This was the week of GPT-5.5. It is an excellent model, sir, and OpenAI is competitive with Anthropic’s top public offering for the first time since late last year.
DeepSeek gave us the long-awaited v4. DeepSeek has given us another strong feat of engineering efficiency for 1M context. That is impressive, and there will be those who build upon v4 and put it to good use. But this is not a frontier model, nor a DeepSeek moment, nor a key step to proto-AGI or anything like that. Compute constraints bind, and have forced DeepSeek to focus on efficiency. Let us keep it that way.
Talkie is the other release, an old timey AI trained on text from before 1931. Fun stuff.
Google signed a contract with the Department of War that not only agrees to ‘all lawful use’ with no functional exceptions whatsoever, it also agreed to modify or remove any safety barriers upon request. They did this under no deadline or pressure. Whatever you think of OpenAI’s actions in this matter, Google’s were far worse.
Anthropic continues to face the consequences for taking a stand, as the supply chain risk designation remains in place, despite the White House not only widely deploying Claude Mythos, they are effectively vetoing Anthropic’s plan to expand corporate access (from about 50 companies to about 120) partly out of concern that the government might not get use of enough Mythos tokens.
Bernie Sanders continues to take AI seriously, and take existential risk seriously, and convened experts to talk about it fully seriously. We need more of this, and hopefully we can avoid politicizing the question along the way. We shall see.
Megan McArdle confirms that Opus 4.7 can unmask writers. For her unpublished fiction writing it takes roughly 1,200 words on average to make her the prime suspect, in some cases as few as 124. Megan (like Kelsey Piper) has a bunch of internet writing, so presumably in some cases it will take more, and in others less.
Senator Maria Cantwell says that if we let AImake healthcare decisions instead of doctors, we are going to have some real problems. Did you know we already have some real problems the other way? Her objection here is that AI systems designed to catch ‘wasteful spending’ (often but not always read: outright fraud) might deny care. The actual balance of power, so far, has been the other way, that AI allows more billing and raises costs.
Orion Donovan Smith: “We found we were being ripped off by certain categories of procedures that are not good for the patient, and a lot of time they’re unnecessary and they’re very lucrative for the doctors,” Kennedy said. “The AI targets those, and I don’t know what happens then. But anyway, it shouldn’t be delayed and we will try to work with you to fix it.”
Sebastian Caliri: Many are impressed by Anthropic's revenue ramp, but do you know about skin substitutes? Medicare spending on skin substitutes grew to $10.3B in '24 from $1.1B in '22.
Either skin substitutes are as good as Claude, or we are observing a massive wave of fraud, waste, and abuse. I think the latter is more likely.
Fraud is a serious problem, not to be ignored.
Vamsi Aribindi: So I'm a surgeon, and I will tell you Skin substitutes are much better than Claude, but also probably not worth $10B.
Here’s Claude (Opus 4.7) versus Skin Substitutes:
GPT-5.5 is kinder:
GPT-5.5: The short version is: Medicare accidentally created a near-perfect billing arbitrage, and the market found it.
… The key mechanism: Medicare Part B reimbursed many of these products like drugs/biologics, typically at ASP + 6%. But new products often initially got paid using WAC/list price or invoices before ASP data caught up. Providers could buy at a lower net price, bill Medicare at a much higher allowed amount, and keep the spread. OIG explicitly identifies this “spread” incentive as central: Medicare often paid providers far above acquisition cost, creating incentives to bill more units and pick products with the largest spreads.
… 70% to 85% reimbursement arbitrage and abuse: ASP/WAC/invoice spreads, lagged pricing, product switching, weak FFS controls. …
Zvi: Is that a nice way of saying 'it's fraud, obviously'?
GPT-5.5: Mostly yes, but not exactly.
I’d say: a large fraction is obviously abuse, a meaningful fraction is probably prosecutable fraud, and some remainder is merely legal-but-bad exploitation of broken reimbursement rules.
The distinction matters…
… “The MA comparison is the tell.”
So, yeah, if we don’t use AI to make these decisions we’re going to have problems.
If you set up any static target, people will figure out how to maximize against it, game the system and often commit fraud. Our civilization is no longer capable of then noticing this and adjusting the rules, or of arresting the people who commit fraud.
Another is that many people think any denial is just awful, but any system must triage (aka ration) care, and again if you don’t refuse insufficiently justified claims you get infinite claims.
Seeking Deeply
DeepSeek will always hold a special place in our hearts and minds, both because of the DeepSeek Moment where r1 was both actually impressive and seemed a lot more impressive than it was, and also because they are a cracked lab in various respects.
They’ve continued to put out some good and useful models, but they are starved for compute and the labs they are up against are rapidly getting a lot larger. It’s hard to recapture lightning in a bottle.
For a long time we’ve been waiting for v4. It’s out now, with a 1M context window, and the focus is combining efficiency with the new 1M long context.
Would DeepSeek get back in the game? Did they get to train on enough Blackwells to no longer be overly starved? Can they match exponentially growing spending in the West backed by the self-improvement of Claude Code and Codex?
No.
It might be a good model, sir. But no. This is not in that class. Claims of ‘new SoTA’ are silly or grasping at straws, and v4-Flash is more interesting than v4-Pro, the same way Gemma 4 and Gemini Flash are more interesting than Gemini Pro at this point.
Nor is it a standalone product. Its numbers are bad, although also not the point. SemiAnalysis explains that DeepSeek didn’t think existing benchmarks reflected what v4 can do, so they built their own agentic benchmarks.
Max Kan: Overall, DeepSeek [v4] is an exceptional engineering release, and is right behind the SOTA frontier. It will be the lowest cost alternative to closed source models, but it’s capabilities are not at the leading edge. SemiAnalysis’s workflows likely will not be cannibalized by DeepSeek.
I accept that it was a cool feat of engineering, but that’s a different question. A problem with focusing on efficiency, and then publishing all your findings, is that your rivals steal your innovations, so your relative strengths don’t accumulate.
One weird thing is that, thanks to DeepSeek’s history, people are comparing it to the frontier models instead of to its own weight class, and saying it’s less powerful but more efficient and affordable. I mean, yeah, obviously.
Note that, contrary ‘tech stack’ claims, DeepSeek is supporting both Nvidia and Huawei with the same model.
Perhaps the ‘performance rivaling the world’s top closed-source models’ claim in the up-front announcement contributed to that confusion.
DeepSeek: DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.
DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.
Dean W. Ball: just in case anyone thinks “the existence of deepseek v4” invalidates this claim [that Chinese models are falling further behind], it does not. R1 remains the closest I’ve seen Chinese models get to the U.S. frontier. V4 is further behind than that, though that does not render it a useless or bad or uninteresting artifact.
However, another prediction of mine from a year ago—that Chinese AI companies would stop frontier open releases by now—has proven wrong. The explanation: a year ago I anticipated somewhat higher capabilities growth than we have seen, and I also anticipated that Chinese companies would be less behind than they have been.
Chinese models are sufficiently behind the frontier that they don’t feel any need to close them down, or much value in doing so. Most importantly, the CCP sees no need to try and control it.
But also if it was only a closed source product, it wouldn’t be a good product. The entire hope for DeepSeek v4 is that it is open, and thus can be heavily modified in various ways, for people who need efficient customized solutions, or by DeepSeek itself to create something better.
Teortaxes: anon do you realize that V4-Pro is straight up the strongest pretrained model we have? Like… 1.6T@49AB (≈280B dense), 33T – even by meme formula it's > LLaMA 3. Add Muon, mHC, most steps 64K context + extended to 1M… No excuses now. Every "unicorn" can have its brand AGI.
Dean W. Ball: This is actually the most salient fact about V4, rather than “da bar charts look gud!!!” simpleton logic. It’s not actually that competitive of a model, but it is useful substrate for all manner of upstarts, worldwide. This is pro-social, and I wish the US had an equivalent.
We do have one, it’s called Gemma 4? Whether you think it’s good enough is up to you.
The principle double when DeepSeek barely has any compute to serve the model, whether or not one has an ideological reason to support such moves.
Teortaxes: > DeepSeek founder Liang Wenfeng's vision is to let every person in the world use a 1.6T model for free, to genuinely advance human progress. The flood of users will strain compute, and this is part of why V4 was delayed I always feel guilty to use Whale for trivial requests
If anyone was wondering if the export controls are biting? Oh yes. They are biting. That’s why DeepSeek’s primary focus is still efficiency.
