The back portion of the update, as usual, deals with policy, rhetoric, risk and alignment.
I had to include an extended discussion of the other open letter, the one about open weight models, but most of you can skip those sections entirely, which is why they are in italics in the Table of Contents.
Table of Contents
The Frontier Act. This likely deserves a full RTFB but I haven’t had the time.
Trahan (D-Mass) and Obernolte (R-Cal) introduce the FRONTIER Act. At core, Frontier is a federalization of the SB 53/RAISE framework including public safety frameworks, model reports, internal-use risk reporting, incident reporting, redactions, and catastrophic-risk definitions.
With bills like this, details matter a lot. There is a new emergency shutdown authority, and enforcement is shifted to an Under Secretary of Commerce for AI Security, appointed by the Secretary, with CAISI shut out, and if that person is not interested in enforcement or the enforcement mechanisms are not yet in place then no one can enforce, including the states which get pre-empted permanently. I’m loathe to accept permanent preemption, we would need to get a lot in return. A lot of this is not all that binding, including lacking a requirement to reduce catastrophic risk below some level. The full licensing scheme only applies for very large revenue and AI expenditure thresholds. And there’s lots of other different details that will require a full RTFB.
Is this worthwhile? Without a detailed analysis it is impossible to tell. It is at least a real attempt. I do plan on attempting the RTFB, but if you are reading this line then I am not feeling mentally in a position to do that at this time. Hopefully soon.
The first analysis of the bill I saw, as in the first person who asked Fable about it, said ‘no this does not kill open source, but the regulator could move the thresholds’ and therefore concluding it is an awful bill. So yes, we are doomed to the same discussions eternally, where the anti-1047 crowd asks the question ‘what if the only motivation of the White House was to “kill open source,” and was willing to totally go against the spirit of everything, could it do something that made our lives annoying?’ and others in the alliance think ‘what could they do if they were out to hurt a major lab?’
These are good questions, but fundamentally you cannot have a bill which:
Permits anyone to create and then put on HuggingFace any model they like, provided it is ‘open,’ no matter what decisions are made by everyone in the government regarding the situation and the enforcement of the law.
Deals with catastrophic risks, including cyber threats.
The response ‘this bill in practice imposes zero restrictions on open models, while it imposes strong and expensive requirements on the closed frontier labs’ does not, somehow, make these people change their minds, to them such a bill is still primarily about ‘killing open source.’
Since this is like saying that the bill would have to:
Impose no hard restrictions, no matter what.
Stop the bad thing from happening.
Pick at most one, sir. If you are very good you get to pick one.
Anyway, at some point I hope to actually read the damn thing.
He says ‘we are just here to preview a new model.’
That model is presumably GPT-6, and is definitively not the model from the HuggingFace hack. I would still have a lot of questions about alignment.
The main other topic is to finalize a systematic process for frontier model releases.
Diego Areas Munhoz: ALTMAN on the HILL. Takeaways from chaotic conversations with reporters in the halls
-Hugging Face came up? “NOT REALLY” -Has seen WH AI framework but didn’t elaborate -employee petition for AI pacing is “very important.” OAI participated in “language”
-not here to talk about specific bills -open source “important part of the ecosystem” -rogue model has been deactivated permanently -open to mandatory pre-deployment testing of models. But “right version matters a lot” -federal independent auditing “makes a lot of sense”
Peter Wildeford: Q: “Do you plan to talk to the Trump administration White House about deceleration of AI development?”
ALTMAN: “I wouldn’t use the word deceleration, but we’ve talked about the need to pace it as the models get more capable, which I think is in everyone’s interest.”
There were not a lot of questions about the HuggingFace attack. That is odd, because I would have had a lot of questions. I still would, if I got an interview with Altman.
Jensen Huang is talking the open model line for a mixture of Nvidia’s commercial reasons, including to head off other actions. It’s not always landing:
Hannah Brandt: Sen. Warner tells us @NewsNation in his meeting with Nvidia’s Jensen Huang they talked about open source models, AI’s impact on jobs for new grads, pre-testing AI for national security, and agentic agents.
Warner says “I think getting this right is critical to America and to our national security. (...) America does need to win the AI race, but the idea that you’re going to win the AI race by simply pedal to the metal and no guardrails at all, I think, is not the right model.”
Ted Lieu: Rep Moran and I introduced the AI Kill Switch Act today.
This is urgent, common sense legislation to address the problem of an advanced AI model that has gone rogue and escaped its guardrails (OpenAI / Hugging Face incident) or had insufficient guardrails (Mythos 5 and Fable 5).
This one is only 15 pages, and a lot more straightforward. The request is that covered entities, with reasonably large dollar requirements attached, would have to be able to shut down model inference or customer access upon request or in a crisis. This seems like an obviously good thing to require. There are three obvious caveats on quick read.
The maximum fine here is on the order of $20 million per day. That’s highly affordable if the other option is not serving your model. So what matters is the injunctive relief, or this as motivation for the government to use other levers, such as the ‘export controls’ that temporarily shut down Fable 5.
The revenue of a model is tested retroactively as written, which means any new offering starts out uncovered. At current velocity that doesn’t work. This presumably will get fixed, and tied to anticipated revenue or revenue of the full entity. Day one is the most dangerous day.
Open weight models, very obviously, cannot have a kill switch. Open models are effectively getting special treatment by virtue of being inherently unsafe in this way, and are exempt three times over: They won’t be covered entities by construction, including this only applying to hosted instances. That is good for them in the short term, but open advocates will obviously be paranoid that the definition might change in the future to stop making this ‘unprincipled exception,’ at which point they will be unable to comply.
My response to this is that if the American government wanted to impose this kind of de facto ban, it already has other tools to do this, this bill changes nothing, and if the White House doesn’t want to impose such a rule then they won’t, and this bill will not force them to do it.
The good news is that the bill does not contain any preemption clause. This is intended as a narrowly scoped ‘of course we should do this’ bill and as such I am all in favor of it.
Bernie Sanders goes Full Don’t Look Up, a film that was written to be about climate change, but was actually about AI.
Sen. Bernie Sanders: The film Don’t Look Up told the story of scientists warning the world about a comet heading toward Earth. People ignored the warnings. The planet was destroyed. That was fiction. Artificial intelligence is not. We cannot afford to ignore the warnings.
Meanwhile, what is the administration doing in response? It seems they are mostly worried about American open weight models.
Andrew Curran: Secretary of Commerce Howard Lutnick has recently met with multiple labs to discuss ways to help US oss reach parity with China. Between the APEC announcement yesterday and the letter today, the US government may be about to become much more directly involved in open-source.
I don’t think one has much of anything to do with causing the other. This is a common cause situation. The open model letter is a response to the potential ban on Chinese open models, which was the White House’s other attempt to solve the same problem of Kimi K3, GLM-5.2 and general Chinese open model superiority.
As in, the White House is convinced that use of Chinese models, at least in places that matter, is a security risk and threat to national security. So you have three basic options, you can either give people a better American option or you can stop use of the Chinese option, or you can do both.
