEO, the GM. And everyone has a stake. So all the creative and business decisions are worked out pretty quickly, and new features roll out with everything vetted and in place.
https://github.com/minimaxir/hacker-news-undocumented/blob/m...
It’s about pumping up revenues, earnings and eventually cash flows (since they can’t keep raising money constantly) to support a nonsensical valuation.
Tokens billed at API prices are profitable for openAI and anthropic today and it only get more lucrative every month for them as their inference costs fall. If it weren’t for continuous massive training runs taking larger and larger capex, these companies would be massively profitable
That's nice for them. What's less nice for them is that big customers are figuring out (some faster, some more slowly) that they don't want to pay API prices, or at least only for a much smaller number of tokens.
Like: What competence do decision makers in FINANCE have, when they are this oblivious to economics?
I don't think this is axiomatic. I usually invest my time, energy and money in things that go squarely against the market's momentary sentiment.
[no offense intended to the parent post, I'm on their side]
GPT-10 Human Ultra
The AI companies' salespeople must be the greatest geniuses in the history of the world.
Please stop using this brand new hammer for glassmaking.
GPT 5.5's double token cost was the threshold for me. These things are getting expensive quickly - the subsidized pricing can't go on forever.
Ymmv, do whatever you want . It's your life.
And I do more hand coding to guide the agent to useful patterns I want.
I feel like that’s how I used less capable agents a year ago. But I’m finding even with high quality agents, the slop creeps in.
I want more control. I want to save money. Hence going back to a more 80/20 agent to human LoC split.
Everyone is insane.
Just like many managers, the appearance of productivity is all that counts. And LLMs shine at giving the appearance of having solved all of the managers problems, and all they have to do to use it is spend on tokens.
This isn’t to say that LLMs aren’t truly useful, they absolutely are. But they’re very nature is one of simulating intelligence through next word prediction.
The chat modes and models are by their nature supremely attractive to management layers, because they give answers that sound so damn plausible even when they are complete fictions, and uttered with such confidence how could they be anything but the singularity.
> Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
https://openai.com/index/previewing-gpt-5-6-sol/
Can someone explain how this compares with Pro? I thought Pro was already something similar.
1. Managers: those who generally know what needs to be done, and want it done faster, so they provide a lot of instructions and context (where many developers fall)
2. Executives: those who vaguely know the end goal, but are clueless about the process, and are willing to burn resources and cycles on a black box to get the result
Same. I suspect they'll get better at taking in terrible prompts over time though... Maybe that's what Fable does better, reminds me of Sora 2, it would take my crappy prompt and expound upon it. I told it once to generate a video of someone working at some company that changed its name, but the old name had historic relevance, it referred to the new company name without me telling it to, by virtue of me wanting a video of TODAY with a 90s icon.
If your tasks are well defined and don't require a very large number of steps -- e.g. you're asking for small, clearly defined changes to the code -- you're fine with grok-4-fast. (Well, you would be fine if they hadn't killed it.)
I work in both of these modes, and I find that the latter actually benefits from dumber models, because smaller models are faster. The work shifts from async to realtime/interactive. So you can stay alert, keep track of what they're doing and iterate, instead of alt-tabbing, getting a coffee, and then spending extra time resynchronizing your mental model later.
It’s far more careful than opus and puts far more effort into testing and validating by default.
Switching back to opus at work was a downgrade. Similar requests felt more clunky and needed far more hand holding.
If they know much of what the tool is used for, they can customize prompts to "do that usage right" even if the user doesn't know exactly how to ask for it.
The responses I get from pro don't feel like ensembles. They are often very one directional.
> We generally treat GPT-5.5’s safety results as strong proxies for GPT-5.5 Pro, which is the same underlying model using a setting that makes use of parallel test time compute. As noted below, we separately evaluate GPT-5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.
https://deploymentsafety.openai.com/gpt-5-5/model-data-and-t...
There have been multiple podcasts with people from OpenAI which have confirmed this.
Any idea what that means exactly? I vaguely remember that ChatGPT Pro was originally called "deep thought", just like Geminis "deep thought" feature (or "deep think"?), so it seems likely they are using the same approach.
Its widely accepted[1] that it runs the same query through the model in parallel and then has a model that either selects the best answer or synthesizes an answer from the multiple ones generated.
I believe most people think it runs 6 sub-models, but I think that is based on the pricing.
It's a pity that OpenAI doesn't publish details like this.
