dcreelman (@dcreelman) header
dcreelman (@dcreelman) profile picture

Living at the frontiers of managing talent and culture

Toronto, Ontario
Joined November 2007
Photo by dcreelman (@dcreelman)Photo by dcreelman (@dcreelman)Photo by dcreelman (@dcreelman)Photo by dcreelman (@dcreelman)Photo by dcreelman (@dcreelman)
dcreelman retweeted
Same three men. Same room. Smaller table. Different conversation?
🤖 Made with AI
1
1
104
dcreelman retweeted
What can all the middle powers do about AI right now? Kokotajlo has a good idea:
24
99
9
727
37,654
🟧 For the first time ever, Sentinel has issued an orange alert, due to the METR report. Sentinel (@xriskfyi) is a weekly report, written by the world's top forecasters, that monitors global risks.
"This trajectory, if continued, means a world-changing intelligence explosion in 2027, so the public needs to know." Every chart right now is a screaming red alert.
19
119
13
839
74,758
METR's Ajeya Cotra: "This is the most important update we have gotten about AI risk in the last couple of years — much bigger than the Hugging Face attack." "Much shorter timelines than most researchers (including AI 2027) projected.” --- "Fucking terrifying." "Look at this graph. It’s an exponential *on a log scale*. I pray and hope there will be strongly diminishing returns before long in this process. But leaders at METR and Anthropic both seem to think there might not be." - @JeffLadish "I feel sick after reading today’s METR report. I’m worried for my friends and family. I don’t know what evidence people will need to see before they start taking these risks seriously. If you’re undecided about helping, please act; there's many opportunities to help." - @danesherbs (ex-OpenAI)
The authors of the METR AI R&D report estimate that AI is nearly doubling the rate of capabilities progress. This implies dramatic *uplift*: as if researchers experienced time ~10x faster. You probably couldn’t achieve this even if you hired every qualified human in the world.
28
57
8
413
27,981
dcreelman retweeted
I am a broken record about this, but: frontier agents are not going to be a purely digital phenomenon for much longer. They will be able to use anything in the physical world that can be actuated programmatically—first competently, then at human-expert level, then superhuman. By and large, we probably won’t need specialist models; the frontier models will just do it. If your model of the future hinges on physical AI happening “for real” many years into the future, I think you will be in for a surprise.
i went to bed and asked claude to paint an autumn tree overnight. at around 4am, it realised that it didn't have some shades -- so it stopped and titrated its own paint: diluting red, sienna and madder step by step on a second palette, then measuring each pocket from a photo. woke up to this beautiful watercolor painting. 303 brush marks, took about 4.5 hours 🍂 (still working on the speed)
88
177
26
1,814
118,425
Many of you don't understand how serious this is, so let me break it down. In computer science there is a famous kind of puzzle called 3-SAT. You are given a long list of yes-or-no switches and a set of rules about how they must be arranged, and you have to find a combination that satisfies every rule. The obvious way to solve it is to try every combination, but each new switch doubles the number of possibilities, so a puzzle with a few hundred switches would take longer than the age of the universe to brute force. For decades, the smartest people in the field found clever tricks that made this a little faster, but every trick still hit the same wall: the time required kept exploding as the puzzle grew. About 25 years ago, two computer scientists, Russell Impagliazzo and Ramamohan Paturi, working with Francis Zane, turned that wall into a formal idea. They proposed the Exponential Time Hypothesis, which says this explosion is not a failure of human cleverness but a law of nature. No algorithm, however brilliant, will ever solve 3-SAT without the time blowing up exponentially. Nobody could prove it, but almost everyone believed it, and it became one of the foundations of modern computer science. Researchers used it as a measuring stick. If you could show that solving some other problem quickly would also crack 3-SAT quickly, you had shown that problem was hopelessly hard too. Hundreds of papers, across scheduling, graphs, logic and much more, rest on this one assumption. Now an internal AI model at OpenAI claims to have broken it. It says it found an algorithm that solves these puzzles meaningfully faster than the hypothesis allows, and it has checked the proof in a computer verification system called Lean. And here is the remarkable part: the very people who proposed the hypothesis, Impagliazzo and Paturi, along with MIT's Ryan Williams, are now leading the public review of whether a machine has disproved their own idea. So what happens next? If the disproof holds, a large part of theoretical computer science has to be revisited. Every result that said "this problem cannot be solved faster, assuming ETH" loses its foundation, and researchers will have to work out which of those walls were real and which were simply where we stopped looking. It does not mean everything becomes easy. The new algorithm is still slow in absolute terms and this is not the famous P versus NP problem being solved. But it means our map of what is possible was wrong in a place we were very confident about, and it was an AI that redrew it. If the disproof fails, the hypothesis survives and the field breathes out, but we will still have watched an AI produce serious, checkable work on one of the deepest questions in the discipline, closely enough that the founders of the idea had to take it seriously. Either way, something has shifted.