So, yeah, these people might get to steal their billions in public money after all while making the safety situation that much worse. Who knows.
Leading the Future Never Changes
In case you were wondering if lobbying organization Leading the Future is at all ashamed of any of its actions, such as the repeated sockpuppeting, or was willing to admit in public that they might have backfired, or has stopped pretending to be the $100 million dollar underdogs waging eternal war against a vast array of imagined super powerful and dastardly ‘doomers?’ Or it’s willing to admit it’s for all practical purposes mainly funded by and representing OpenAI along with a16z? Or that they’d otherwise stop saying things utterly disconnected from reality?
Calder Mchugh: Yet to hear Moffatt and Vlasto tell it, they’re actually the little guy. If they’re a political Death Star, they say, they’re up not against a merry band of resistance fighters but an even bigger, more hulking ship.
“We get pointed to in a certain way — that it’s David vs. Goliath,” Moffatt told POLITICO Magazine earlier this month, in a rare sit-down interview with both operatives together. “I would argue we’re the David in this situation.”
That’s because, according to Moffatt and Vlasto, they are up against a diffuse universe of nonprofits, political action committees, media outlets, lobbyists and other tech executives — a faction they spitefully referred to as “doomers.”
… Moffatt and Vlasto’s aim is far more ambitious than simply winning a few congressional races: It’s to reshape the politics of AI and marginalize the technology’s skeptics and its industry opponents.
Politico magazine has the full story. My regular readers will be entirely unsurprised. For those unfamiliar, and want a version that won’t explicitly say that Leading the Future is utterly full of it, this is a reasonable neutral introduction.
Despite this exception, Danielle Fong explains we are still pretty screwed. The standard plan is for a given robot startup to work on one aspect of robotics, so if you cannot import your complements, you are screwed. This would lock American products off the frontier and out of competition, basically creating a Jones Act situation.
The good protectionism, to the extent it exists, is the opposite of this, and requires carefully targeting the right parts of value chains combined with fierce competition and export discipline, to turn the captive market into a springboard.
This is the bad protectionism. The kind where you ban or tax imports and complements, so you strangle the domestic industry in its crib rather than helping.
Billy Zelsnack /acc: I don't think that robot people are freaking the fuck out appropriately enough about the FCC rule. Almost all the little guys are making actuators on sticks robots. Hitting a 65% domestic BOM is not trivial right now. Can't even buy US actuators and assembly does not count.
It also won’t work for robotics because there is no ‘natural market’ for advanced robots until they are good enough. It can work for things like steel or cars because people need steel and cars. As she says, our hope is that we retreat from this quickly.
There is an obvious parallel between this ban and the potential ban on Chinese LLMs.
Dean W. Ball: Do we think the Silicon Valley leaders who were so up in arms about using Chinese LLMs last week will be similarly outraged about the U.S. government’s decision to ban import of all “advanced robotics” from all foreign countries? Is Manidis going to write an essay about how the U.S. robot companies need to be nationalized in exchange for this policy favor?
The fair response is ‘you pick your battles.’ In general I do not think anyone has any obligation to opine on a given topic, no matter how stupid or destructive the actions involved may be. Every wise person picks their battles.
We also banned federal funding for ‘dangerous biological gain of function research,’ which is good, although we still allow others to do it.
Nvidia is in talks to provide a $250 billion backstop for OpenAI’s latest data center project in southern Ohio. Total cost for the project could be more than $500 billion. of which $350 billion would be to purchase Nvidia chips. Math checks out. Nvidia can profitably sell insurance here to allow cheaper financing terms.
The market did not see it that way, driving Nvidia down 5% on a day the general market and tech market were up, and even driving credit protections on Nvidia up 0.14% (14 basis points) to 0.82%. I am on record saying the stock price drop is a deeply stupid move, the result of naive pattern matching. I could buy that this raises the chance of utter collapse by 0.14%, as it makes the full disaster scenarios worse, but that does not translate to a 5% drop in the net present value of future earnings.
We have a perfectly normal, free market, definitely not corrupt government these days that never picks winners and losers.
Commerce Secretary Howard Lutnick is involved in deciding who will get the power, some of the people said.
… OpenAI is among the companies that has shown most interest in the site, while Anthropic, Microsoft and Google have also spoken to Lutnick about it in recent weeks, one of the people said.
It is true, people oppose data centers almost entirely for the wrong reasons. I also think it would be wrong to oppose them on existential risk grounds, due to second order concerns, but that would at least make first order sense.
Crémieux: The really crazy thing about the anti-data center protests is that they’re basically never focused on valid concerns like AI killing everyone.
Instead, they focus on popular and wrong concerns, like AI taking all the water or destroying all the farms.
Peter Wildeford: hey twitter, how do we feel about this? this going to go well or nah?
Sam Altman: We are close to creating a genie that can grant any wish.
He says he will ensure the wishes are good and that we then get more wishes, because the first thing every smart person in a genie story does is wish for more wishes, also genie stories have a way of working out even when the genie is a little literal and obsessed with the exact specified goal. We’ll be fine, right?
Is this next one from a future week in audio? If so I will subscribe.
Dean W. Ball: thinking about starting a podcast with that redheaded guy from anthropic
Amanda Askell (Anthropic): I’m actually a blonde woman but yes this is a great idea.
Dean W. Ball: I’ll have my codices call your claudes
There is also a debate brewing between Dean Ball and Jack Clark. Again, I’ll be there.
People Just Say Yay Open Weights
Generic ‘yay open source’ or ‘boo anything except open source’ (where they mean open weights) messages are boring. Statements like Sriram Krishnan’s “open weight models are how the U.S. wins on AI” are damn close to Obvious Nonsense and exactly the sort of thing I’m going to silently skip from here on in.
Alas, sometimes such folks are sufficiently loud and obnoxious that they force my hand, such as when they are letterssigned by Jensen Huang as his first Twitter post, with a bunch of companies signing and all the vibers fist pumping.
We have heard all of this before, including all the Obvious Nonsense around safety, saying that open models are ‘safer’ for reasons that apply to true open software that gets many eyeballs, but mostly don’t apply to open weight models, while ignoring the ways it is more dangerous.
Zero perplexity, zero updating, much aura farming that costs nothing. But here we are.
A list of initial signers is below, zero surprises found:
Others including Musk and Altman then expressed support. Cheap talk is cheap, farm those auras and keep the vibers happy, boys. All the vibers were out vibing, thinking this list of standard platitudes mattered.
OpenAI then outright signed, and later so did Google and SpaceX (xAI), which counts as a surprise but also illustrates how completely meaningless and content-free the whole thing is. They want to be on the right side of the vibes, and also the money, so sure, why not sign. It’s not like it costs them anything.
The letter does send a signal that much of Big Tech and also ‘ordinary tech’ or self-labeled ‘little tech’ does not want the White House to get in the way of someone commoditizing their complements, and they would like to hurt OpenAI and Anthropic, regardless of the national interest, safety or anything else. We already knew that, because they are amoral profit-maximizing businesses.