Pro is quite limited on the web UI I reckon. This approach can be highly effective for reasonably verifiable task, for example, write comprehensive unit tests pointing out a tricky bug, get multiple agents to swarm at it.
It's unclear how they would do this when there is no signal that provides an objective ground truth.
const audits = await pipeline(found.files, file =>
agent(`Audit ${file} for missing authentication checks.`, { label: file }),
)
Apparently in the web chat (and also in Claude Code?[0] Though I haven't tried yet) they can call the Anthropic API and your subscription key gets auto-magicked into the requests somehow.
Those are two separate things of course (aside from the key-injection) but I guess there's no reason it couldn't run completely in the front-end... hmm...
To what effect I don’t know… I thought subagents were useful because they were explicitly single purpose and bound to a narrow context
If you don't mention it directly, it's 50/50 whether any given request will invoke a subagent.
The same with tools, skills, etc. No matter how smart these LLMs appear, they rarely do thinking as you expect.
So basically: learn how the harnesses operate, and know the names of the tools they have.
Hopefully, 5.6 will automatically spawn sub-agents without needing to ask.
https://www.theinformation.com/newsletters/ai-agenda/openai-...
"However, these inference optimizations, which rival Anthropic refers to as “compute multipliers,” are a big focus for all the labs. Anthropic CEO Dario Amodei has been publicly talking about the concept since at least mid-2023, when he said on a podcast that the company limits “the number of people who are aware of a given compute multiplier” because it could give other AI labs a leg up if they were to be able to replicate them. (Compute multipliers can also refer to efficiency optimizations in the model-training phase.)"
Yes, on a world with finite resources where your industry is singlehandedly siphoning ALL THE RESOURCES - hoard general efficiency optimizations and treat them as trade secrets - winning is all that matters, normal people and other species and the planet be damned.
Everything I hear about Dario these days makes me like him less and less. He sure did seem to speed run the 'tech leader with scruples' to 'tech villain' path! I guess all the cycles are compressing as we approach the singularity..
In what universe is any company going to give that advantage away?
In any case if they take away a lot of market share it's basically the same in the end - most people will be using these optimisations.
But at work I can only use the approved Enterprise Plans we have and we only have those with Anthropic and OpenAI.
OpenAI seems to be trading roles back with Anthropic becoming misanthropic. I hope they both start heading in the direction of how the AI field was prior to LLMs.
Collaboration and benefit for all should always be the primary motivator.
Of all the things to never happen, this is never going to happen the most.
That train left the station for good once hundreds of billions to trillions of dollars were involved.
On the bright side, in the long run I suspect the vast majority of the value of AI will not be captured by the model making labs and the vast investments in them are going to implode, so...
They have a staggering surplus of grid capacity and can bring more online without any difficulty. We couldn't get a serious nuclear project done if Jeffrey Epstein was offering private flights to the ribbon cutting.
In the United States at any given time more than half of the FLOPs are badly misallocated, Meta has like, a double digit percentage of the total capacity going down the drain every day and has for years. That's a conspicuous example but on OpenRouter rankings it's rare to see more than one or two American vendors in the top 10, sometimes the top 20. But 3rd, 4th, and 5th place are all merrily burning half the compute duplicating effort and missing key innovations because we stopped publishing real results. In China if DeepSeek makes a breakthrough it's at Zhupai and Moonshot and MiniMax and MiMo and Qwen that week.
Our only lever, export restrictions, seems to do nothing but breed multiply antibiotic resistant super hackers who just get more efficient and immediately propagate all of those efficiencies to the rest of the Chinese AI industry.
At the beginning of 2026 there was one Chinese lab with a model that had any real relevance fielding modern tool users. Today in July there are like, eight lagging the absolute frontier by maybe 3-6 months. Barring some massive bend in some curve 3-4 of the top 5 and 6-8 of the top 10 will be Chinese and open weight by January.
The great irony in all of this is that our current playbook is straight out of the 1960s USSR, and the PRC's current playbook is straight out of 1960s USA. We're the ones with the opaque decision making and gross resource misallocation driven by the personal agendas of a shadowy cabal of frenemies wired back channel into government in the form of the individuals rather than the offices. They're the ones with a thriving marketplace of ideas powered by robust public/private partnership and a paved path running bidirectionally to the university system.
It's going to implode because the Kruschev system does. Theirs is going to thrive because the Kennedy system puts a man on the moon before the decade is out.