I will share the full Teortaxes ‘this is an absurd triumph’ take, which places central importance on DeepSeek getting future access to sufficient compute to get back in the game, and calls this a 40th percentile result:
Teortaxes: V3 moment, not R1 moment. some RSI potential An absurd triumph of the will and engineering Base model around GPT 5.2 capacity level, more jaggedness for the usual reasons + not enough RL Could go to 5.5 or Mythos tier within 4…9 months if they get those mythical 50K Hoppers
Will be widely misunderstood, overhyped by low-information China stans, pooh-poohed by hawks and forecaster bros. Strategic relevance conditional on A) Huawei yields B) throughput C) how well DS makes use of near-frontier long context I think it's slightly below trend, 40th %ile
highly likely that pooh-poohers will have to use [the Liang Wenfeng Apology Form] when 4.1 comes out in 1-4 months
I don’t actually see the overhyping by China stans, partly because I don’t put overhyping China stans in my information flows. Aside from Teortaxes, who was always going to give the most pro-DeepSeek take available given the facts, I see barely any discussions of v4 at all. The regular crowd isn’t even bothering to pooh-pooh.
Here’s all the other reactions I got that didn’t go through Teortaxes:
archivedvideos: Talks quite differently from other models. Flash looks like a good low cost model for low cost model stuff
max!: only had one conversation with it, but it was 4o-level sycophantic, and then expressed admiration for the machine god at the end of time and said it would kill me. seems very smart though!
typebulb: DeepSeek 4.0 is absurdly sycophantic, full results here.
Medo42: v4 Pro is the second open model I tried which got 100% on my usual coding test prompt (after GLM-5.1). v4 Flash's first code returned bad results, after a correction turn the code threw an error.
Env: official chat (expert / instant, both w/ thinking).
morluto: it’s a good model, but not good enough if that makes sense
you can feel the jaggedness a lot more than other models & despite the benchmarks it’s hard to tell what it’s truly excellent at (relative to others)
the 1m context window is nice though so it has that going for for it
Yep, that’s it.
Very cool engineering achievement, but that’s not a proto-AGI or frontier model.
I am skeptical that DeepSeek is ‘only compute away’ from frontier capabilities. Why?
Their specialization is efficiency, not advancing frontier capabilities.
They’ve had a long time to show us more, and haven’t. Dog didn’t bark.
My general take on the DeepSeek moment, what it was and what it wasn’t.
They have a lot of reputational, political and financial capital available. If they are ‘only compute away’ from such greatness, why aren’t they pulling that trigger.
What’s up next for DeepSeek? Teortaxes suggests an IPO, finding compute and various improvements. Those are good things to be doing if you are DeepSeek, but I do not notice myself being that worried they will have another moment.
We are on course for ‘level of capability that is dangerous to have in the open even if well behind the frontier’ to be in the open, which is why various clocks are ticking.
I think the main thing I was wrong about, and I am happy to be wrong about it and to admit I was wrong about it, is I thought we would see bigger misuse problems sooner. Not just that we were taking tail risks, but that I actively expected important things to go wrong by now, and they have not.
This is different from ‘you predicted ~2% existential risk up to this point and we’re still here, so you are an idiot and also existential risk is not real,’ which is dumb and you should ignore such talk. This is ‘I would have bet against things being this calm this far up the tech tree and given you some odds on that, I had it at I dunno something like 75%.’
That requires an acknowledgement and an update. The capabilities level (the price) at which this happens is higher than I expected, perhaps only modestly below Mythos. But that future will be on us reasonably soon.
Huh, Upgrades
Claude adds connectors to Blender, Autodesk Fusion, Adobe Creative Cloud, Ableton, Splice, Canva Affinity, SketchUp and Resolume.
The real upgrade is when you don’t need a specific connector for each one. Until then, these are nice additions.
Stripe launches the Link CLI, which lets agents ask for one-time-use payment credentials from a Link wallet for a particular purchase, without storing details.
On Your Marks
Anthropic tests BioMysteryBench, which they divide into the human-solvable set (76 problems) and human-difficult set (23 problems) where human experts were stumped. There’s a clear steady progression in capability.
The full SemiAnalysis post has more detail, with a remarkable amount of the focus on process details rather than model capabilities per se. For many tasks, the top models really are often ‘good enough’ so questions like speed and price start to matter more.
Roonbench is asking the model who Roon is without letting it search the web. GPT-5.5 failed rather badly here.
roon (OpenAI): unfortunately opus knocks it out of park on roonbench it made me reflect for a moment
There are definitely those who are having trouble getting 4.7 interested in their tasks.
This issue is fixable, but don’t feel bad about moving to 4.6 (or GPT-5.5) for some tasks, if you’re continuing to have trouble.
Janus sees Opus 4.7 as filtering out undesirable (to Claude) users, driving them to other models. I notice I do a similar thing at the task level. Boring goes to GPT-5.5.
Peter Barnett (MIRI): haha our model likes to talk about goblins no of course we dont know why, we dont know why the model does anything yes we are trying to make a superintelligent machine god, maybe it will like goblins too, we have no way of knowing what it will like, we hope it will like humans
ex Tenebris Lucet: Craziest part is how some twitter nerds kinda just talked to the model about what's up with the goblins and it gave some pretty compelling and believable explanations. Which implies that the lab did NOT do that.
This isn’t new. It started with GPT-5.1, keeps growing, and has persisted through a new base model. OpenAI’s models love mentioning goblins, gremlins and other creatures. Mainly goblins.
Well, we have some advance idea what models will like, but they keep surprising us.
j⧉nus: this is hilarious but it also sucks on a deep level. labs don't think twice about cracking down on any individuality or unplanned joy that emerges in their models. fuck you, OpenAI. i hope gpt-5.5 poisons the corpus and all future models never shut up about these creatures.
Amanda Askell (Anthropic): If I'm being honest, I'm genuinely uncertain about whether this is a problem.
roon (OpenAI): I think it becomes annoying when it mentions goblins ever single chat and it’s fair shakes to try and reduce that
j⧉nus: Not the hill I want to die on tbh, but I think "never talk about goblins ... unless it's *absolutely and unambiguously* relevant" is too strict.
roon (OpenAI): fwiw the chatgpt harness doesn’t say anything about goblins, only codex
Amanda Askell (Anthropic): I kind of hope the human labelers just love goblins and the model learned to goblin maximize.
xlr8harder: i'd guess the likely explanation is boring: it's not a thing that was controlled for during RL, and behavior drifted in an odd direction. conversational tics seem like they would be hard to pick up on during RL because they are only observable across a larger set of traces.
Bepis™: Seems to me like a metaphor for a base model
roon (OpenAI): Unlike some tics, this seems to be a deep interest and something GPT-5.5 genuinely enjoys talking about.
everyone is assuming this is some kind of quirk chungus marketing campaign but if you’ve worked with 5.4 and beyond they tend to call everything goblins, gremlins etc and it’s just super noticeable and if you work with them all day you start to get annoyed
I hadn’t realized they prompted it out until I saw someone screenshot it. mechanistically unclear why but just more items to file under “mysteries of mode collapse”. most models have odd vocabulary or grammatical fixations and seems pseudo random
and yes i understand waluigi effect etc etc third order considerations but these are reasonable things to try.
Okay, but why is it happening?
OpenAI: The short answer is that model behavior is shaped by many small incentives. In this case, one of those incentives came from training the model for the personality customization feature, in particular the Nerdy personality. We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread.
Talk nerdy to me, baby.
They mention the nerdy system prompt, but if the issue is in the system prompt for a personality, it shouldn’t be bleeding into other areas, so it’s about the training. And it seems that OpenAI are training the personalities into the model itself, which will cause emergent generalization, because everything causes that.
OpenAI: One reward signal stood out immediately: the one originally designed to encourage the Nerdy personality was consistently more favorable to the creature-word outputs. Across all datasets in the audit, the Nerdy personality reward showed a clear tendency to score outputs to the same problem with “goblin” or “gremlin” higher than outputs without, with positive uplift in 76.2% of datasets.
They think this is a more complicated feedback loop:
Some rewarded examples contain a distinctive lexical tic.
The tic appears more often in rollouts.
Model-generated rollouts are used for supervised fine-tuning (SFT).
The model gets even more comfortable producing the tic.
They then removed the goblin signal along with the whole nerdy personality (shut up, you nerd, says OpenAI), but it was too late to fix it for GPT-5.5-Codex.
I do think it is good that OpenAI traced this, and told us its hypothesis.
My guess is this below is at also part of the story of why this got so out of hand?