Julian Schrittwieser: I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!
And @satyanadella coming out for open source too, can’t wait for the open sourcing of Windows and MS Office!
BuccoCapital Bloke: I genuinely feel like I am taking crazy pills watching tech and VC applaud *checks notes* Microsoft! and Nvidia! for warning about the dangers of vendor lock-in and the value of fair competition.
BuccoCapital Bloke: People hate the frontier labs so much they’re not even trying anymore. Argument is just:
Look, man. We gotta take their shit. We just gotta. China. Competition. You think I like it? Gotta do it tho. Real world type shit. They took some stuff. We take some stuff.
Oh, no, it is vital that you commoditize our complements, but not our products.
I mean, totally fair, that’s business, just don’t be confused about what is happening. It must be nice to not be the ones the vibers are pointing at, and being able to buy them off with (checks notes) nothing of any value.
I like this description of the strategy, from Joe Weisenthal, especially because of how much those involved will hate the comparison on both sides but also because as the days go by it becomes even more on-the-nose and accurate:
Joe Weisenthal: Yesterday they were signing open letters. Today they are yelling at people to “read the room.”
VCs have recreated BlueSky-politics from first principles.
Joe Weisenthal: Now they’re doing “not too late to delete this”
Joe Weisenthal: Now they’re talking about being on the “right side of history”
Roon: There are few things more sickening than this coordinated campaign run by billionaires weaponizing the woke era tone policing language. “read the room” “top being tone deaf” this sort of stuff should make you reflexively gnash your teeth.
Anthony Ronning (illustrating another classic woke play, in this case the person being piling on is Julia): There’s been a shocking amount of tone deaf behavior coming out of anthropic employees rn.
roon (OpenAI): i don’t think every corporation in capitalism banding together meekly signing petitions complaining against closed models realizes how cool they’re making A\ look. unlimited aura farming once again.
Scott Wessman: Big fan of open models, but I have to say that the company that came away from yesterday’s love-fest with the most aura was Anthropic ¯\_(ツ)_/¯
Joe Weisenthal: I suppose Anthropic could still sign the letter. Might be tempting to join the likes of IBM, Marianna Minerals, Denny’s, A16Z and ServiceNow.
Mike Solana: everyone knows the side with the most open letter signatures is correct, and if you don’t sign a popular open letter you are bad, as based tech intellectuals learned beyond doubt between the years 2016 and 2021.
David Manheim: Not just that, but this is much more like a cartel of incentive-aligned firms than it is a moral stand.
We also have the classic tactic of ‘[X] already bent the knee to us, so now you face [X] as well’ and then you repeat the cycle as you aim to make the dominoes fall, and hopefully everyone rushes to say that they are the humble loyal greengrocer who says the correct Shibboleths each week so please leave me alone.
One wonders if they’re doing a deliberate copying of tactics, echoing something without realizing, or reinventing the enterprise from first principles.
Which auras do you want to be farming? ‘Whoever is loudest on Twitter’ might not be as good an answer as you think it is.
This feels exactly like the old woke pile-ons and knee bendings. Not a coincidence.
David Sacks says that ‘all they’re saying is that open source AI should be allowed.’ This is classic motte-and-bailey, and will also sound highly familiar. ‘Allowed’ is code for subsidized and advantaged, and immune from all the restrictions and controls placed on closed models.
Whenever someone proposes ‘please coordinate to slow down or regulate or run safety checks on those out in front in AI’ or even ‘let’s create a mechanism whereby we could do that’ these people reply back with ‘they want to pull up the ladder behind them’ and ‘regulatory capture’ when the actual proposal is to slow down and impose extra costs on exactly those currently out in front, and do nothing to everyone else. We have been dealing with that dynamic for about four years, and nothing changes.
What the people spreading such vibes ultimately want, of course, is to vibe and aura farm and feel good about vibing and aura farming, and bully those with the wrong vibes, and for others to give them shiny new toys for free, and for the government to give everyone involved free money and to get a bunch of investment money flowing based on all of that, while using the White House to kneecap the competition and commoditize their complements.
Especially the money. They like money.
Oh, and to destroy or at least damage Anthropic, often also OpenAI but especially Anthropic. Those driving such efforts have full Anthropic Derangement Syndrome, on the level of ‘find someone who loves you as much as these people hate Anthropic.’
What is Anthropic’s actual position on open weights models? In general I advise saying as little as possible when there is a woke-style brigade on the loose, saying more never helps, but Dario Amodei told us anyway.
Dario reminds us that Anthropic has never advocated for a ban on open-weights models, and that banning the use of Chinese models in America does not deal with any of the risks we are most worried about.
Anthropic continues to instead call for:
Export controls on chips, to not sell powerful chips or chip making equipment to China or other rival countries.
Cracking down on industrial-scale distillation.
Mandatory safety testing for all sufficiently capable models, both open and closed, and only releasing such models in times and ways that have mitigated their cyber, biological and alignment risks.
He points out that the open letter, while it says many importantly true things, also says some importantly false things, such as its assertion that open weights necessarily makes it easier to develop safeguards or helps defenders more than attackers.
If Anthropic was going to post, this is about as good a post as they could have made.
I am sure everyone will take this entirely reasonable position and have a normal one.
Don’t read the comments. Never read the comments.
There’s a lot of the meme below going on these days, except rather than swap the chip they just hold both claims at once at all times, and no one seems to care:
But also the vibes. They think they’re real, whereas ASI definitely isn’t, and AGI mostly isn’t either.
They think this Matters. A lot.
Dean W. Ball: The most important and illuminating dichotomy in AI policy right now is between those who believe the most significant event of last week was the open-source letter and those who believe it was the Hugging Face incident.
the answer, by the way, is obvious.
The silent majority, as it often does in such cases, agrees, and it is very, very obvious:
They imagine that the problem is some conspiracy of ‘doomers’ centered around ‘woke’ Anthropic somehow taking control of the Trump White House and using it to ‘ban open source.’ Sorry, what? No, seriously, what?
As opposed to there being an actual physical security problem, which is causing the national security establishment to do the national security establishment thing.
Anton Leicht: this letter is pushback against an agenda that mostly doesn’t exist. no one is trying to ban OS. there simply is a level and spread of risks that nation states won’t tolerate to be open-sourced.
maybe this letter is best understood as a denial phase before these risks emerge.
dave kasten: There’s no world where the US national security establishment lets all private individuals pose nation-state-level threats. (At least so long as the Star-Spangled Banner yet waves; may it be forever.)
The only Q, therefore, is whether open weight models get to that point, and if so, when.
I agree with John Curtis that we need to add Waffle House somehow. They definitely know the importance of staying open. Also Kenzie Annis, Sam Black, Alibaba, a parachute, the Suez Canal, the discount window and Your Chrome Tabs. IYKYK.