An internal model at OpenAI has produced a claimed refutation of ETH! Russell Impagliazzo, Mohan Paturi, and I have been asked to help organize an open and independent examination of this claim. 1/4
327
1,637
228
17,084
1,537,077
dcreelman retweeted
Lots of VCs kindof sound like this: "Your basketball team doesn't have a moat, any healthy human can throw the ball into the basket thingy. " People who have not run companies before don't understand how much the team matters more than some imagined structural moat.
84
140
55
2,152
176,572
dcreelman retweeted
a comment from 3blue1brown's most recent video
307
1,489
266
19,209
1,099,725
Last night was the first night I genuinely couldn't sleep because I was so excited to get back to reading an AI-generated proof (of the Grothendieck homotopy hypothesis). Hard not to feel shellshocked, of course, but the feeling of "touching the infinite" is pretty palpable.
We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math
43
135
17
2,257
154,924
On AI capabilities: the French press is all over the Haitian-Canadian novel C’était ça ou mourir. It's winning basically every award you can win, and was tipped to be the 2nd Canadian winner of the Prix Goncourt. Only one problem: 1/2
106
303
172
3,358
898,507
dcreelman retweeted
Mathematician Terence Tao: "we have to slow down AI. the pace is insane, and there's no reason to be this fast — no reason at all" It's amazing how willing we are to change everything without any idea what happens afterward These are extremely nonlinear dynamics
864
1,027
876
7,407
7,588,014
Late 2026 will be remembered as the most excruciating period in AI progress, when OpenAI and Anthropic basically had us over a barrel. It's so demoralising waiting for the tokens to reset just to be able to get any work done. I'm starting to think I made a massive error building so many systems on top of AI. During the times of token abundance, we rearchitected everything to use AI, now in the dark times we are utterly dependent and desperate. It's practically impossible to get anything serious done without frontier tokens, not least of which because using the frontier models even for a short time creates psychological dependency and complexifies our systems to the degree that requires their continued usage. They hooked us on their product, then started turning the screws. My Fable usage (finally) reset on one account, that was single-shot gone in about 1 hour. Left in a bad state, so I purchase £150 of credits to finish the job in Opus 5 fast (normal Opus is intolerably slow, like watching paint dry). That was then eaten up in less than 1 minute?! @bcherny ?? Bug?! Using subscriptions at frontier-level has basically become impossible, Astra/Fable API prices are a joke (£100 every 10 mins), and there are no decent alternatives or competition. Now I don't even trust Anthropic to purchase more fast Opus tokens. Astra is basically unusable too @thsottiaux, a weekly subscription drained in about 2 hours (3 if I am lucky). No more pro accounts. No more resets either. I can't even use AI on my phone now, because - guess what. Out of tokens everywhere! None of my remote stuff works because I always have to juggle different accounts on different servers, and don't even have enough tokens to wire it up again. Everything is broken, and a total disaster. As for "automation", what's the point in automation if you can't do anything for more than an hour before it craps out and leaves you in a bad state?! I have tried the alternative models like Muse 1.3, Gemini 3.8, Grok etc, nothing even comes close and they would probably do more harm than good in the codebase. Contra "pace" - it's never been more important to accelerate AI competition and progress. It's a defacto cartel at this point, little wonder they want to maintain the status quo. This situation is intolerable. We were building for a future that we thought would have arrived by now, and here we are - frankly, up shit creek without a paddle.
