Jack Clark

Here’s what I’ve been working on recently: @anthropicai. I’ll be spending a lot of my time on measurement and assessment of our AI systems, as well as thinking of ways govs/others can assess AI tech. There’s a lot to do!

@sudoraohacker would be very interested in you taking time to write up a take here! part of why I wrote this essay is I think it's good to have more public discussion about this issue outside of the usual places (e.g, LessWrong)

I think there’s a lot more uncertainty about AI recursive self improvement happening than what Jack suggests. There are 3-4 good benchmarks on generic AI research that suggest progress, but the ultimate gating factor is whether software will be enough for AGI or if we need faster hardware codesign and deployment (eg EUV machine construction and more chip fab capacity), energy system upgrades and DC construction, monitoring and control for a wide range of non-SWE and non-AI jobs (whether computer based or physical), humanoid progress, etc. I think those bottlenecks will take many more years to work through and will be a practical resistor to RSI and AGI takeoff scenarios. Don’t get me wrong - I think the future will happen faster than most humans today expect and we’re in a singularity-like transition, but maybe not as fast as current lab researchers think.

@andrksl @edwardmarkdown @deredleritt3r thanks for taking the time to ask me some specific, constructive questions here. Have a good day!

@jackclarkSF @edwardmarkdown @deredleritt3r that's awesome. definitely silly to use academic credentialism as a complete proxy for technical understanding esp. for an ant founder lol. enjoyed your ezra klein interview.

Jack Clark believes that there's a ~60% likelihood that we will reach fully automated AI research in 2028, but only a ~30% likelihood we will reach it in 2027. The full Substack post (linked in the tweet below) explains his rationale and is worth reading in its entirety.

@andrksl @edwardmarkdown @deredleritt3r I also am reasonably well calibrated - I have a bunch of public technical predictions in my newsletter about things like distributed training or error rates and I do pretty well.

@andrksl @edwardmarkdown @deredleritt3r It's 90% my own understanding, 10% colleagues. My hobby is sitting around and reading arxiv papers. I read the papers until I understand them, then I write summaries, have claude rate my understanding, etc. For Import AI, I've now read something like ~5,000 papers over 10 yrs

@edwardmarkdown @deredleritt3r I'm not sure what makes you so confident in being able to discount what I think and how I think. I wish you luck in your endeavors in life and hope you treat those close to you with more generosity and grace than you have treated me.

ofc i know he’s anthropomorphic’s cofounder. just saying that he’s not a technical person so he just saying what his employees tell him multiplied by marketing + fundraising goals this is not to criticize your post. i thank you for distributing the info because it’s interesting what jack got to say. just think that what he says should probably be discounted by a large margin because he doesn’t have the necessary expertise to make such claims and his obvious bias.

@JacquesThibs 🤣

> 40% chance RSI doesn’t happen by 2028 🚨 Anthropic co-founder believes we’re in a long timelines world!!

@S_OhEigeartaigh Yes, completely agree.

(there will be a time of course where you may need to alert us partly on the basis of non-public data - it would be surprising if that didn't happen at some point - but wherever and for as long as it is possible to make these arguments based on public data that can be scrutinised, challenged, checked, along with the data sources not used, that seems very useful)

A tragedy of the growing tech industry/lobby spin campaigns is that there are honest brokers in their midst trying to speak truth (at least, what they see as honest truth) about incredibly important, but deeply unintuitive things. And with every day the world has less reason to take them seriously. I firmly believe Jack is such an honest broker. He may be wrong about many aspects of this (indeed, I hope he is, especially about the timeline). But he's not trying to hype or otherwise mislead you. Please read this, assess it as honest communication, consider its arguments fairly. As I've said before, the core question in this essay is IMO the most important question in the world.

@karinanguyen My whole experience doing this project was finding endless "up and to the right" graphs at all resolutions of AI R&D, from the well known (e.g., SWE-Bench) to more niche (like those above). It's a fractal, but at all the resolutions you see the same trend of meaningful progress.

@karinanguyen There's also MLE-Bench, which is ecologically valid (tasks come from real kaggle competitions) and involves building a very diverse set of ML apps to solve specific problems. The same progress shows up here.

I've spent the past few weeks reading 100s of public data sources about AI development. I now believe that recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves.

@nwilliams030 consider: babies

may I recommend to you: babies ?

Announcing the 2026 Cosmos Lecture. Anthropic co-founder and Cosmos Founding Fellow @jackclarkSF will be speaking at Oxford on how to live a self-directed life in the AI age. Apply for tickets here: blog.cosmos-institute.org/p/the-2026-cos…

@AlexPalcuie there's been a chance in business strategy I'm afraid

went to the new anthropic king’s cross office but my badge didn’t work… what gives???

@PhilippKoralus yes this essay is a prelude to some of what I'll be speaking about in May in Oxford!

Looking forward to this !

@TimSchnabel Will be curious to hear your feedback!

Definitely waiting with anticipation for this "big essay"-- had the privilege of being at The Curve for Jack's keynote last fall (youtube.com/watch?v=e0L29m…), and am anticipating something just as riveting.

Import AI will be skipping this week. I have a big essay I've been working on relating to AI progress and it will come out in the next issue.