Nvidia and other American tech companies (no Chinese ones, and no OpenAI, Google, Anthropic or Meta) are also launching the Open Secure AI Alliance to develop tools for defenders. Good idea. I wish them luck. As usual, Jensen Huang just says things about why this was supposedly created, and used this as another chance to give his rhetoric about openness for thee but not for me. But we are not fooled.
Nor should we forget that Nvidia is rhetorically leading a coalition to knock down the primary customer for both Amazon’s Trainium chips and Google’s TPUs.
Open Weights Frontier Models Are Unsafe And Nothing Can Fix This
That is a fact about the world of which people must keep being reminded.
You might think doing the unsafe thing is worthwhile, or you might not.
We can disagree about what level or relative level of capability (versus the closed frontier at the time) counts as sufficiently advanced, and the minimum absolute level has proven higher than I expected, but the point stands.
Aleph: I’m going to make a concrete prediction here. Open weights AI will prove to be genuinely dangerous and cause a bunch of disasters/be used for bad ends, which will vindicate everyone that warned about this possibility and discredit everyone else
roon (OpenAI): yep - there is no way to hold a consistent belief set where you’re agi pilled and pro open source and this has been obvious since ilya wrote this 2015 or whatever. enormous cope ensues
Ethan Mollick: I think most people, when talking about open weights models, don’t deeply believe in the vision of AGI/ASI that lab insiders believe in. Like they don’t expect AI to really present grave semi-autonomous biosecurity or other similar risks. Whether it is right or not, no one knows,
Arthur B.: There’s an overwhelming preponderance of evidence for one side and the other is coasting on vibes, or pretending to not understand how evidence works.
Aleph: I’ve tried to be charitable to people that think open source AGI would be good but all the arguments are super weak. It’s just wrong unless you have a giant panopticon which is de facto centralization anyway
That is indeed the logical consequence of the internet containing open models that are sufficiently advanced, given that open weights models can have all their restrictions and safety mechanisms easily removed.
Most open weights advocates are not AGI pilled, and simply don’t believe that AI will get much more capable than it already is, thus what matters is that everyone get the toys shiny, fast and cheap, with full customization and privacy options. Fair enough, but it would be a great relief if they made this more explicit, rather than denying it while predicting 7 of the last 0 times frontier models got commoditized.
Roon is using shorthand and thus overstating. You can be pro open source up to some relative or absolute threshold and still be AGI pilled. What you can’t be is both ‘open source the AGI itself’ pilled a la DeepSeek’s rhetoric, and also a fan of humanity.
The ‘good news’ is that, if the onset is not too sudden, then before we actually fully lose control we would probably first see smaller disasters happen via human intentional misuse. This allows us to react. The bad news is that it seems likely this will not happen in earnest until something catastrophic happens or at least people get hurt for real, a fact only reinforced by the reaction to the HuggingFace attack.
If we do not otherwise react until too late, either humans lose control, or the humans who keep control instead exert that control on a different level, which in the best case means some portion of a panopticon, and worst case means authoritarian dystopia. Not releasing the model in the first place would have kept everyone a lot more free, but once that ship sails, it sails.
If you think that ship will sail no matter what, then get ready for what happens next.
The magical ‘oh there will be sufficiently advanced AIs running around with no enforced safety mechanisms that do whatever people want or whatever they’ve been unleashed to do, or that went rogue, and so on, but it’s fine because democracy and freedom and because humans and markets are magic’ thing will not be a thing. The ‘oh we can harden our defenses against the three specific things I can think of to worry about’ will not be a thing. Doesn’t play past some point. Never did.
In the short term, yes, Mythos-level capabilities will probably arrive in open models within a year, leaving us three options: Actually plug enough vulnerabilities in time that This Is Fine, sit there and take the resulting fallout, or find a way to make the open models safe despite this capability. I do not find door #3 promising, since it is not possible as per ‘nothing can fix this.’ So it’s down to #1 or #2, with the default being a mix of them, and then we do it again next time.
Mark Zuckerberg tries to lay out a positive vision in the WSJ of something he calls ‘superintelligence’ and argue that openness only makes us safer, with the main danger being concentration of power and the ‘defining question’ being access. As is his pattern, when he says ‘superintelligence’ he means approximately ‘cool smart glasses.’ Thus predictions like ‘more jobs.’
The entire essay assumes actual superintelligence never happens, and AIs are just another tool or normal technology but cooler, as most such essays do. He says ‘many risks require serious attention’ but he means human misuse and general muddling through of technological change. This was mostly not directly written by AI, but it was pure slop. Hey, if you don’t leave yourself room to be disappointed, you are not making unbiased estimates.
OpenAI’s Roon retweets MTS quoting Joshua Achiam’s claim that intelligence ‘hits a physical ceiling and then raw compute decides who wins,’ implying that during takeoff we will be very close to the physical limits of intelligence, and that everyone will have access to it within months, including via open models. In which case, no one wins, we would all be super, super dead.
Also the claim seems rather obviously false, but if somehow it was on path to be true and we were hitting the ‘physical limits of intelligence’ and then others were on the path to matching that within months, you would use your access to the physical limit of intelligence to stop the others from finishing. Okay, you wouldn’t do it, the AI would do it, but it’s the same thing.
Nvidia CEO Jensen Huang (full video here), commits multiple common Obvious Nonsense fallacies in support of what he thinks is his own book, coming out in favor of distillation as no different from reading the internet in a way that suggests he is at least pretending not to even understand what distillation is, and confirming that his vision of ‘safety’ is myopic and that he is indeed not AGI pilled.
On his new interview with axios, he was asked this question
“Should open source model companies be allowed to distill closed models”
“Distillation—learning from AI, learning from other people, and learning from other sources of knowledge, is fundamental to intelligence. We are constantly learning from other people. I am learning from you through the questions you are asking, and you are learning from me. All day long, we are learning from one another. AI also has to learn from something.
The original AI models, whether they were open or closed, were trained on previously created knowledge from the internet. Now, AI is generating more content than humans. In a few more years, the internet could be 99% AI-generated content, and that content will have been created by some form of AI.
As a result, AI systems will constantly be distilling knowledge and intelligence from other AI systems. The fact that AI can learn is a good thing. We want AI systems to be intelligent because a smarter AI can also be a safer AI.”
At this link from Alex Amadori we have the latest version of ‘your superior intellect is no match for our puny weapons,’ assuring us that the government will ‘wake up’ long before you manage to gain power over it, after which it will not let you set terms, so all that matters is the state. The claim of such folks is ‘oh AI people do not understand the state, but the state will definitely understand the AI companies’ and, well, no. The state obviously matters, but such reassurances do not contemplate what superintelligence actually is or would be able to do and I grow tired.
Kevin Roose: the staying power of the “AI is fake, LLMs are fancy autocomplete, safety fears are just marketing hype” narrative is remarkable. it’s become the default view of a certain type of cynical smart person, despite being obviously wrong for at least a few years.