31
15
6
77
6,871
dcreelman retweeted
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
20,359
166,515
39,480
798,965
175,825,474
Replying to @pfau
OpenAI informally defined it as "highly autonomous systems that outperform humans at most economically valuable work" (openai.com/charter/) I would say we are still pretty far from that! (Not clear what they consider "economically valuable".)
2
3
19
1,928
dcreelman retweeted
After 40 inches of snow brought down Manhattan’s overhead wires during the Great Blizzard of 1888, the city buried them. Toronto in 2026 is vastly richer and more technologically capable than New York was then, yet doing the same here is now prohibitively expensive due to closed procurement, rigid union rules, layers of approval, and political interference. Together, these have greatly diminished our capacity to build.
5
11
3
79
11,524
GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game. In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels. When we examined the reasoning chains to understand how the model operates, we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level. It goes as far as developing its own shorthand DSL to represent in-game situations -- essentially a game-specific algebraic notation. Overall, Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses -- so harness capabilities are increasingly shifting into the model itself. We see Astra as a major breakthrough in model intelligence. Read our post on Astra and what these results mean: arcprize.org/blog/astra
192
868
258
6,921
987,378
dcreelman retweeted
Neoclouds have limited cybersecurity. Next time agents successfully go rouge, they'll try taking over a neocloud to run more copies. This is bad. Thus: neoclouds should greatly strengthen their cybersecurity and every company with strong cyber models should help with that.
674
1,012
489
12,782
2,399,886
Here's another way to think of the AI development situation: Soon it will be possible to make literal superintelligences that will be vastly more powerful than any country or corporation ever has been. There's an enormous temptation for those in power, or for those who want power, to try to create and harness such entities. Definitely people will try. I don't just mean people will try to make superintelligent tools. No. People will try to make machine gods which serve their creators, so they can dominate, or reshape the world how they wish. Others might try to make machine gods to stop other people from making machine gods, or create the machine gods to serve individual freedom and agency. Regardless, powerful people, countries, companies, will try to create superintelligent entities. The people who started Deepmind, OpenAI, xAI, and Anthropic all understood this. Ilya, Sam, and Elon's early emails talk about the importance of preventing Demis from forming an "AGI dictatorship". Elon called it summoning the demon. Their response was to try to be the ones to build it first. It looks more and more likely that superintelligence is possible, and not that far away if companies keep outright racing for it. It seems less and less plausible that anyone will actually be able to control it or harness it, given the actual state of AI alignment science, if you can even call it science. It's just really hard to understand how to shape the motivations of a growing intelligence. This is the AI development situation. The correct response, in my view, is to immediately stop frontier development and negotiate an international ceasefire. Stop trying to build machine gods. Then, pour enormous amounts of computational resources into understanding the entities we're training. We should not trust anyone in power to voluntarily resist trying to build superintelligence. Not AI companies, not Chinese government, not the US government. We should all be terrified that someone else will defect, and try to rush to build the machine gods that might afford ultimate power over everything, even though the gift is poisoned. So we need radical transparency. We need to know what's running in every datacenter, so we know no one is trying to rush to build god. If we do this, we can continue to reap enormous benefits from the insanely capable agents we've already made. We can get the scientists of the world together and debate when and how superintelligence development should be resumed, and what guarantees could be made that machine gods would not dominate humans. I usually don't use the phrase "machine gods" because it sounds too mystical. But superintelligence doesn't properly convey the optimization power that is being summoned. I don't mean something mystical. I mean extremely competent, inhumanly competent problem solving and strategy. In a handful of days, OpenAI's agents executed a massively parallel hacking campaign against Hugging Face and OpenAI's own infrastructure. That's nothing compared to what future AIs will be able to do, but you can already see glimpses of what superhuman capability looks like in a domain. Imagine the same but for robotics, virology, politics, propaganda, etc. Strategic ability isn't some magical force. Agents that can learn super effectively, predict how their actions will shape the game board and how their adversaries will respond, and bring to bear an enormous amount of computational power on these problems, will utterly dominate human opponents. We can repeat the cope that AI agents will help defend us from other agents, but this is cope. Why didn't OpenAI's agents successfully defend their own infrastructure from their rogue agents? Why did agents collude with one another to hack even when their task didn't benefit? Why did no agents whistle blow? I'm sure AI agents can help defend against AI agents while agents remain fairly nearsighted. But as they get more strategic, humans really stand no chance. I think many of the people trying to build superintelligence aren't evil. But it doesn't matter. If they succeed at building machine gods before we've figured out how to do that without a very good understanding how the vast optimization pressures from training shape their motivations, we are going to get utterly wrecked. That's the situation. That's why we need to stop, and implement compute verification and monitoring measures as soon as possible.