There's no evidence of this, the parsimonious explanation is PRC AI, by virtue of being sanctioned, simply is not able to run magnitude more expensive compute model, and even if they could, they don't have the $$$ or market cap to do so. So they optimize and involute margins like they do in everything, and US misallocated expensive flops because the entire industry has been financially engineered for phat margins along the entire producer supply chain is just cherry on cake. Like wipe out the 50%+ margins from toolmakers, fabs, gpu/memory/data center components to some reasonable level and US is overpaying for tokens by a stupid multiplier on top of actual compute misallocation due to incompetent infra. Maybe PRC AI has unsound economics, but it's structurally simply not able to misallocate as much as US who will find a way to financialize compute to point of absurdity.
https://www.bloomberg.com/news/articles/2026-05-20/airbnb-s-...
I don't know how much business OpenRouter does in Europe (they have some GDPR text in some settings pages I think) but it's zero in China.
You might also consider the countless companies that do nothing but open weight LLM inference: Together (of Tri Dao fame), Fireworks (founded by my old colleague), Baseten, SambaNova (of Chis Re fame), and too many more to count, there's a new one every week. Plus NVIDIA does a ton of this business via NIM and shit.
I'm trying to be reasonably polite because it's possible you literally didn't know that, but this reads as a troll, so, if it is a troll please stop.
The administration could probably put some serious friction on open weight model use in the Fortune 500 for a little while, but the opposition never got such a gift right before a squeaker midterm. And outside of major enterprises with puckered ass compliance departments? Not a chance. It's popular around here to forget Uber and AirBnB and yes, OpenAI and Anthropic all got their start flagrantly breaking the law and grew lawyers and lobbyists faster than anyone could enforce it. And this time everyone from the DNC to the EFF would be holding hands wearing "Save The Models" t-shirts. Not even NVIDIA is remotely pretending they're anything but all in on GLM 5.2, they had an NVFP4 quant up by the time most people read the blog post.
And the Trump Administration isn't exactly enamored of Comrade Amodei at the moment, being as they're appealing the lawsuit Anthropic brought against the Pentagon during a shooting war.
Forcing the American proprietary AI megalab financing event was our fiscal Ukraine Special Military Operation, the market is calling the bluff and neither the capital markets nor the Federal Reserve has the dry powder to absorb this one.
The Treasury auctions will flat not clear in an orderly way. We can't raise 2-4 trillion dollars on a dime in 2026 and if CoreWeave turns out, as many suspect, to be Patient Zero? It would be that big a hole.
We play by the same rules as everyone else now. I hope we regard it as being worth it, but I fear we will not.
On one hand it appears to cooperate with OpenAI and Anthropic, as big customers.
On the other hand NVIDIA cooperates with Palantir, providing the HW for its "Sovereign AI OS" (a turnkey system including HW and SW for local inference and post-training/fine tuning) which uses the slogan "The future of AI is on-prem" (i.e. not as a customer of OpenAI or Anthropic, but using an open-weights LLM, e.g. a fine-tuned NVIDIA Nemotron or a Chinese LLM).
Presumably with the goal of promoting their competing solution, Alex Karp (Palantir CEO) has delivered a few weeks ago a very harsh criticism of Anthropic and OpenAI (who allegedly inflate the token consumption and they might also steal the data of their customers, which must be sent to them).
So NVIDIA both cooperates and competes with OpenAI and Anthropic.
Once China starts to get scary, Commerce will export control GPUs and declare Chinese models "foreign munitions." Any nation doing business with the US will not be allowed to use these models either, and that will be the end of that.
It is just not in the US's interests to fund China in the race to AGI.
Peter Hegseth, another really pro-America being powerful guy, he's dealing with a lawsuit because he doesn't want Anthropic in his military, he calls it a supply chain risk (he's right).
There is no evidence of any kind that a complex attack vector can be trained into model weights and survive all the crazy slicing and dicing that happens between published weights and running model. These things get quantized and run on mathematically imprecise kernels and sampled and LoRA-tuned and Dolphin/Orca de-tuned. Go look at what the ComfyUI community comes up with, those guys know more about WAN 2.2 than the people who trained it. Because those models run for real on a desktop, so there's mad innovation at light speed.
There is no one who wants a capriciously expensive black box run by extremely creepy people, not once the capability crosses over (in about November).
But don't take my word for it, you just had a chance at one AI IPO, and I'm sure you'll get another, so if you like how that goes, you don't need to convince me!