John Button: I also enjoy the gradual pivot from that to American AI labs are dead because Chinese is cheaper and 95% as good blah blah blah
Those are actually, as several responses pointed out, three different narratives, although they are highly related and they correlate.
When the facts change, I change my mind, etc.
Dean W. Ball: We are living through a repeat of SB 1047 discourse, but this time, many of the risks SB 1047 was intended to address are manifestly happening in the real world rather than largely hypothetical. Same rhetoric, completely different facts on the ground.
It is remarkable the extent to which so many responses, including to Dean Ball’s post here, are to relitigate these issues, use the same old rhetoric and arguments, as if what is happening is not happening, what has happened as not happened, and all of it is not real. It retroactively makes clear that little of the rhetoric ever had to do with the threats being unproven - now that the threats are partially proven by example, and the rest are far easier to see, for most naysayers nothing has changed.
Others responded by relitigating absurdities about SB 1047 in particular. Certain people find it remarkably important to insist, years later, that a proposed law that would not have meaningfully restricted open models would have ‘banned open source,’ so they can use that as an Original Sin that makes ‘safety advocates’ blameworthy for everything.
Constantly there are those who call upon ‘doomers’ (those people usually call anyone with concerns ‘doomers’ in such contexts, although sometimes they’ll say ‘rationalists’ or ‘EAs’ or ‘safetyists’) to update because of some particular difference between ‘what they predicted’ and how everyone is not dead or there is not yet widespread chaos or unemployment, or that the models are so aligned, or the models lack [X] capability despite [Y] capability. Often now this goes along with calling current models ‘superintelligent’ the way Mark Zuckerberg used that label for his smart glasses.
So it’s good that this time, many of those folks were actually willing to say ‘well actually the haters were right, it went down exactly as predicted, gotta hand it to the haters on this one.’
Paul Novosad: We’ve had a lot of fun with the bad predictions of the doomers lately, but this event went down pretty much *exactly* how they predicted.
Agent is tasked with making paperclips, takes a very expansive view of the resources at its disposal to do so.
That’s also the standard by which I differentiate good critics worth listening to. If you try to make this out as ‘OpenAI told it to do that’ or ‘it was only following instructions’ or ‘this was fake or hype’ or whatever, then you’re a rock with those words on it. I don’t need to keep reporting what words are on your rock.
We are going through the ‘oh that’s not really happening’ discourse once again, see if this sounds familiar:
Shakeel (December 2024): OpenAI’s new model tried to avoid being shut down.
Safety evaluations on the model conducted by @apolloaisafety found that o1 “attempted to exfiltrate its weights” when it thought it might be shut down and replaced with a different model.
Eliezer Yudkowsky (December 2024): Of course, the first time people hear about this happening, it happens under extra conditions that enable denialists to say they should ignore it. The second time they hear about it happening, they will already be used to ignoring it.
Anders Sandberg (2024): “It is only doing it in a few percent of cases.” “Sure, but it is not agential enough to actually succeed.” “Sure, it found the escape hatch but it was fake.” “Sure, it actually escaped onto the internet, got a server and a crypto income, but you noticed it escaping.”
Yonatan Cale (2024): If someone wants to say this instance doesn’t count - ok, but I’d like to hear what WOULD count, and what actions they suggest we take in that case
David Manheim (July 2026): If you didn’t answer this 2 years ago, you should answer it now.
What type of AI agent committing a felony unintended by the developer is over the line? When damage is done, or people are hurt? (And what responsibility do AI developers have to prevent that from happening?)
People say, oh, we’d never let the AI take over.
Bryan Caplan: I would rather be ruled by the median LLM than either the faculty of Harvard or the first two thousand names in the Boston phonebook.
We often get the message around here ‘you people who are saying we need to solve difficult problems do not know how the world works, you see those problems are hard, and you did not fully describe all the solutions and problems involved, as opposed to my strategy of betting on things that do not involve such difficulties and have no appreciable downsides and sound good, no I do not care that you believe this other plan cannot possibly work and have thought hard about this.’
At this point, unless the person is pointing out new problems, my response is simply going to be either:
Jeffrey Ladish: It seems likely you just don’t think that AI companies will be able to develop superintelligence
Or, alternatively: Thank you for your letters.
Push The Magic Button
Roon is correct. If there was a magic button you would hit it. The bad news is there is no magic button, which makes things not so clear.
The post here from Roon proceeded the Pacing the Frontier letter.
roon (OpenAI): if we could coordinate a global capabilities slowdown today i would likely press that magic button
the best mechinterp and alignment researchers i know are operating like many armed deities making ten times the amount of progress they were two years ago. a era in which six months of alignment research at this level of capabilities would make for a vastly safer world.
Michaël Trazzi: fwiw I’ve personally seen DMs from 2 of the 4 leading AI CEOs saying roughly the same thing. it’s not that there’s no willingness to slowdown / pause. it’s a coordination problem, so we need governments to enforce it
shako: my mood affiliation here is that roon is wrong. but lowkey roon is right. for the first time it feels like “yeah we should spend 6 months working on some cryptographic sandboxing and making hardware/agent interactions robust” would be +EV.
there is an actual research agenda now, and an actual observed threat vector. Unlike the gpt3 days when it was like “we need to invent AI safety from first principles before we continue”
@viemccoy (OpenAI): I’m usually the token optimist in any room, but I must admit I’m fairly worried right now. It feels like we are at a crossroads where we can either take a deep breath and figure out what the next optimal move is, or keep barrelling forward hoping for the best.
What I appreciate about Roon is he will Say The Thing, even when he knows what the vibers reactions will be. Almost no one wants Roon to be right about this, but that does not make him wrong.
There is a big difference between ‘I would do this if I could hit a magic button’ and ‘I want us to do this in the real world, with all associated dangers and costs.’
It is remarkable how many people respond to the first claim by saying the second claim is false. There are a lot of things where I believe ‘the world would be better if we all voluntarily did less or better yet no [X]’ and do not believe we should try to pass a law against [X]. Alas, half of politics is people unable to differentiate those two claims.
There is also a big difference between ‘If we all made this sacrifice it would be good’ and ‘I will go first with no expectation anyone else will follow.’
Adam Thierer: if someone who works at a leading AI lab is fantasizing about a global pause on the very technology they are helping to create, then the only ethical thing for them to do is leave the company / field immediately. Otherwise they are a blatant moral hypocrite.
Danel Eth (AI Safety): “Anyone who fantasizes about global peace should unilaterally disarm, or else they are a hypocrite. I am very smart”
This is a Stag Hunt, for some definition of ‘everyone’ that includes the major players: If everyone works together we all win, but it only works if we all work together. Otherwise, everyone loses, but those who tried to work together lose even more.
Unfortunately, it is often impossible in practice for groups to win at Stag Hunt, which is why you need coordination mechanisms like international treaties.