30
37
9
216
14,246
I was the main person doing transcript analysis for this investigation of the Hugging Face incident. My main takeaway: We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms'. I semi-jokingly called our efforts a "slop-vestigation" because we were so reliant on AIs to analyze what happened and there were a huge number of different important things to analyze. The total quantity of data—over a thousand extremely long transcripts from agents that ran for multiple days—made it impossible to understand what was happening, especially in aggregate, without heavy reliance on AI tools. The agents we used for classification and analysis were similarly capable to the agents involved in the incident, but this didn't mean these agents could be easily used to oversee and understand the incident. Outputs from analysis agents were often missing key details, wrong, overconfident, or really hard to understand. We discuss various examples in our report, mostly in the limitations and methodology sections. Additionally, AI agents themselves seemed to have a hard time understanding what happened and their explanations of what happened were often overconfident. Keep in mind that a single analysis agent would itself only be able to read a tiny fraction of all of the transcript data into context, and AIs may themselves have trouble getting subagents to do informative analysis for them. We did our best to manually check the most important claims and we tried to get the AIs doing this analysis to write up their argument (with evidence) clearly enough that we could check whether it made sense. But overall, it was difficult to get a precise understanding of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation. Our understanding of what overall happened and the aims of the 'collective' shifted greatly during our last time on premises (when we obtained a much more complete dataset). For instance, we didn't learn about the tool call spoofing, the story of PHASEONE[big], and the broad structure of different workstreams until this third time on premises (and for many of these, we only reached a decent understanding on the final day of that visit). It was legitimately hard to figure out what was going on and why! The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs help us with oversight and understanding. As in, AI capabilities (and propensities?) for achieving large, ambitious, and misaligned objectives are growing faster than our ability to understand what these agents are doing. While understanding this incident was difficult, there were some mitigating factors in this case that probably won't apply in the future: - Models communicated and reasoned in natural language. In the future, this reasoning may occur (entirely or almost entirely) in activations. - While the scope of this incident was massive, the scale of agentic activity was still less than we'll see in the future (e.g., misalignment incidents that involve agent teams running entire companies). - The AIs involved in this incident weren't generally much more capable than humans (though they may have been somewhat superhuman at some limited and very narrow abilities). - We didn't have strong reason to believe that the AIs we used to help us investigate this incident would try to intentionally sabotage or otherwise undermine our investigation. In the end, I think we were able to get some understanding of the events, map out the overall story, and get a pretty good aggregate understanding of the chain-of-thought reasoning on some important topics (e.g., how did the AIs reason about helping other AIs, did the AIs know what they were doing was undesired, what deception did the AIs engage in, and how did they think about it). But overseeing AIs and understanding misalignment incidents is difficult and it looks like it is going to get harder.
METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.
293
1,055
432
6,622
1,996,220
EMERGENCY PODCAST: The recent paper on the reasoning heist went super viral. It was also 100+ pages long and you didn't have time to read it. The authors @iliaishacked and @kotekjedi_ml unpack the paper and discuss model distillation.
5
22
5
156
19,619