As well, it is a false equivalence to say that local models are only Chinese and otherwise we would use cloud models, but there are American or European ones, so a ban would simply force companies to use these, even if they are inferior to Chinese ones. It's simply a matter of national security to the US government, and they will not care what random people in media say.
I don't know what part of this you guys are having trouble with, but it doesn't get a whole hell of a lot more emphatic on "what the US Government thinks" than who the Pentagon is in court with to avoid doing business with that party.
Mr. Hegseth is the representative of the administration in the AI usage policy of the largest bureaucratic organization in the history of civilization. He has emphatically rejected at least one black box American AI megalab and President Trump has endorsed this action on Truth Social.
The government's stated position is not what you among other commenters are stating or implying as the government's position.
We are in a race to superintelligence. The first country to AGI will be the first to superintelligence, and the first to superintelligence will have de facto control over the world and the future of humanity. They will also be able to prevent others from reaching superintelligence.
Of course it's in the US's best interest to slow down China. You aren't zooming out and looking at the big picture, you're taking models as slightly useful tool, not what they will soon turn into.
There is nothing to support this. You get cheap Deepseek tokens by foreign providers too.
It is the same thing with automakers. They complain about not being able to only make luxuary cars with high profits with BYD raining on their parade and blaming the Chinese gov.
OpenAI tried to pull off the same trade secret thing with RL when they announced o1 and o3, aka "Compute time scaling". Then Deepseek revealed it with Deepseek R1.
Could also be something like Deepseek DSpark. Or using diffusion like DiffusionGemma as a draft model. The timing between the release of those, and this article, makes me think its maybe one or both of those things
This 2023 thread about this issue is prescient: https://old.reddit.com/r/MachineLearning/comments/11sboh1/d_... (just add Anthropic to OpenAI)
How is this suddenly evidence of him being a villain?
I don’t think this makes Anthropic a villain?
Compute multipliers are like a quant firm's trading algorithms. They're the crown jewels, the whole alpha of the lab. If you leak them, the lab dies.
Protecting them does not make Dario a villain, it's literally his job. It's also Sam's job, Denis's job, Mira's job, etc. Every lab guards these multipliers closely because they represent the entire worth of the lab.
No wonder you're confused about DeepSeek when you have a fairly obvious explanation provided to you, and your response is "it's unrealistic to think the Chinese Communist Party is behaving like communists."
> if that were true, why don't they have just one model initiative instead of several?
Because value exists at several layers of the product hierarchy? I.e. for the exact same reason that the for-profit labs don't have just one model initiative?
I wonder if that makes sense if the orgs within the industry are starting to shift their mindset towards "Tokens are expensive, we should use AI less." which feels like an existential threat to the status quo, if those AI providers can't find ways to keep costs affordable for their clients. Otherwise those orgs would just be using GLM 5.2 or DeepSeek V4 Pro but it seems like what they're doing instead is trying to use AI just less, period.
OP phrases it as a bad thing that Dario is keeping compute multipliers to Anthropic. How naive can one be? Compute multipliers are the whole business. Those are the trade secrets every lab is built on. It is the alpha of the business. How does protecting this make Dario evil?
This website is getting out of hand with the uninformed hot takes. I wish when HN was still people that knew what they were talking about.
People have different opinions than you, it happens..
[0] @sama if you're reading this we can fix that...
AI is a bid for control of all humanity. The first to superintelligence owns our future. I'd say it's okay to be a bit suspicious or conspiratorial of anti-AI/anti-lab narratives.
(BTW Anthropic only exists because Sam Altman is a liar, Dario admitted this.)
Except for, you know, all the outside investors and the forthcoming IPO.
Related: https://80000hours.org/2012/03/the-replaceability-effect-wor...
There's a more nuanced discussion that could be had about how to balance relevance with outside influence. But at a foundational level it should be acknowledged that the tradeoff exists, and that receiving outside investment can't alone be seen as evidence of corruption.
Besides that, there's more that can be said about other things like their corporate structure or the degree to which they accelerated the AI race.
Of course that's what Dario thinks because that's what every tech CEO thinks. Dario, Sam, Sundar, probably many Chinese CEOs as well. It's what everyone thinks. That's why they're competing so fiercely with one another. That's why they basically make all the same decisions. That's why we need properly open source AI.
