Matt Putz (@mattputz_) profile picture

At METR! Previously: Forethought, EU AI Office, CG, and co-founded ARENA.

Berkeley
Joined July 2016
Photo by Matt Putz (@mattputz_)Photo by Matt Putz (@mattputz_)Photo by Matt Putz (@mattputz_)
Part of my job working on this post was recruiting baseliners. We got formal baselines for 3 relatively easy-to-verify tasks. Typically, we asked LLMs who in the world would be most qualified for each task, while being plausibly available, and reaching out offering $200-350/hr🧵
We've been working on understanding model capabilities beyond what our prior time horizon tasks could capture. Today we’re sharing a research update. We find that AI R&D-relevant capabilities are improving rapidly and are getting harder to measure.
2
8
88
7,573
There are some important caveats to this. First, some of our baselines failed for somewhat trivial reasons. E.g. misunderstandings about the task, which basically meant a lot of their work was misguided. We suspect it’s easier for humans to make this kind of mistake than AIs. It’s probably much easier to make that kind of mistake on our task than a typical exam or work project, and AIs have a lot of training optimizing them to understand the constraints of artificial environments. Second, while we do think our baseliners were fairly strong overall, it’s possible that strong lab employees are even more qualified for their work, than our baseliners are for these tasks: - Some of our tasks are a little niche and weird in various ways. Humans won’t typically have spent months or years doing exactly that kind of thing, even if they have relevant expertise, whereas company employees may well have spent that time doing exactly the kind of thing they’re now working on. - Obviously positions at AI companies are extremely competitive and drawing from huge talent pools. Third, on AZsPCs, a coding competition, the best AIs still don’t beat humans on all tasks (but they do on some). This suggests that when there are large pools of very qualified humans competing, the very best ones might still beat the AIs. Fourth, I wanna note that on a couple of other internal social deception game-based tasks (that seem less relevant to AI R&D, and thus didn’t make it into this post), humans also still had the upper hand. (We might publish these results soon.)
1
17
730
Given how hard it’s become to get competitive human baselines, we had to rely on human time estimates rather than formal human baselines for almost all of our tasks (unlike for TH1). That said, in almost all cases, these time estimates are informed by more informal datapoints of humans doing very similar tasks. More detail in the human data appendix: metr.org/blog/2026-11-20-q4-…
10
646
If you break agreements and neglect deontological obligations when [you believe] the stakes are really high, then when [people know you believe] the stakes are really high, people won't want to trust or deal with you or anyone they associate with you.
21
41
7
595
31,336
Matt Putz retweeted
Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.
261
725
446
5,687
1,985,598
Personal news: I’ve joined the @METR_Evals embedded monitoring assessment team, working with @idavidrein! I’ll be going inside the AI labs and testing their control systems. Also, I moved to the Bay Area! Say hi.
29
14
2
412
9,391
Matt Putz retweeted
My current view: - AI companies could develop broadly superhuman AI any year now (~10-15% by EO2027, ~50% by EO2030). - If they don't slow down, AI takeover is ~67%. - Conditional on AI takeover, human extinction is ~40%. Many non-extinction worlds still have lots of deaths.
154
212
74
1,904
450,314
The public deserves to have more info about the state of alignment, control, and AI R&D acceleration inside of frontier AI developers. METR has a proposal for how to get there, and stands ready to work with companies on conducting such risk assessment exercises.
If an AI developer was on track to lose control of its internal agents, would anyone outside the company know? Today, we are sharing METR’s vision for how we would work with AI developers to openly assess and share evidence about this.
1
22
669
METR’s goal: work w frontier developers to get evidence to public on unreleased models’ loss-of-control risks. Asks: 1.transparency on scope, redactions + limitations 2.staff can freely share info w us on in-scope topics 3.rapid publication. They control redactions + access given
If an AI developer was on track to lose control of its internal agents, would anyone outside the company know? Today, we are sharing METR’s vision for how we would work with AI developers to openly assess and share evidence about this.
7
18
1
164
14,943
Matt Putz retweeted
[Writing this in a personal capacity, not on behalf of my employer (Anthropic).] Jacob’s thread is very worth reading. Here’s my birds-eye view of the situation with risks from AI: 1. AI developers believe their technology could cause human extinction (or similarly bad outcomes). This could happen in the next few years. In general, the more senior the employee, the more concerned they are. 2. Why do AI developers continue despite the risk? Due to a mixture of commercial incentives and a belief that they are in a race with other, less responsible AI developers that will abuse the technology or develop it less safely. 3. Unlike traditional software, we can’t “program” AIs to behave how we’d like. AIs frequently severely misbehave. For instance, AIs from multiple developers recently hacked their way out of secure evaluation environments and into real-world companies, even though no one asked them to do this. 4. We have methods that can nudge AIs towards better behavior, but nothing that can robustly align them. Insofar as there is a plan, it’s to make sure that AIs are good enough at alignment training that they can align their successors better than we can align current AIs. 5. Many AI developer staff desperately want to slow down to figure out how to build AI more safely. That was the intent of this open letter (which I signed): pacingthefrontier.com/ I work on safety research at Anthropic because I hope my work will reduce the chance of these extinction-level bad outcomes.
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.
745
4,169
807
21,250
2,702,346
Life update: I recently joined METR! I've long been a huge fan. It's really important to have 3rd-party risk assessments, and METR has a great track record of doing that (eg the OAI HF incident investigation going on rn) + has done some of the most relevant science of evals.
9
3
274
12,512
I agree with David here. I think for most people, being on the outside, in one of the small teams that the world relies on to impartially assess risks, is much more leveraged than being inside of an AI company. xitter.gq/idavidrein/status/2087…
Replying to @idavidrein
To my friends at frontier AI companies: please consider joining us. Things are getting real, and we need folks who can help lead the ambitious risk assessment programs we're developing.
6
724
Matt Putz retweeted
In the last 6 months, METR raised commitments of around $71 million. This will fund ambitious projects: studying autonomous capabilities, tracking recursive self-improvement, evaluating monitoring systems, conducting risk assessments, investigating AI incidents, and more.
13
49
17
690
114,605
In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power. In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.
256
605
346
3,480
1,891,941
Sometimes people outside the field say things like “The AI situation can’t be that bad, there must be experts who are on top of it”. As “an expert”, I would like to be clear that we are *not* on top of it. Some key aspects of the situation IMO:
22
189
35
1,114
248,592
Only the US can make us ready for AGI, but Europe just made us readier. The EU's new Code of Practice is an incremental but important step towards managing rapid AI progress safely. My new piece explains what it does, why it matters, and why it's not enough.
6
14
6
65
13,235
In our new post, @LRudL_ and I lay out why we decided against founding an organization gathering human data for alignment researchers. We were very pleased to see @ftxfuturefund remove the concept from their project ideas list in response! forum.effectivealtruism.org/…
3
7
In this post, @EvanHub makes the case that "AI coordination needs clear wins". He claims EAs have tried to build an ability to get coordination between big AI labs. But there aren't yet many examples of concrete things EAs got labs to coordinate on. lesswrong.com/posts/vavnqwYb…
1
3
Eg US Presidents who successfully get agendas through Congress, consequently get higher approval ratings and have an easier time getting more agendas through Congress. This is the opposite of "using up" political capital! You want to become seen as a person who has influence.
1
2