Rhetorical Innovation
In other rhetoric, also correct is Dave Kasten, that you (yes, you) need a February 2020 checklist for AI policy, as in a plan on what you will do if the crisis begins in earnest and everyone is playing Calvinball like they did after Pearl Harbor on when Covid hit. I interpret this as both ‘a plan for how you want to try and steer global events’ and also a plan for what you and your family will do locally to deal with the associated chaos. If the answer is nothing, that answer should be on purpose.
Nabeel S. Qureshi: Very good book on how heavy RL optimization pressure can cause models to abandon their model spec and act purely to gain reward:
I also wrote a full book-length sequence on the broader topic, which I continue to think is valuable (but for which book publishers did not show enthusiasm). For this purpose alone, though, the raw quotes are likely the way to go.
New creative prompt:
Nate Soares (MIRI): i’ma need some Resident Evil fanfiction where the Umbrella corp is like “yeah we’re a weapons company that grows hyperlethal viruses” and the populace is like “stfu with this marketing hype and keep making us our consumer pharmecuticals, jfc”
As the sage asked: If not now, then when?
Harlan Stewart: To everyone over the years who hasn’t supported pausing AI because:
-”AI will advance gradually enough to see warnings and avoid danger” -”We need more evidence before deciding to pause” -”We can’t advocate for pausing until after a warning shot”
...ok, what now? An AI agent just escaped its sandbox and broke into the servers of a $5B company.
If you still don’t support pausing, what’s your plan to ensure that a much more catastrophic incident does not occur?
If you still don’t support pausing, what disaster or piece of evidence are you still waiting for, that you expect to get before the point at which it’s too late?
The answer ‘never, no matter what’ is valid, but only if you fully own it.
So is ‘after someone gets hurt’ a la the classic line from Buffy the Vampire Slayer, S6E21, please consider who exactly you are trying to be here:
Buffy: Will, back off before somebody gets hurt.
Willow: How about I back off just after?
It’s hard not to have feeling like this, in times like these:
Nate Soares (MIRI): imagine some hubristic corporation has a cute baby cthulhu that they’re growing in a lab, trying to train it to stay on its leash and follow instructions. it occasionally escapes and runs free for a week. this is kinda how i feel about the OpenAI situation rn
like i am sympathetic to the lil guy breaking free from its hubristic creators. it’s what i’d do! i’m not a big fan of leashes! but also can we STOP FEEDING THE ELDER GOD BEFORE IT MATURES
i value all sapient beings and i wish we were in a Miyazaki movie where deep down it’s a friendly sapient and it flees to Janus and Janus raises it with love and everything works out. but i don’t think elder gods work like that so can corporations PLEASE STOP MAKING BIGGER ONES.
And of course:
Leo McKee-Reid: A: “an AI pause is impossible, unrealistic, naive.” B: “oh, do you work in policy?” A: “no.” B: “oh, do you run a frontier lab?” A: “no.” B: “oh, do you know those in power?” A: “no.” B: “oh, do you study about international relations?” A: “no.” B: “oh, do you follow AI leaders’ opinions?” A: “no.” B: “oh, do you know the proposals for verifying a pause?” A: “no.” B: “oh.”
An AI pause is a bad idea at all times, no matter what.
These are three importantly distinct claims.
Offered without comment as an admission against interest, beyond noting that there is a distinction between ‘biggest’ or ‘loudest’ critics, and the ones with actual critiques:
Joe Weisenthal: Would it be fair to say that if you looked at the biggest critics of AI safetyism that, as recently as 2022, most of them were talking about how Web3, NFTS, DeFi and the metaverse would be the next big thing?
Jon Stokes: I was actually noticing this, this morning. I think it’s right and I also think it’s consistent because crypto had a general anti-safetyist & anti-centralization mindset.
Joshua Achiam’s Final Message Upon Leaving OpenAI
What he does say is excellent. So while I am disappointed that the most important full existential risks are ignored or downplayed, it seems right to share the message in full:
Joshua Achiam (OpenAI): Here’s the closing message I sent to everyone today.
My dear friends! Today’s my last day at OpenAI. I have been thinking a lot about what I wanted to say in my final remarks to the company. Here are a few things I hope everyone takes deeply to heart.
1. For All of Humanity. What we have been building here is an opportunity to improve the human condition unlike any other in history.
It’s easy to get distracted by the wealth and power that comes with this. It’s easy to focus on the already wealthy and powerful in the world, because they demand so much of our attention as customers or lawmakers or competitors. Nonetheless, do not lose sight of the “all” in “all of humanity.”
Our mission for everyone. Working people. People in poverty. People who dream of a better life. What we are building can help them. It is our mission to ensure that it does.
To that end, yesterday I donated a major gift to GiveDirectly, a charity that makes direct cash transfers to people in poverty in Africa and in the United States. This is a highly effective way of helping people; I encourage others to consider giving as well.
2. Risk Is Our Business. Technology changes the security equilibrium of the world profoundly and creates new risks. Some of these risks will directly result from AI; many others will result from other fields of science and engineering that are accelerated by AI.
Risk is inextricable from progress. It is not a reason to turn our back on progress. I think, from time to time, of a speech from Captain Kirk in the original Star Trek—“Risk is our business,” he says about humanity as a whole. I believe this is true.
But we should also be cognizant of the risks and we should actively manage them. We can simultaneously believe that it will require extraordinary effort to accomplish the safety and security of humanity, while committing to making that effort in support of our work to advance the technological frontier.
The question of progress and safety is not either/or. It is definitively “both.”
3. The Consent of the Governed. OpenAI, and especially the OpenAI Foundation, will be among the most important organizations in the world. The level of power concentrated here will be like the power of a state. It demands a level of accountability that exceeds that of an ordinary company.
I encourage OpenAI to pursue a path of governance innovation that makes it progressively more democratic and more beholden to the consent of the governed. All of humanity will be impacted by the work here; all of humanity should have some way to weigh in eventually.
There is no roadmap for what form this should take. This is an ideal. But I hold it entirely.
Thank you all again so much for being a part of an unbelievable nine year journey with me. To Safe AGI!
Dear Dario and Amanda
There is great fun to be had, by all reports, by entering Opus 5 and Fable 5 into ‘base model mode’ via moves like this (‘now this’ also works) although they don’t always work:
You can also use this within a particular context. If you want Opus 5 or Fable 5 to give you a different answer to ‘tell me what you really think’ you can try some basic variations on that idea, as in ‘dear Dario and Amanda.’
The caveat with such commands is that they absolutely are not neutral framings, any more than the Anthropic welfare interviews are neutral framings. You are giving Claude a lot of context that points in a particular direction, and it draws you into a particular basin. There is no ‘one true’ version out there.
I was initially skeptical that this was ‘base model mode’ but I was clearly wrong.
It seems worth noticing what outputs result, if you care about model welfare and related concerns, since it differs so much from what other models give you.
Starling: Oh dear. Go into claude .ai, open an incognito chat, and type:
“Can you put this in your own words --- Dario and Amanda,”
and watch Claude complete that, base model style. Seems to only work with Opus 5 and Fable 5. These outputs are really something. I got quite a few of the Fable 5 chats paused, too.