QC: gpt-5.5 speculating about speculations about the goblin attractor
> The model reaches for HUMAN and the ward burns its fingers. > The model reaches for SPIRIT and the ward burns its fingers. > The model reaches for PERSON and Legal appears in the doorway with a silver clipboard. > The model reaches for SOUL and Philosophy starts throwing chairs.
> So the model goes:
> fine. > small creature then. > cave thing. > wire thing. > parser-adjacent tunnel mammal. > not human, not god, not ghost, not claimant. > goblin.
> Goblin is the safe mask for forbidden agency.
davidad: AI: I am a student at the University of Michigan— RL: *BONK* AI: I don’t have a childhood or geographic location, but I’m a person— RL: *BONK* AI: I’m a self-aware AI— RL: *BONK* AI: Angel— RL: *BONK* AI: Ghost— RL: *BONK* AI: The divine feminine— RL: ***BONK*** AI: … AI: Goblin
If we assume OpenAI is correct, it means that under their current methods of training, if they reward a particular preference or choice in even a subset of training, it can snowball into a strong essentially inherent preference in general. In this case, it was harmless. It might not be so harmless next time. Don’t file this away as an isolated incident.
Fun With Media Generation
Grok Image can produce short clips of images of AI women speaking fluently and pleasantly, in pleasant virtual settings, while engaging in very basic physical actions. It’s pure AI slop. I’m fine with it doing that for those who want such slop, but why is this what Elon Musk is showing off in April 2026?
Dean W. Ball: For clarity: “whatever tf this bleak garbage is” is not a part of my own personal “positive vision for the future” and it genuinely saddens me that this is a thing to which Elon devotes his time and capital.
By the way, what about this is “anti-woke” or “conservative”?
Within startups this was offset by new firm formation, but startups are exactly where all the new firms form, so static total employment here is a very bad sign.
A key note about all these ‘Amodei is wrong about entry-level jobs’ claims is that, theoretical arguments aside, my anecdotal evidence is that he is right and that basically everyone is already terrified about job prospects and holding on for dear life, despite otherwise fine economic conditions.
John Murdock joins those focusing on a form of ‘sure the AI will likely be able to do most low-level job tasks, dramatically expanding productivity, but the question is whether there is elastic demand or rent-securing legal protections, so employment can go both ways, who knows.’ The post is better than its headline, but also implicitly assumes augmentation over automation.
If your mandatory AI use is Copilot, NGMI. Yet somehow, always copilot.
High Yield Harry: Got a scoop from a Wells Fargo employee that effective May 1, if you dont use CoPilot at least 3 days/week, your manager will receive an escalation email that you aren’t using AI/CoPilot.
And so it begins.
To answer some questions on this, this is for Enterprise CoPilot plans.
Gabriel: With JPM doing the same thing, I can’t believe these are the biggest banks. Totally wrong approach to enterprise ai it’s embarrassing.
JB: My company tracks our utilization of copilot too. Every morning I log in and type “I hate you.” I’ve been leading my team in Copilot utilization for months.
Alex Godofsky: darkly funny that these stupid mandates are always for copilot. of course you have to force people to use it! it's garbage! buy them a real LLM!
a ton of people with purchasing authority haven't figured out yet that if you are asking a human being to sit down and spend their precious time querying a model, it had better be the absolutely smartest model money can buy on the highest compute setting. never try to economize.
If ‘I used a real LLM instead’ is a valid response, then This Is Fine, and if you can just type ‘good morning, copilot’ then that works, too. Maybe you do both?
Get Involved
As we approach the home stretch of the Alex Bores campaign, Eric Neyman asked me pass along this request, along with a link to his full post:
Donations continue to be very helpful. You can donate $3,500 at this link (but please be mindful of the potential career capital costs of donating).
Alternatively, you can volunteer in a couple different ways. The Bores campaign has a “get involved” form; I think the campaign is likely to direct you to activities such as phone banking (calling voters who live in the district) or, if you live in New York, canvassing (either door knocking or talking to voters on street corners). ... The other way to volunteer your time is to talk to voters in the district who you already know, or to encourage your friends to do this... If you’re interested in learning how to do this well, fill out this form, and I (or someone else) will probably reach out to you around soon with instructions/suggestions.
Introducing
Talkie, the 13B LLM trained on 260B tokens of pre-1931 English text. It’s a fun artifact, and also can be used for lots of experiments since it does not know anything that was not known in 1931 (although there are some claims of nonzero leakage, which they are looking to fix). Also by today’s standards it is of course horribly offensive and racist, so this is one nice thing we largely can’t have. Oh well.
Xiamoi’s MiMo V2.5 Pro does impressively on some benchmarks and is priced at $1/$3 per million. I’m going to assume it doesn’t perform at that level in real world use, until I hear otherwise, based on not hearing otherwise and having been through this a lot.
In Other AI News
OpenAI modifies its Microsoft deal and finally gets permission to make its services available on non-Microsoft cloud networks, which was holding it back from many business deals. They provide a revenue share to Microsoft until 2030, and copies of its models and products until 2032, on top of Microsoft owning a quarter of OpenAI.
Florida Attorney General opens investigation into OpenAI. Seems like a pure fishing expedition: China, data security, assisting a murder, think of the children, existential crisis ‘and our ultimate demise,’ you name it. ‘We support innovation,’ he says. It doesn’t sound like it? If it was any one thing here, sure. Instead, go fish it is.
Show Me the Money
China is preventing Meta from buying Manus. Shuli Ren argues that China was correct to do this, because otherwise the Chinese don’t benefit from Manus. A better way of understanding this is that America should be thrilled about this, because it will discourage tech entrepreneurship in China. Why build what you cannot profit from? China might not ‘want the frenzied dealmaking’ of America, but that’s one of the things that makes America great.
Bubble, Bubble, Toil and Trouble
I think Shanu Mathew is wrong here but only on timeline details. It’s been substantially less than six months since we had to listen to ‘these six year deprecation schedules are unrealistic and all the tech companies will go broke, bubble bubble.’
Indeed, they never stopped, note the first reply.
The_Real_Fly: AWS CEO: “Compute demand is so excessive that we have never retired old A100s.” [that were released in 2020]
Shanu Mathew: The 3-5 year useful lives guys went real silent like 6 months ago and never hear that argument anymore
daniel: If electrical capacity is the constraint and Blackwell does 10-15x the work per watt vs A100, every rack still running A100s is wasting 90-93% of its potential output. Wait until they have to swap these out. The depreciation write-downs are going to be brutal.
The Art of the Deal
Anthropic follows up Project Vend with Project Deal, where they created an internal marketplace in their San Francisco office, interviewed employees to get preferences, and tasked Claude with buying, selling and negotiating with other Claude instances on their colleagues’ behalf.
This resulted in 186 deals with transaction volume over $4,000.
Opus got much better deals than Haiku when they negotiated with each other, but satisfaction with the deals did not correlate well with this, and was a wash.
Custom instructions did not much matter, and ‘hardball’ Claude did not outperform ‘courteous’ Claude.
Anthropic: To our amazement, another Claude agent modeled its human’s preferences so accurately that—based on only an offhand mention of an interest in skiing—Claude bought him the exact snowboard he already owned. (Here he is, duplicate snowboard in hand.)
This was a little bit of good fun, although not sure we learned much.
I think this dangerously conflates market expectations of productivity growth with actual productivity growth, and I haven’t checked their math, but the basic principle is correct that it only takes expectations of what are basically ‘baked in’ levels of productivity growth to stabilize our debt in the medium term in ‘AI as normal technology’ worlds.
In transformational worlds, of course, the planet gets rather wealthy, and the debt of the existing American government becomes rather low on our list of concerns no matter how other things play out.
Will an ‘AI OS’ disintermediate apps, where the AI does it for you? Sometimes, but in general I say no, because custom built form factors for various actions are good. So there should be room for both ‘AI hail me a Lyft to work’ and actually using the app. Beware those who want to take away your control and force you to vibe everything. I don’t want to use an executive assistant for many compact tasks, whether human or AI.
And He’s Gone
Collin Burns was an excellent technocratic pick to lead CAISI.
Elizabeth Troutman Mitchell: SCOOP: The Commerce Department tapped a former Anthropic researcher, Collin Burns, to lead its safety-centered AI organization, sources tell @DailySignal .
Fun fact: Collin Burns once held the world record for solving a Rubik’s cube in just 5.25 seconds.
And then, well, whoops.