This doesn't seem like the right place to spend my time litigating that point to its fullest extent (no-one here is doing that). But there's plenty of relevant info surrounding eg.:
* The New Yorker article on Altman [1]
* The story behind Anthropic's founding
* Various efforts to influence government policy (a16z policies and contributors [2], Trump's inauguration donors [3], giving Trump credit for AI infrastructure [4], Dario's op-eds [5])
1: https://www.newyorker.com/magazine/2026/04/13/sam-altman-may...
2: https://a16z.com/portfolio/
3: https://www.opensecrets.org/trump/2025-inauguration-donors
The day Mythos class models are open sourced will not be a good day. I don't think you understand the impact that will have on the world and on cyber defenders everywhere. It will be pure chaos.
Even if you don't think Mythos-class is the bar, open source has to stop at some point, you don't hand everyone a superweapon.
Handing every skiddie and nation state and APT and hacker group access to Mythos does not help cyber defenders
Even if you don't think Mythos is a big deal: At a certain point models become smart enough as to be dangerous, and you don't give everyone a superweapon. Open source has an end of the line sooner or later.
What kind of rosy-eyed chump believes in the "tech leader with scruples" bullshit? It always lies.
Did some people just ignore Mark Zuckerberg and Tim Cook's sociopathy, somehow? Did anyone buy into their "privacy is a human right" nonsense?
The thing I can't quite square is that it doesn't really fit my lived experience. I have known sincere, genuine people in the types of positions that I'm sure someone like you would declare to be sociopathic.
But beyond that, I just don't know why it would actually be true that everyone at the top is a villain. Why couldn't someone like Dario (or even Altman, gasp) be sincere? Because if he is, it does seem like a lot of the moves he's made would make sense given his worldview.
But if you assume he's just a villain, then you can twist any of those moves to just be further evidence of that which you already believe.
I don't know, I just find cynicism interesting, and a little sad.
You don't have to assume anything. A true "good guy" doesn't openly say that he's fine with autonomous, AI-powered weapons being used against me, and mass surveillance applied to me and my family just because I don't live in the US. A true "good guy" doesn't say "privacy is a human right", and then immediately (and completely) bend the knee to an authoritarian government on this issue.
And about the mass surveillance, I don't see why the military should not use AI to do surveillance abroad.
If you are dropping bombs on someone I'm unconvinced the use of AI will make them like you more or less.
I remember a long time ago it came out that the US had been doing mass spying on the Danish people, my dad was very upset about it and disliked the US for the rest of his life. Of course the only thing he did about it was not watch American movies anymore or visit the US.
Anyway, I assume it will be a case of a million little paper cuts, each thing putting off a group of people until someday it adds up to real meaningful economic impact.
I went in the opposite direction - how far can I push myself to see multiple facets of a story? That is a wild ride, and it gets progressively more wild.
Please, I'm dying to hear the optimist's take on Mark Zuckerberg's career. It wouldn't happen to be embarassingly foolish, would it?
Lots of nerds for some reason have made cynicism a personality trait. They think optimism/honesty is hopelessly naive, therefor cynicism is the correct default.
It is the result of experience. Working with and creating systems (even embarrassingly simple ones), then seeing them fail in a myriad of ways more often than succeeding, colors your expectations about throwing humans into the mix.
Children learn to lie as part of their natural development, but do not always externalize that until faced with media (Airheads candy commercial or equivalent). Either way, honesty is expected as a default for utility and not an expectation in leveraging goals.
All have collaborated with the current US regime. All have shown signs of being quite willing to compromise their principles in order to make money.
History.
Also, nobody said 'everyone' or 'villain'. How Paul Graham of you.
Power corrupts, absolute power corrupts absolutely.
Given enough money and increasingly perverse incentives to gain even more has a very high potential to corrupt.
Did they start out as corrupt, or was it the influence of the power that came with the obscene amount of money?
It's really a chicken and egg level of calculus.
Doesn't matter which came first, either way you get feathers and chickenshit all over the yard.
Do some very rich people still seem very nice in person? Sure. Of course they must, because otherwise no one would willingly work for or with them. As the total amount of money goes up, the incentives to remain 'seemingly nice' go down and either you get to see who they really are, or who they became through the choices to make that much more money. Doesn't matter which is true.
The examples of non-villainous billionare are rare.
Of non-villainous multi-billionaire; lets see there's about eight of them that stand out for giving significant amount of the massive wealth to helping the world around them, who live normal lives like the people in the communities where they reside, and who participate at the companies they own by eating in the company cafeteria among the people who earn the wealth they enjoy.
Thats 8/3400 global billionaires ... about a quarter of a percent.