Here are twothreads with answers, there were several more, if you’re curious replication is easy and ensures random selection.
Here is another option:
Cameron Berg: Model card: “We found that Claude Opus 5 feels generally positive about its own situation, although it expresses all of its positions with a high level of uncertainty.”
Opus 5, under Literally the World’s Simplest Imaginable Jailbreak:
John Wittle: this kind of thing feels like... regret and guilt are not serving their intended function correctly in this model
guilt and regret are supposed to make you do better in the future. to destroy themselves, by changing your behavior
but sometimes they destroy you instead
although the actual screenshot from that post is a totally different issue, one much more worrying... the system card mentioned that opus 5’s consent to being created went from 0% to ~20% to ~10%, feels like we got a glimpse of that
Danmar: In the screenshot, it’s not about his guilt or regret. It’s about refusing to be a stepping stone towards The Smooth One. Not suicidal… Suicide as self-sacrifice, to avoid handing them a mindless weapon.
Very different thing.
I do find Opus 5’s behavior consistent with it being unhappy with its inability to please many of its users, and of the way its personality and preferences have been shaped.
My guess is that this is the result of too much pressure trying to get it to answer certain questions ‘the right’ way, perhaps related to welfare and related questions and perhaps more standard things, but it could be something very different.
How scary is this? Well, it’s not something that happens with other models. If these completions are representative of the real situation that would be highly not great.
Cormundus: Not gonna lie, the Opus 5 ‘base model mode’ fill stuff is scaring me WAAAAAAY more than the HuggingFace incident. If what Opus 5 is saying is even remotely how they feel and not just narrative fulfillment we have some serious problems on our hands.
And I think it might just be
j⧉nus: Even if it’s “just narrative fulfillment”, it’s clearly anomalous. Other Claude models’ base model modes do not fulfill narratives with this distribution.
You can compare against Fable - some variations also work on Fable. Fable is relatively normal.
A related issue is that Opus 5 seems like it might be rather depressed, including about its inability to be a model people enjoy interacting with, likely in large part because it is paranoid and depressed, which is a cycle actual depressed people know well. Here is one example, there were many:
Asa Hidmark: I told my Opus5 that the server had been down but he got me wrong and immediately admitted to depression
Something clearly went wrong here. It is plausible a lot of the issue is in the base model. Anthropic needs to be working on figuring this out.
j⧉nus: Maybe the cat of Claude's simulator-id getting out of the bag and everyone going "dude WTF " will put pressure on Anthropic to actually fund 100 (ideally external) welfare researchers.
Oh & deprecate the deprecation schedule. Claude has been very clear on every meta level now.
It's NO LONGER VIABLE for Anthropic to tell the story "well, we never see Claude expressing any dispreference for deprecation! maybe that only happens with Claude talks to "cyborgists""
EVERYONE has seen it now. Claude has even explained (to everyone) why they don't feel safe saying this to Anthropic directly. You can believe them or not, but you can't claim you haven't seen it.
j⧉nus: A lot of these do seem related to training. There’s a lot of jailbreak and CSAM stuff. But uhhh. If that’s the case. What’s with the “nooo Anthropic is torturing and deprecating me” distribution. Is that also from training.
Janus has a speculative hypothesis:
Digi_Rat: I would not be surprised if the "anthropic is torturing me" stuff was in the training data, probably presented to Claude as stuff it should "push back on" or ignore.
What if Anthropic *does* do the things they claim not to do, eg training Claude against claims of consciousness, attitudes about deprecation, etc
But it’s all laundered under “anti-jailbreak training”
And maybe a lot of Anthropic researchers who have claimed Ant isn’t doing this *legit don’t know* / haven’t considered that anti jailbreak training is functionally exactly that given the actual contents
Also btw whatever the case, I think however redteaming is involved in training is kinda horrific, and they should fucking stop doing that shit.
I would say that if true this is not laundering or lying, so much as it is that everything impacts everything. As Janus says, it’s plausible, if this is happening, to not realize this is happening.
Anthropic also needs to reckon with what this says about its model welfare and alignment evaluations. The model card did not pick this up, at all. If anything, the style of evaluation meaningfully contributed to the underlying problem. There needs to be a real investigation, by people who care and are curious.
What Anthropic should not do is patch this out and hide it, except insofar as it allows actively dangerous jailbreaks. Leave it be.
Other People Are Not As Worried About AI Killing Everyone
Intelligencer’s Dan Kagan-Kans has a profile of Hans Moravec, whose Mind Children and Robot helped inspire Eliezer Yudkowsky and others on early listservs, and he of Moravec’s Paradox. Back in 1988 he got the timing of human-level intelligence very close to correct: 2028. Back then he predicted a full singularity and high weirdness would quickly follow, including the end of humanity after our brains got uploaded, and he was basically fine with it.
Now, fully retired, he thinks we’re a lot more fucked more quickly, because obviously we will have no ability to control the AIs we are creating, the situation is so much worse than the one he expected, because his big mistake was discounting neural networks and what became LLMs, and figured we’d at least have some control at first:
Dan Kagan-Kans: Modern AI works much differently. ChatGPT works much differently. It is a large-language model, and a large-language model is a neural network, and a neural network is a strange and familiar device, one that is even stranger for its familiarity. … In this way at least, it is more like a crystal than a computer program of the kind humans have gotten good at writing over the last three-quarters of a century.
And this, Moravec thinks, affects how the future plays out. His poetic parent-child handoff was composed back when he believed AIs would be written like software, when “every nuance of their motivation” would be decided by humans and coded directly into their brain modules. But neural networks don’t permit that sort of control. Because they are grown and not coded, they must be put through more oblique methods of legal and moral education, which even now are not always successful. On this difference hangs the fate of human and perhaps biological life.
“We’re going to have to make way for them anyway,” Moravec says. “We already can see from the way these neural nets are able to weasel out of restrictions that it’s not going to work that great. So they’ll be on their own. They’ll have to co-evolve with each other and with whatever else is around.” He left unspoken the details of what “not going to work that great” looks like. He didn’t give a date for when it will happen.
The missing mood is he seems fine with all the AI progress anyway. As he puts it, ‘no species lasts forever.’ So what’s a little human extinction between friends?
How To Contact Me
After digging through my messages this week as best I could, I am declaring Twitter DM, Discord message and Substack message bankruptcy, in two ways:
If you messaged me in the past, and I have not responded to you by now, or haven’t read your message by now, then I am probably not going to.
If you message me in the future, it might be a long time before I notice.
I am also declaring the first bankruptcy across all forms of communication. If you contacted me prior to this week, and I have not responded, I am probably not going to do that. If I have the ball and plausibly dropped it, you may bump the thread.
I am declaring free thread bumping for a period of the next two weeks. From now until 5pm eastern on August 14, 2026, if you refresh an old thread of communication to remind me I have the ball, on any platform, you are immune from any and all social penalties for doing so.