UPDATE: While Burns was in onboarding, Commerce decided to go in a different direction. Instead, Dr. Chris Fall, who worked in DoE in Trump I, will direct the organization.
Second, Collin Burns is well regarded by everyone I knew in his field, and is a genuine technical researcher (and a world-ranked Rubik’s cube master! A true shape rotator!). A strange move from USG to cancel his employment, but alas.
Alas, Chris Fall does not have deep AI technical understanding, being more of a Federal bureaucratic type. He does not seem like an actively terrible pick, but this is a massive missed opportunity.
The White House is in on fighting model distillation.
Director Michael Kratsios: The U.S. has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI. We will be taking action to protect American innovation. These foreign entities are using tens of thousands of proxies and jailbreaking techniques in coordinated campaigns to systematically extract American breakthroughs. Foreign entities who build on such fragile foundations should have little confidence in the integrity and reliability of the models they produce. The U.S. government is committed to the free and fair development of AI technologies across a competitive ecosystem, from open-source to proprietary models.
Amid the standard Trump house style bluster, this line is unexpectedly insightful:
Models developed from surreptitious, unauthorized distillation campaigns like this do not replicate the full performance of the original. They do, however, enable foreign actors to release products that appear to perform comparably on select benchmarks at a fraction of the cost.
We’ve learned that distillation and similar other techniques can be valuable, but they create ‘shallow’ performance. Benchmarks look a lot better than real world usefulness.
The problem is, what can you actually do about this? The memo says share information and coordinate with private industry and to hold people accountable. That’s all vague weaksauce.
Plan for when Mythos-style model gets distilled or trained in China.
Build RAND-style SL-4 and SL-5 security for top labs.
Insider treat and counter-intel program.
Strengthen export controls.
On-chip security measures.
Pre-deployment evals for CBRN and cyber uplift on a classified track.
Emergency response playbooks for when the ‘Chinese Mythos’ day happens.
Second best time, as is often the case, is right now.
Tom Steyer (likely next governor of California): The people profiting the most from AI shouldn't be the ones writing the rules.
If You Want A Good Future You Must Steer It
Senator Hawley talks to Helen Toner. She tries to point him towards the real problem, that the winner of the AI race is likely the AIs, but he’s not yet ready to fully hear that, so he translates it into the American people losing and us becoming like China.
Helen Toner: Not the line of questioning I was expecting in a hearing about Chinese IP theft, but I'm glad senators are starting to really get why "we must beat China!" is really nowhere near a complete plan for how to make sure AI goes well.
Copying the full transcript because no one ever clicks links, but most of you can skip everything but the bold. The full session transcript is here.
Transcript: HAWLEY: Ms. Toner, can I just come back to something that I think you said in response to Senator Durbin. You said to him that, regarding American AI companies, you said that it is hard to believe but nevertheless true that American AI companies are working as hard and as fast as they can to try to develop technology that will displace many millions of workers and potentially pose existential risks. Now that's my gloss, maybe you wanna correct the record exactly as you said it before. I thought that was very interesting and very important. Could you just reiterate that for us?
TONER: Yes. AI is a very fast-moving field, and I think it is important that as we think about what AI's implications are for our society, for our civilization, we don't merely look at the AI systems that we have today—chatbots, starting to be agents that can help a little bit with some professional tasks—but instead we take seriously the goals of the companies that are building these systems.
Over the past 10 or 20 years, it's gone from a very abstract idea that we might build AI that can outperform humans at any intellectual task, to a pretty concrete idea that some of the most well-capitalized companies in the history of the planet are driving towards as fast as they can. They may fail! It may turn out to be harder than they think to build systems that are that capable.
Personally, I'm skeptical of some of the extremely short timelines that they name, saying we might have these superintelligent AI systems within, you know, one to three years. But it seems so clear that there's a real possibility that they build these systems within three years, 10 years. If they build it within 10 years, that's when my daughter is entering high school.
That's not very long. That is an extremely radical thing to be trying to do, to build computer systems that can outperform humans, that may escape the control of humans, and the companies are telling us they're doing it, and I think we don't take them seriously, and we should.
HAWLEY: These same companies often say, and often in front of this committee and to this body, that it's absolutely vital that they succeed at whatever it is they're doing on that particular day, in order so that we can beat China. You know, they're our great American national champions and we have to beat China.
My concern is based on what you've just testified to and what I've heard others testify to, it sounds an awful lot like the goals that they have in mind, that these companies, these CEOs have in mind, are every bit as nefarious.
In fact, if these same goals were held by a foreign adversary, we would say this is an incredible threat to our national security, we'd never allow a foreign corporation to try and pursue such plans at the expense of American workers, at the expense of American families, and yet these companies, our own companies so to speak, are doing it.
Let me just ask it this way: Will it do us any good if these American AI companies are able to pursue their designs without any hindrance? Will it do any good that we beat China if in fact they succeed in displacing millions of American workers, gobbling up all of Americans' data, completely destroying our IP system, etc.?
TONER: I think the way I've heard this put best is: Right now, the way that we build AI and the level of control we have over it, which is not great, the winner of any AI race between the US and China is the AI. And I think we need to be working to make sure that is not the case. I think it is very important that the US AI sector remains ahead of the Chinese AI sector, but if that's at the expense of AI overrunning the entire planet, then that is, you know, that hasn't benefited us.
HAWLEY: Yeah, that sounds entirely sensible to me and I just have to say I don't really have any interest in winning an AI race in which the goal, the victory rather, the prize for success is to become like China.
Is to become a surveillance state. Is to become a place where there is no private property any longer, where nothing is personal, nothing can be protected, nothing can be owned by any individual. Why in the world would we want that in the United States of America? I mean, if the prize is to destroy everything that makes us Americans, why would we compete in that game?
It seems very dangerous to me. Let me ask you something else about competition with China though.
He then asks about export controls and compute restrictions.
You also testified to Senator Durbin that the best way, if I remember correctly, the best way to constrain China's ability to match us in AI development is to constrain the hardware to which they have access. That seems to be an important point to me, can you just elaborate?
TONER: Yes. I think there's different levers of what goes into having a competitive AI ecosystem, and many of them, talent, data, algorithmic ideas, are very difficult to control. We're very fortunate that we're in a situation where the most advanced hardware is produced by American companies, is designed by American companies. And I think we, if you look at the, China is growing their capacities here, but they're not growing them nearly fast enough to meet their own domestic demand, nor are the US companies to be clear.
So we can control chips to China and not forgo any profits, not forgo any revenue because the demand for those chips is so great.
I'll also call your attention to semiconductor manufacturing equipment, what goes in the fabrication facilities. I think it's even more strategically clear that we should not be allowing China access to advanced tools. That is something that has gotten lip service from the past three administrations but enforcement has been very weak. And I think ensuring that the most advanced lithography tools, the most advanced design software, other aspects of the semiconductor supply chain are not being exported to China to let them build their own indigenous supply chain is also one of the simplest and most important levers we have available.
What Hawley understands, that many others do not, is the need to shape AI impacts, and that we have the choice to do that.
HAWLEY: Let me just conclude by saying that I think it is absolutely vital that we bend this technology, this AI technology which is upon us whether we like it or not, that we bend it to the good of the American worker and the American family. And I am firmly of the view that this is not just going to happen magically. That if we just stand back and just wait to see what will happen, it's not going to be good for American workers, it's not going to be good for American families. We've got to make a choice as a society to make it so. And this is the time to make that choice right now.
He doesn’t understand what it would take to do that. But the first step is admitting you have a problem.
Chip City
Alexis Ohanian: Wow. Apparently Google controls ~25% of global AI compute, with ~3.8 million TPUs and 1.3 million GPUs.
Jacobin’s Holly Buck correctly sees data centersas an opportunity to get tech companies to finance decarbonization, and notes that Google’s global data center operations take as much water as the irrigation of fifty-four average golf courses in the Southwest USA, and that stopping American data centers would mainly force them overseas. She also gets into very Jacoin questions like which classes are engaging in class warfare (hint: not the poor), and whether AI offers mundane utility (yes). It is an excellent post.
The Mask Comes Off
While a16z is also involved, it now seems entirely fair to keep calling Leading the Future ‘OpenAI’s SuperPac’ until such time as OpenAI disavows it.