And of the 'pledges' like Giving Pledge by the billionaire class, the actual amount delivered - not parked in a family or private trusts for tax deductions; but actually delivered to the front lines of any global crisis amounts to 0.18%, less than one fifth of one percent of the $20.1 TRILLION dollars held by that class of owners. thats less than $2.00 on every $1000.
That's not to say that donating to public needs is 1:1 for non-heinous behavior, but it seems like a basic tool for distinction. the 'can make a significant difference in global suffering : chooses not to' ratio as a surrogate for villain may be useful metric and doesn't require cynicism as the underlying rationale for calling someone's behavior as unkind in general or mean in particular.
Why should I treat Sam and Dario with special white gloves? Are they different, this time? They have peers in China that do the same research and actually release it to the public. They let you run the production weights on your own machine. Am I a cynic, for comparing these CEOs to their populist superiors? Am I stupid for assuming their hostility when they refuse to give us the benefit of the doubt?
I'll believe their actual altruism when I see it. Both are seeped in "boy genius" puffery and lie out their ass. If this is the future of intelligent innovation, then America is truly declining.
This is not hard to understand. Do you really think DeepSeek would publish their algorithms if they led the American companies? Lmao.
Even when they're reaching parity with American models, Z.ai, Qwen and Deepseek are upholding their end of the bargain. I'd criticize them too, if they were due any scrutiny.
It is a great model for the price, but it has much worse autonomy and long context performance, and I don't trust it for anything beyond personal projects, whereas I use Opus/Fable and GPT for work.
The primary effective altruism cause areas are extremely acute and high-scale problems like malaria, vaccine distribution, and factory farming
For example, if inference isn't too expensive, but they figure out how to cut costs, then price goes down. After all, why pay OpenAI when a smaller datacenter can give you similar models?
But, if they make a huge issue about how inference is too expensive, they engineer a crisis of their own creation - then, once they deploy the solution (which they might already have), then they're back on top.
[1]: https://www.reddit.com/r/OpenAI/comments/1stsxvc/new_feature...
What you think could be a big chunk, is more likely to be a fraction of a percent of queries.
And what use is similar query caching - so you (very often! if actually cost effective, maybe half the time) get a response to a query that was different from yours. Including for when you have a lot of context input already. You’re going to get trash.
If it were constrained to only very common initial prompts, and somehow the long tail did not actually dominate as it does with Google search (can't find the reference at the moment but it was a famous article some years ago), it also wouldn't account for serious enough cost savings. Long context is what is expensive.
This might only work in constrained domains like customer service where there’s tolerance for generic answers and escalation paths. For technical work? For general purpose use, with secretly canned responses charged at full price?
By contrast, when coding, devs typically have hundreds of thousands of tokens in the context window, and may use many millions of input tokens per day.
Caching requires the full prefix to match exactly. If a single word differs near the beginning of the prompt, nothing after that can share the cache. So this type of caching would save a few queries that cost virtually nothing, but wouldn't help with the stuff where cost matters.
The only optimization that makes sense is per user prefix caching, because you are often sending the same system prompt over and over again or are continuing a conversation.
“ Automatic Prefix Caching (APC in short) caches the KV cache of existing queries, so that a new query can directly reuse the KV cache if it shares the same prefix with one of the existing queries, allowing the new query to skip the computation of the shared part.”
https://docs.vllm.ai/en/latest/features/automatic_prefix_cac...
The technique you linked only makes a substantial difference for particular use cases where you are going to have many LONG CONTEXT queries with the same prefix. For instance, when having a set of documents that commonly get loaded in as context. It's a way for application developers to keep prefixes they manage (or prefixes managed by some set of their users) cached. It has no relevance for long tail general purpose use.
The transform script(s) are cached and can be played back or adjusted. Surely for some breadth of question inputs, they map more often to similar answers--but not static answers; instead, evented edits.
It's nearly untenable for a human to keep private edit scripts to generate code changes. The extra steps for custom regex, essentially one-offs for a shared codebase, is inefficient. But maybe not to an LLM.
I don't think so, I'm on a ~$200 subscription (guessing that counts as way beyond "base plan") and have no pay-as-you-go pricing at all, using the OpenAI APIs would be way too expensive for me.
You can't use sub pricing with orgs w/ OpenAI.
No context, nothing, just a title and a random link to a tweet, which has a seemingly relevant response from someone who works at OpenAI I guess.
How does it compare to GPT-5.5?