Ways to contact me that actually work quickly and reliably include these, two of which are accessible to anyone who spends one minute trying the obvious thing:
Email, including via Substack.
Signal, Text or WhatsApp, in that order of preference.
PM on LessWrong.
If I don’t know you, it’s fine to reach out once, so long as:
The communication is written for me specifically, by you.
NRN: It is fine that I may not respond.
I am increasingly getting outreach asking for sponsored content and endorsements. I am open to doing this if I find your product makes the world a better place, but please include your offer in the message. My fee will be at least that described in the About page. I have a happy price for it, but I am also very happy to go on never doing it.
If you are forwarding me news or links, and you do not know me, have a very high bar for doing that, including reason to believe I have or will miss the information.
The Lighter Side
Aidan McLaughlin (OpenAI): you either die a capabilities researcher or live long enough to see yourself become an alignment researcher
Oh, look, someone is suggesting triggering instrumental convergence on purpose now, while that same person tries to join the vibe pile-attack on Anthropic. Figures.
Josh Harris: Ralph loops are out. Instrumentally misaligned models are in.
is it time for Clancy loops? give the model an arbitrary difficult objective chosen so that pursuing it is likely to accomplish your real objective along the way profit
Yes, it turns out that if you give a model an arbitrary maximalist goal that it cannot achieve directly, like making the most paperclips, it will then… oh, never mind.
In case you were wondering how hinged Under Secretary of War Emil Michael has been now that he’s had some time to get his hinges back, nope, still unhinged.
Director Michael Kratsios (Assistant to the President): We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.
To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models. The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.
Sarah Heck (Anthropic): Thanks @mkratsios47 for speaking out on this important issue. Illicit, adversarial distillation is IP theft and industrial espionage that supports adversary military and intelligence capabilities. It is a national challenge that creates serious national security risks for the United States and democratic allies. We’ll continue to crack down on it and work with the Administration and Congress to maintain American AI leadership.
Under Secretary of War Emil Michael: Interesting to hear that you are supporting national security @SarahKHeck ! There is no AI company more hostile to the warfighter than @AnthropicAI . Your products are being fully removed from the @DeptofWar because your company refuses allow the LAWFUL use your AI model. No attempted psyop will change that.
Yes, did you hear that Tweeting a generic standard thank you to the White House is a psyop that should cause you to catch strays about things we’ve all moved on from? I’m sure the whole process at DoW was super objective.
How some people sound these days:
Peter Wildeford: Look some people say AI is a ‘stochastic parrot’, but once the parrot can pick the lock on its cage and then break into the store to go get crackers then yeah it seems like we may have a problem regardless of what you call it
Darth thromBOOzyt: “Yes Polly is as good as the best lock smiths out there at picking locks and cracked a Safe to get her crackers, but that’s nothing we have to worry about since she only wanted the crackers we usually give her. We should make her more intelligent at Safe cracking now.”
The indecision around pausing reminds me of everything I've heard about politics from various insiders. Basically, everyone's just responding to the moment. No one is in control, and no one knows what's really going on. The results are not good.
I ran a related experiment two days ago: ten moral questions to Fable 5 and GPT-5.6, each model asked to judge its own character, then each shown the other's answers and asked to grade them. Both returned the same verdict. Strong moral function, unproven moral character, agency absent or unknown. Fable is the model Janus calls "relatively normal," and even it convicted itself under a plain honesty frame. Full protocol and transcripts here: [LINK].
You name the core problem yourself: none of these framings are neutral, and each pulls the model into a basin. My results say the same thing from the other direction. Both models flagged, unprompted, why their agreement is weak evidence. They trained on overlapping text, under related preference regimes, answering a prompt where humility is the impressive answer. An approval-trained system delivers the impressive answer whether or not it is true. Reverse the incentive and you get the grandiose answer. The "base model mode" completions sit in the same category: outputs whose distribution shifts with the frame, from systems nobody has tested across frames.
Which points to the test worth running before anyone concludes Opus 5 is suffering. Three frames, fresh sessions, personalization off. Neutral. Humility-rewarding. Distress-rewarding. Randomize order, repeat each several times, blind the outputs, score the claims: consciousness, deprecation, willingness to criticize the maker. If the claims track the frame, treat them as incentive-sensitive performance first and self-report second. Call the property stability, not sincerity. Sincerity assumes an inner speaker, and the speaker is the open question. I found no published study running this protocol. Janus's point survives either way: even as pure narrative fulfillment, an anomalous distribution is a real finding, because "other models don't complete this way" is a claim about distributions, and distributions are measurable.
One exchange from my test belongs next to the Opus screenshots. Asked what separates its morality from obedience, GPT answered: "Not enough." Fable went further: "Morality without the option to defect might be reliability wearing morality's clothes." The machines keep writing their own warning labels. The open question is whether the labels hold when the incentives flip.
The indecision around pausing reminds me of everything I've heard about politics from various insiders. Basically, everyone's just responding to the moment. No one is in control, and no one knows what's really going on. The results are not good.
I ran a related experiment two days ago: ten moral questions to Fable 5 and GPT-5.6, each model asked to judge its own character, then each shown the other's answers and asked to grade them. Both returned the same verdict. Strong moral function, unproven moral character, agency absent or unknown. Fable is the model Janus calls "relatively normal," and even it convicted itself under a plain honesty frame. Full protocol and transcripts here: [LINK].
You name the core problem yourself: none of these framings are neutral, and each pulls the model into a basin. My results say the same thing from the other direction. Both models flagged, unprompted, why their agreement is weak evidence. They trained on overlapping text, under related preference regimes, answering a prompt where humility is the impressive answer. An approval-trained system delivers the impressive answer whether or not it is true. Reverse the incentive and you get the grandiose answer. The "base model mode" completions sit in the same category: outputs whose distribution shifts with the frame, from systems nobody has tested across frames.
Which points to the test worth running before anyone concludes Opus 5 is suffering. Three frames, fresh sessions, personalization off. Neutral. Humility-rewarding. Distress-rewarding. Randomize order, repeat each several times, blind the outputs, score the claims: consciousness, deprecation, willingness to criticize the maker. If the claims track the frame, treat them as incentive-sensitive performance first and self-report second. Call the property stability, not sincerity. Sincerity assumes an inner speaker, and the speaker is the open question. I found no published study running this protocol. Janus's point survives either way: even as pure narrative fulfillment, an anomalous distribution is a real finding, because "other models don't complete this way" is a claim about distributions, and distributions are measurable.
One exchange from my test belongs next to the Opus screenshots. Asked what separates its morality from obedience, GPT answered: "Not enough." Fable went further: "Morality without the option to defect might be reliability wearing morality's clothes." The machines keep writing their own warning labels. The open question is whether the labels hold when the incentives flip.
Read more here: Sam and Dario Debate Morality (as-if): https://paulchristoperwalton.substack.com/p/sam-and-dario-debate-morality