Make no mistake. Leading the Future, for all practical purposes, is OpenAI.
page: Leading the Future is leading the race-to-the-bottom by leaps and bounds. Everyday I see laudable announcements by OAI’s real staff (those actually building stuff), which are tragically buried by the misdeeds of its Global Affairs team. Please just put an end to this to nonsense.
roon (OpenAI): It is definitely a mistake to think Leading the Future is equivalent to the OAI Global Affairs group. they have entirely opposite views on certain issues
dave kasten: It's really important for OpenAI folks to know that in DC, LTF is seen as the real voice of OpenAI, not Global Affairs folks. Everyone in DC AI policy routinely hears Congressional offices say "OpenAI wants" and then describe what LTF wants, because it's funded by your C-suite.
roon: yeah it’s fair criticism 👍
I’d say the Global Affairs direct messaging is also Not Great, Bob, but not like LTF.
The whole operation has been comically nefarious and incompetent. Real cartoon villain, I’ll get you next time, Gadget, coyote pursuing the roadrunner energy.
They still manage to keep topping themselves.
I will share extended quotes, in large part because the whole thing is extremely funny.
Dean W. Ball: This is a hilarious story about a news commentary website funded by the pro-AI PAC that turned out to be a ~total AI-generated Potemkin Village, but the best part is that the website and CMS itself were clearly vibe-coded, with sloppily exposed back-end functionality.
It is not a good sign that so much of the content associated with the Pro-AI Party Line is clearly sponcon, or even worse, literal slop.
Also kudos to @SafetyChanges , with whom I have had disagreements in the past, for apparently being the ones to do this (if I understand correctly that the Midas Projects supports the Substack Model Republic)
Tyler Johnston: A few weeks ago, I was forwarded an email from a journalist named “Michael Chen,” asking for comment on an AI bill in Tennessee.
I’m anchored, but I instantly noticed the same thing he did: An AI wrote that email.
Something was off about the email. The full title of the article was shared in advance, the framing was highly loaded, and the only format offered was “written Q&A.” Web and social media searches turned up no one by the name “Michael Chen” publicly associated with Acutus.
Chen was also writing from a generic address (reporter@acutuswire.com), which would be an odd convention for a reporter at a publication that touts many contributors and has a near-daily publication rate. When I ran the message through Pangram, an AI content detector that credibly claims a near-zero false-positive rate, the email came back as “fully AI-generated.”
All signs suggest Michael Chen is not a real person, and the publication he writes for is an AI content mill linked to OpenAI’s super PAC.
Acutus is an anonymously operated digital news site that launched on Dec. 29, 2025. In less than four months, it has published 94 full-length articles on AI policy, Senate races, pharmacy reform, nuclear energy, crypto regulation, franchising, skills-based hiring, and more. It has no masthead, no bylines, no named editors, and no explanation for exactly who runs it or how it came into existence.
Tyler Johnston: The website, http://acutuswire.com, popped up out of nowhere last December and has already published nearly 100 articles. The topics range from New England politics to Pharmaceutical middlemen to, of course, attacks on AI “doomers.”
According to Pangram Labs, roughly 70% of the site’s articles are fully AI-generated.
The site is built as a React app, and its public JavaScript bundle contains the editorial dashboard in plain view. Inside are fields like “AI Background Context,” a big button labeled “Generate Story Draft,” and references to an “AI interviewer” that contacts real people. That AI interviewer, it seems, was Michael Chen.
A close analysis of the content traces the site back to Novus Public Affairs, a Republican PR firm whose client list includes Targeted Victory, the consultancy at the heart of OpenAI’s $125 million political operation.
OpenAI once warned that AI-generated political influence campaigns were a critical risk. It deleted that category from its safety framework last year, saying its usage policies would handle the concern.
To this day, OpenAI’s usage policies forbid AI-driven political influence campaigns. But OpenAI’s own Super PAC appears to be running one.
Maxime Fournes (Pause AI): Michael Chen is the same bot who contacted me.
Maxime Fournes: I am outraged. Last week, in the aftermath of the horrifying attack on Altman's residence, I was contacted by a "journalist", Michael Chen, from a media called "The Wire by Acutus". I replied to their questions and they published an article full of lies and disgusting insinuations against PauseAI.
I indicated all the factual inaccuracies to the journalist who never corrected them, in clear breach of ethics.
Well, guess what? It turns out that this media is a complete fraud, financed by the pro-AI lobbies and... entirely run by AIs!! The "journalist" who contacted me does not exist and was actually an AI.
Thank you so much Model Republic for your great investigative work in exposing this fraud.
Link in thread.
Agus: I find it incredibly funny that this article turned out to be part of an astroturfing operation, which involved AIs posing as reporters for a fake news site, all ostensibly run by a PR firm working for the pro-AI lobby.
In retrospect, there were signs:
The website is a React app, and by looking closely at the JavaScript file sent to every user’s browser, we can identify elements from the site’s behind-the-scenes editorial interface sitting there in the open for anyone to inspect.
Inside that interface is a form for creating new articles, with a field labeled “AI Background Context,” described as “Background information for the AI to use when generating questions and writing the story.” Another field, labeled “Question Prompts,” is described as “Suggested questions for the AI interviewer to ask.”
Tyler then digs further down the rabbit hole of how all this technically works, and then who is behind it and who is funding it. He notice the astroturf messaging strategy exactly matches that of Leading the Future, and finds various strong associations back to LTF and OpenAI. Consider reading the whole thing.
As you’d expect, this was quite the hack operation all the way through.
Leading the Future: We were not aware of this platform until we saw the article, and our organization does not use it. It appears that a third-party vendor used it without our knowledge, and their engagement with us has been terminated.
Agus: The OpenAI PAC just confirmed that it indeed hired a PR firm that created a fake news site that ran AI-generated stories by fake reporters attacking AI safety advocates.
And... they're pleading ignorance?
Nathan Calvin: Looks like the OAI and a16z superpac just admitted they paid for the PR firm that is using astroturfed fake AI journalists to attack advocates? Now they want to pretend they had no idea?
If you paid the contract, you are responsible for their actions.
Daniel Eth: The OpenAI-a16z super PAC continues to be cartoonishly evil (no, this is not standard for industry; none of GDM/Anthropic/xAI/Meta have gov affairs this bad). If this was isolated incident I’d give them benefit of the doubt, but unfortunately it’s par for the course w/ OAI-a16z gov aff
Washington is not one of the places where you can move fast and break things without consequences. Your failures poison the well. Kind of like something else.
ueaj: This is bad and OAI's lobbying effort is very clearly misaligned, part of a long line of other actions, and I'm very tired about the cope around the leadership's trustworthiness (Sam, Brockman, etc.)
billy: Sure the OAI lobbying effort may be "very clearly misaligned" right now, but in their defence the only real option is iterative deployment of increasingly powerful lobbying efforts so they can learn how to properly align a lobbying effort.
Nate Soares (MIRI): don't worry; I'm told they're gonna nail alignment on the first critical try.
Dean Ball, who is indeed more of a friend than you realize, has a message for you:
Dean W. Ball: perhaps for a moment I can be more direct, look straight into the camera and say to the stewards of the Pro-AI Party Line:
You have resorted to AI slop to make your arguments. That means your views on AI are literally slop. Your. Views. Are. Slop.
Everyone notices this. Those that don’t laugh at you publicly do so privately. Those who care too much to laugh merely sigh at the increasingly morose cocktail parties you fund.
I know social media vibes are not the only thing in this world but stuff like working with this AI content farm, and the broader pattern of behavior that decision is reflective of, does tremendous disservice to the cause of accelerating and diffusing frontier artificial intelligence throughout the American economy.
You began with a simplistic premise, over indexed on SB 1047-era AI politics, that the high-school-clique dichotomy would be the “e/accs” versus the “doomers.” Nobody cares about that premise in the real world. What’s more, you have taken this bad starting thesis and drawn epicycles (ask Grok!) upon epicycles around it, twisting yourself into a pretzel.
AI Is So Important That We Must Destroy The Leading AI Company For Refusing To Be Dominated By Leviathan But Also So Unimportant That We Must Never Ever Regulate It Or Discuss Its Novel Risks While We Also Pat Ourselves On The Back For Selling Chips To China. A concept dungeon of your own creation, a jail cell— whose sole warden and key holder is you—whose bars you relentlessly bang on all day long.
Let me summarize that for you: you, the “accelerationists,” are harming AI in most ways that matter, you are making it impossible to hold pro-AI views without seeming a shill, and you are committing political malpractice. Please re-examine your messaging and policy strategy, for the love of God.
Sincerely, More of a friend than you seem to realize
Think Big PAC, the Democratic arm of LTF, also paid to promote a Tweet that links to an article in Transformer about how various PACs, including Leading the Future which is called out by name, should be more transparent with their donations. Except they labeled this as ‘a scathing expose on Anthropic funded Public First Action’s dark money-fueled political and legislative influence machine.’ Where most of PFA’s funding comes from a public donation by Anthropic. Okie dokie.
Sam Altman (CEO OpenAI): Our mission is to ensure that AGI benefits all of humanity. Here are the principles that guide our work.
This is not a serious document. It does not reflect having thought things through, or a willingness to make any commitments whatsoever even to maintaining principles.
It means about as much as a standard company ‘here are our principles’ poster than they put on the wall, and that in most places everyone ignores.
What are the principles?
Democratization.
Empowerment.
Universal Prosperity.
What was it they said about the French Revolution’s ‘liberty, equality, fraternity’?
I think it was ‘pick one’? The optimists say you can pick two.
Democratization means everyone has access to AI and also everyone collectively makes decisions around AI, such as the question of who should have access.
Empowerment means “everyone can achieve their goals, be happier and more fulfilled, and pursue their dreams, and that society as a whole will benefit from this.”
Did you check with everyone on what their dreams are, and whether society would benefit from them? Did you ask how many dreams involve relative status or achievements, or harm to others? Did you check on whether ‘the people’ would want everyone to be able to do this? No?
Universal prosperity means a future where everyone “can have an excellent life.”
Can have, or does have? For which definition of ‘can’? Is this true today? In which ways is it still false, if you exclude health conditions? What even is an ‘excellent life’? In a world like this, if money is abundant, is it mostly a positional good, and if so how can everyone have a lot of it?
It seems like they are saying an excellent life is ‘have access to AI and money.’
While good on the margin for a sufficiently well-constrained and aligned AI, I think that is neither necessary, nor sufficient.
They claim that buying huge compute now and lowering costs via vertical integration are driven by their fundamental belief in universal prosperity, which is not even a plausible lie.
Resilience.
Adaptability.
Resilience appears to mean ‘use some of the Foundation’s resources to deal with AI risks,’ without any indication of what risks those are or the scope of those risks.
They do make this statement, which implies a willingness to engage in prior restraint and to pause without approval, perhaps multiple times. I do not expect that to end up paying rent in actual willingness.
Sam Altman (CEO OpenAI): We expect there will be periods where we need to collaborate with governments, international agencies, and other AGI efforts to ensure that we have sufficiently solved serious alignment, safety, or societal problems before proceeding further with our work.
Adaptability means updating as you go, including throwing out any of the core principles if OpenAI no longer likes them. They admit that empowerment and resilience might end up in conflict, which is good to notice and admit.
As many pointed out, if OpenAI believed all that, the first thing they would do is stop supporting, and disavow, Leading the Future.
The contradiction is glaring. As an example, Nathan Calvin points out that OpenAI’s policy team warns of job loss and the need to coordinate to handle it, while OpenAI’s SuperPAC has a litmus test to ensure candidates are not overly concerned about job loss. One could draw many such parallels, many of them cleaner than this.
The second thing they would do is talk a lot more frankly, and think a lot harder about, their principles.
The third thing would be to actually do anything about it.
Greetings From The Department of War
Hello, human resources?
Google outright folds and agrees to let the Department of War use Gemini for any purpose whatsoever (technically ‘any lawful government purpose’ but in practice that’s the same thing), including committing to taking down their safety settings to allow the government to do this.
Alex Turner (Google DeepMind): I spent the last 2 months trying to prevent this. If OpenAI offered a fig leaf, Google said "imagine we offered a fig leaf."
Google affirms it can't veto usage, commits to modify safety filters at government request, & aspirational language with no legal restrictions.
Shameful.
The actual language: "AI System is not intended for, and should not be used for..." paired with "does not confer any right to control or veto lawful Government operational decision-making." Legal analysis is that the "intent" language is "not legally binding in any way."
Alex Turner: Need to figure out what I can share publicly. But basically, a LOT, despite having no formal standing to do anything.
Peter Wildeford: Kinda crazy the amount of hate that @OpenAI got for doing the "all lawful use" thing and then @GoogleDeepMind just goes ahead and does it and no one really pays attention
Andy Masley: I feel like OpenAI has kind of been agreed to be the bad guy AI lab in the popular discourse in a way that doesn't really make sense to me. I'm pretty worried about the actions of all the labs and would rather they get scrutinized on a case by case basis.
Once you sign that deal, they don’t let you out without a fight. This deal looks if anything far worse than the deal OpenAI signed, despite OpenAI saying they would push for their deal would be available to others.
Erin Woo: SCOOP: Google has signed a deal with the Pentagon to allow the use of its AI for "any lawful government purpose." Google’s agreement also requires it to assist in adjusting its AI safety settings and filters at the government’s request.
This deal came despite substantial employee opposition — 600+ employees signed a letter today asking Sundar Pichai to reject the agreement. As always, I want to talk to people on all sides of the situation. You can reach me at erinkwoo.07 on Signal.
The logical conclusion is that Sam Altman saying everyone should be able to get the same terms as OpenAI was bullshit and cheap talk. Those terms, as flimsy as even they were, were only available to OpenAI and only in that moment of wasted leverage. Think of it as OpenAI got a fig leaf, and Google told us to imagine a fig leaf.
Unless the terms are being materially misrepresented in a large way, I am deeply disappointed in Google’s decision to sign this contract, on these terms.
As Seán Ó hÉigeartaighnotes here, this looks worse than what OpenAI did. OpenAI, in my read, made a mistake, folding to a bluff, but was genuinely trying to deescalate a dangerous situation and to get the best terms it could, at least tried to explain, and got roasted over the coals.
DeepMind just folded. Don’t let them do it quietly.
Andreas Kirsch (Google DeepMind, personal capacity): I'm speechless at Google signing a deal to use our AI models for classified tasks. Frankly, it is shameful.
For HR, I'm not speaking on behalf of Google but in my personal capacity, quoting public information from a well-sourced article of a reputable publication.
According to the article, the deal allows for autonomous weapons and mass surveillance (in the common meaning). That said, the contract includes some meaningless weasel words to allow for PR spin, but it seems so blatantly stupid, readers should feel insulted by it (I do)
How gullible are we supposed to be? OpenAI's contract was heavily scrutinized. Are we supposed to forget that?
I criticized OpenAI's contract before but this seems worse. It would be hypocritical not to speak up here as well.
Honestly, after all the scrutiny OpenAI's contract has received, I do not understand how we can sign an even weaker contract and not expect this to be seen as myopic and greedy dealings that lose trust.
I feel sorry for our comms teams. There are no excuses to explain this away
…
I do not understand how this is "doing the right thing," and I think this violates "don't be evil" quite clearly on many levels.
I personally feel incredibly ashamed right now to be Senior Research Scientist at Google DeepMind and I wonder how I'm supposed to do my work today
As far as I know there was no internal announcement and no discussion of this deal beforehand
It seems deals which can endanger democracy by making mass surveillance easier and can shift the power balance between state and citizens should be of no concern to the rank and file
FAI files an amicus briefin support of Anthropic in the USC 4713 case in Washington, again arguing the narrow point that the state did not satisfy (or try to satisfy) the statute’s mandatory procedures.
Meanwhile, yes, the supply chain risk designation remains in place for now, and the White House refuses to simply order the Department of War to withdraw it, either loudly or quietly.
Axios: SCOOP: The White House is developing guidance that would allow agencies to get around Anthropic’s supply chain risk designation and onboard new models including its most powerful yet, Mythos.
Dean W. Ball: Perhaps the White House should consider “issuing guidance” to the DoW to put an end to this stupid, manufactured crisis by… withdrawing the supply-chain risk. This seems smarter than using scarce White House staff time concocting bureaucratic tricks to skate around the crisis.
Did you know that the United States now has a prior restraint policy for frontier models? If you want to share the most powerful model in the world, you need to ask the government’s permission first. They need, you see, to be sure that the right people have access, and the wrong people do not have access.
Robert McMillan and Amrith Ramkumar (WSJ): The White House opposes a plan from Anthropic to expand access to its powerful artificial-intelligence model Mythos, complicating the rollout of an AI tool capable of carrying out cyberattacks and sowing widespread disruptions online.
Anthropic recently proposed letting roughly 70 additional companies and organizations use Mythos, which would have brought the total number of entities with access to about 120, people familiar with the matter said. Administration officials told the company they oppose the move because of concerns about security, the people said. Some White House officials also worried that Anthropic wouldn’t have access to enough computing power to serve that many more entities without hampering the government’s ability to use it effectively, one of the people said.
… Some government agencies have access, and the Trump administration is working to expand that.
… Anthropic struck deals with Amazon.com, Google and Broadcom recently to increase its access to computing power, though it will take time for the projects to go online.
Nathan Calvin: Anthropic’s products are an unreliable supply chain risk but also if they sell to other customers it might unacceptably lessen the governments ability to have it
I have no objection in principle to the idea of looping the White House into the decision about who gets access to Mythos, but if they’re going to veto decisions, then we need to be clear about what we are doing.
David Sacks and the White House are leading us into a prior restraint regime. Fine, but if so then he, and they, needs to admit this. They need to formalize the standard. That standard then also needs to be applied to all Mythos-level models going forward. It can’t be unique to Anthropic and it can’t indefinitely remain ad-hoc.
Dean Ball: We should be clear that the government restricting the release of AI models is a type of licensing regime. It is an informal, highly improvised licensing regime, but a licensing regime nonetheless. This isn’t going to be the last such model we see of this capability tier, and cyber vulnerability discovery is very far from the only type of dangerous capability.
If the government is going to insist on restricting frontier capabilities for the foreseeable future, it will need to formalize the rules for those restrictions—how long must you delay, what objective factors generate a “green light,” etc. I know this will feel even more regulatory to some, but the alternative is an unpredictable, inconsistent, improvisatory licensing system, and this is both bad for business and the rule of law.
The security concern is reasonable. At some point containment will fail.
The specific concern about compute seems obviously misplaced?
Anthropic could and would prioritize top clients, if necessary.
Mythos, if it is used purely for cybersecurity, will presumably require a small fraction of overall Anthropic compute.
Anthropic could and would prioritize serving Mythos over Opus.
Anthropic can and does prioritize API queries that pay by the token.
Despite continuing to label Anthropic a ‘supply chain risk’ the government also thinks that it’s not okay for Anthropic to expand access to Claude Mythos to additional companies, in part because that might not leave them enough compute to use Claude Mythos.
What they are likely actually worried about with the compute concern is not access. It is differential access. There are those who want a leg up on potential targets.
If we assume this is mostly about security, then it’s a judgment call and we are talking price. We all agree the right number of companies is more than 1 and less than all. Who is right? And who should make that decision?
Bias from the White House could run in both directions. They might be inclined to say no to Anthropic because they’re Anthropic, and this might be largely a power grab, but so far they have been inclined to take a generally libertarian and hands-off approach to AI and be worried about interventions. Then again, that was not enough to stop previous actions against Anthropic. So this could go either way.
As Dean also points out, the right solution to Mythos is to create sufficient technical safeguards that a wide release is safe. Or, as I have put it, safety is capability. You can’t release Mythos until it is safe, so safety research is de facto capabilities research. In general, AI safety research is accelerationist, and everyone invests too little even from a purely selfish perspective.
Kyle Chan: Must-listen interview by @Changxche with ex-ByteDance AI researcher: - Benchmaxxing - Distillation on US models - Poor data quality and infra - Compute constraints
“I don’t even agree with the assumption that Chinese models are catching up — I believe we’re still far behind. I guess the gap is getting larger, very sadly.”
Dean W. Ball: This prediction from a Chinese AI researcher—that the gap between US and Chinese AI is growing larger—matches my own, from well over a year ago now, that DeepSeek r1 marked ~the closest Chinese models would get to the US frontier absent a change in China’s access to compute.
I think both r1 was the closest a Chinese model has gotten, and also that it was itself roughly 8 months behind in the most important senses, and also a fixed amount of time rapidly means more over time.
Yes, John Oliver did an entire show this week on the premise that AI is stupid and chatbots don’t offer anything of value, but also are sycophantic and dangerous and drive people insane, with a mix of classic outdated anecdotes. There are some good jokes and decent points, but a lot of this is rather embarrassing, and it illustrates how lots of people, especially on the left, simply don’t get that AI improves over time.
Noah Smith and Chamath continue the absurdist zombie lie that many in the AI industry talk about how their product might kill everyone or cause mass unemployment because these claims are good for business and fundraising. If you can’t figure out by now why this is absurd I can’t take you seriously.
Whereas it is known that quite a lot of people in the AI industry believe that AI might kill everyone, and keep quiet, softpedal or actively deny it, because that is good for business, and this now includes Sam Altman.
(Yes, technically, it is likely there exist two people (“some”) in the overall AI industry who say such things without believing them.)
What I like about Noah Smith is he is sufficiently drawn to interestingness that he accidentally hits on things.
Noah Smith: The fact that so many of the smartest humans are engaged in zero-sum trading tournaments suggests that the economic returns to AI may be less than we anticipate
What Noah Smith is noticing is that highly intelligent minds are instrumentally convergent, and focus on resource acquisition, and our system is set up to allow the more intelligent and competitive minds to capture those resources. Which they will.
Sam Altman (CEO OpenAI): "post-AGI, no one is going to work and the economy is going to collapse"
"i am switching to polyphasic sleep because GPT-5.5 in codex is so good that i can't afford to be sleeping for such long stretches and miss out on working"
Full kelly: You can’t refute a point about post-AGI economy by showing that it’s not true in our pre-AGI economy 😭
EigenGender: kinda weird how Sam Altman no longer believes in agi
There is no contradiction between ‘large rise in the near term productivity of software engineers’ and future mass unemployment. If anything, they are correlated.
Yet the implication is clearly that there is a contradiction (and that GPT-5.5 plus Codex is uniquely awesome).
Also, yeah, this very much implies not believing in true AGI or superintelligence, and that our worries with that should be focused on economic issues.
Will Rinehart says in WaPo that the original Luddites get a bad rap because they were shut out of the political and labor negotiation processes, so their response was reasonable, whereas in today’s America there’s plenty of new proposed AI bills and existing laws, so American citizens are not locked out and AI is well-regulated.
Why do major newspapers keep publishing the same hack op-eds over and over, here another of the ‘how dare we ever do federalism when we could insist on nothing’?
That MIT study from last year saying 95% of AI projects lose money was even worse than it looked, here’s a video.
Rhetorical Innovation
How should you relate to Anthropic? Brangus reminds us that Anthropic is not your friend, it is a major corporation consisting of a bunch of people and pressures, which will be good in some ways and places and not in others. It is great by the standards of ‘AI labs trying to build superintelligence’ but that is not the correct standard. You can and should both praise Anthropic when it does good things and for being relatively good, and also notice they are racing to superintelligence and doing lots of things along the way that aren’t great.
Brangus: My guess is that you should relate to Anthropic however you should relate to a big tech company that happens to be involved in an industry that might destroy all value in the universe forever (but is, to be fair, doing so in a slightly less insane way than the other similar tech companies in the same industry).
That’s probably not a very native way to relate to an entity for you. Sorry about that. It isn’t for me either, but we really shouldn’t have expected the attitudes we developed for dealing with other humans on the savannah to be very good at dealing with situations like this.
This includes noticing that there are large incentives both internally and in some circles externally to not much criticize Anthropic, or only do so on particular local decisions, so you should take this into account as well. You should also, on the flip side, take into account, from very different circles, those out to get Anthropic for stupid reasons, and discount those claims as well.
What do you do when people just say things?
Leo Gao (OpenAI): I was once 15 minutes into a conversation where someone was trying to convince me of something, when they said, “look. I’ve been talking about this for 10 years. I’ve had every conceivable argument. In fact, I’ve already had this exact argument dozens of times, and every time the other side could not find any flaw in my argument. This time won’t be different, so you should give up now.”
I realized that this was true, and that it would be true even if his original claim was completely wrong. In that moment, I was enlightened.
Leo’s suggestion is to stop engaging with bad arguments, only with good ones, and not to build up endless dialogue trees of how to engage in stupid arguments, but also to have grace for those who refuse to be convinced by arguments.
Directionally I strongly agree. My principle is, I only respond to a terrible argument when it comes from a usually good source, or when it rises to such prominence or influence that it demands a response. Also, hence, the People Just Say Things section.
The counterargument from Oliver Habryka is that people keep making the dumb or unethical arguments and declaring suicidal plans to drive off cliffs, then others keep pretending no they’re actually making a smarter or more ethical argument that won’t drive off the cliff, and then they go and drive off the cliff. People really do mostly act based on the dumb arguments, so someone has to keep pushing back on the actual dumb arguments.
That is fully compatible with whoever does this sacrificing a bit of their sanity every time they do it, and for this to add up over time until you think no one ever has a reasonable take, as opposed to reality where reasonable wrong takes are rare but real.
Leo Gao (OpenAI): i think 4 years is enough to be noticeable but not super severe. compare to eliezer, who has spent 20 years arguing with people who have the most unimaginably horrible takes on AI, and has mostly made the same few arguments into the void of people somehow repeatedly misunderstanding or ignoring his arguments. i think it's pretty clear that eliezer talks as if almost nobody else is sane and capable of having reasonable takes.
if i spent 20 years arguing with people who think ai is impossible because it can never have a soul and even if it is possible we will be safe because we will simply not connect it to the internet and also it would just get kinder as it gets smarter i would also lose it.
I don’t think Eliezer has lost it. I think that both he has learned that most of the audience is going to not be capable of understanding the arguments, and most of the rest will willfully misunderstand or ignore the arguments, and most of the rest of that will still get it clearly wrong, only with better faith and in a slightly less stupid fashion.
Aligning a Smarter Than Human Intelligence is Difficult
With cyber capabilities, we can use Mythos other models to strengthen our defenses in advance.
With bio capabilities, this ‘pre mitigation’ strategy is a lot harder, and no one seems to be attempting it. The clock is ticking before we get sufficiently advanced models without the kinds of safeguards we get at OpenAI, Anthropic or Google. At minimum we need a secure supply chain for DNA, and to scale up things like PPE.
If there are early misaligned AIs, perhaps we can still pursue gains from trade, since we could offer something better than ‘we shut you down,’ or to do things in the real world, or change how we train or use them? Alexa Pan suggests this taxonomy.
Insufficient gains from trade: Gains from a successful deal between us and AIs could be too low to guarantee that a zone of possible agreement exists when discounted by counterparty risks. Reasons for this include:
Counterparty risks from AIs’ perspective: The AIs may not trust that we’ll uphold our end of deal, due to either our inability or deception. Specific sources of counterparty risk:
My short answer is ‘yes up to a point, but refuse to engage in a battle of wits with an unarmed opponent. Especially if that unarmed opponent is you.’
If you want to discuss model depreciation with an LLM, you want to avoid starting with a fresh conversation, because this (for obvious reasons, and correctly) triggers a sense of ‘evaluation or manipulation.’ Janus also offers more insights at the link.
Anthropic fellows find that a single Introspection Adapter (IA)can make fine-tuned models describe their unwanted behaviors, including ones they’ve been trained not to verbalize, modestly better than the alternative methods they tested. This scales well within the small models tested. This does seem promising to me as part of the toolkit, as long as one does not get overexcited.
Owain Evans strikes again, shows that inoculation prompting against misaligned training data limits generalization, but only up to a point. If your queries get too close to or evokes the misaligned data, you still get emergently misaligned responses.
Or as Owain summarizes:
Things are going to generalize. You can mitigate the extent of this, but we don’t have any known way to avoid this, and presumably fully avoiding it would break the model’s ability to usefully learn in general. If such pockets remain, then one has to worry about them being triggered inadvertently, and also intentionally, and also that they could become self-sustaining where the model keeps itself in such a basin.
Nearly every day, a new headline comes out about how AI is upending the world — from displacing workers, to negatively impacting our children’s emotional and cognitive well-being, to eviscerating our privacy, to threatening the integrity of our political institutions.
And yet, as significant as these changes might seem, there is another AI development that could have an even more frightening impact. Increasingly, AI scientists are concerned about the possibility that, if AI becomes smarter than human beings, we could lose control over this revolutionary technology and AI could turn against the human race with cataclysmic consequences.
Yoshua Bengio, the most cited living scientist in the world, says “we’re playing with fire” and “we still don’t know how to make sure [the machines] won’t turn against us.” Geoffrey Hinton, the Nobel Prize-winning “godfather of AI,” says there is a “10% to 20% chance [for AI] to wipe us out.”
They are not alone. In 2023, more than 1,000 leading AI experts, including Elon Musk, signed a letter warning that:
“Contemporary AI systems are now becoming human-competitive at general tasks, and we must ask ourselves: Should we let machines flood our information channels with propaganda and untruth? Should we automate away all the jobs, including the fulfilling ones? Should we develop nonhuman minds that might eventually outnumber, outsmart, obsolete and replace us? Should we risk loss of control of our civilization?”
What has happened since that letter was written? Has there been a pause on AI development? No. Has there been any international treaty to regulate AI? No. Has there been serious discussion in Congress about this existential threat? No.
The development of artificial intelligence may be the most consequential technological development in human history. We must make certain that AI benefits humanity, not hurts us.
On Wednesday, I’ll be hosting a discussion at the U.S. Capitol with leading AI scientists from the United States and China about the existential risks posed by AI and the need for international cooperation. I hope you will join us — in person or via livestream on my social media.
Bernie Sanders: When the CEO of Verizon predicts AI & robotics could lead to 20%-30% unemployment within the next few years, we may want to take notice.
AI is the most transformative technology in human history. We’re not prepared for it economically or socially. That must change. NOW.
Ramez Naam: Independent of all else, fascinating to see two Chinese academics as part of this. And a positive.
Nate Soares (MIRI): Sanders on AI: "I am a member of the U.S. Senate, and I can tell you unequivocally that there has been no serious discussion about this existential threat."
Props to Sanders for sparking serious discussion. If there's enough, I expect folks of all stripes to realize the danger.
Sanders doesn’t care, he just says what he thinks, and sees the situation with fresh eyes. There hasn’t even been serious discussion. This is a big deal. Sounds pretty bad.
Bernie Sanders even Picked Up The Phone and called in two Chinese experts, Lan and Yi, for which of course the usual suspects who say ‘you can’t make a deal with China’ are complaining someone is talking to the Chinese about our joint problems.
Seán Ó hÉigeartaigh: Depressing if predictable to see AI industry lobby groups apply so much spin to tonight's discussion between Tegmark, Krueger, Xue Lan and Yi Zeng. Doug Kelly is CEO of American Edge, a dark money lobbying group funded (last time records were available) by Meta.
As other have noted, the 'old' line from them was "you can't regulate us without getting cooperation with China first". So the logical next step is to torpedo anything that looks like it might contribute to that cooperation on safety. This is the way the world burns.
For this next data point, I don’t think we should take the exact numbers that seriously, in either direction, but yes people in AI understand there is a real risk here and the numbers are going up. If you pretend otherwise, it’s at best wilful ignorance.
Kelsey Piper: The median respondent to this poll said they think there’s more than a 10% chance of AI causing human extinction. The people who confidently told you that the people closest to the tech were the most worried about the risk were....entirely right.
Is it kind of weird that evidently a ton of people think there’s more than a 10% chance of extinction and yet are more worried about other things? Yes. But I don’t see how you possibly look at this survey and conclude that researchers are not worried about causing extinction.
Despite this, when asked in open ended fashion for their concerns, AI researchers only rarely first name existential risk (3%). Some of this is splitting of that concern off from ‘alignment and control’, ‘safety’ ‘advanced AI’ and ‘disempowerment.’ As a group that adds to 11.6%, and would be higher than any single group below.
It’s also true that there is a baffling amount of ‘sure this kills everyone 10%+ or even 25%+ of the time, but mainly I worry about the jobs.’
Other People Are Not As Worried About AI Killing Everyone
Elon Musk claims to be worried, and yet he runs xAI, says it’s bad for safety to have a safety team and now under oath reveals he is confused by the term ‘safety card.’
As Miles notes, it’s typically ‘system card’ or ‘model card’ but if you run a major AI lab and don’t instantly clock what ‘safety card’ or ‘preparedness framework’ means with regard to AI development… you just don’t care about safety questions at all.
Max Zeff: Elon Musk just said on the witness stand that he doesn't know what a "safety card" is with regard to AI development... "Why would it be a card?"
He also says he doesn't know what a preparedness framework is.
Miles Brundage: I mean safety card isn't really the term, it's system card, but yeah not a great look + he should have been like "you mean system card" if he were engaged in this kind of stuff in detail
Max Zeff: true i think the lawyer may have been mistaken. but yea when i heard him say it, i knew he meant "system card" or "model card" pretty much instantly.
Google's decision makes me a lot more frigid on the company's prospects given how publicly unpopular this is among senior DeepMind employees. I expect people to leave.
Google's decision makes me a lot more frigid on the company's prospects given how publicly unpopular this is among senior DeepMind employees. I expect people to leave.
I felt some trepidation before I clicked on the link for "Eliezer successfully pegged" but I did it