We are really excited to build independent evaluations of RSI progress and are actively hiring. We constantly see conflicting evidence about RSI progress. Most hinges on information coming from labs. To illustrate this, let's just look at evidence discussed *this week* that conflicts in whether it suggests RSI progress is fast or slow. xitter.cf/tobyordoxford/status/2… Results from OpenAI and Google evals indicate the time horizons for R&D tasks being automated are around ~15 minutes, which are much slower than METR time horizons for software tasks. This suggests we should have *slower* RSI timelines. xitter.cf/bneyshabur/status/2112… Results from Mirendil indicate that pretraining at a certain scale can be heavily automated such that it can be managed by a single pretraining researcher. This suggests we should have *faster* RSI timelines. xitter.cf/SemiAnalysis_/status/2… Results from Nvidia indicate that frontier models provide significant uplift for CUDA optimization. This suggests we should have *faster* RSI timelines. Each piece of evidence is more complex than what I write here. Interpreting them all together is confusing even if you trusted them all fully. And factoring in the underlying level of credibility is basically impossible: each result is basically just some plots and numbers selectively disclosed by labs in a tweet/talk/report. So deep sense-making about RSI progress is very hard. If you are interested in building independent RSI evals, apply here: cruxevals.com/careers/senior…

Firms are doing a lot to capture value from their own AI adoption, and yet often don't succeed. We are trying to understand the bottlenecks in practice better. If you have visibility into firm adoption whether as an employee, FDE, consultant, VC, or in some other capacity, we would love to hear from you. docs.google.com/forms/d/e/1F…

I have written extensively about the value of independent experts external to the frontier labs. I personally have chosen to remain independent at great financial opportunity cost. So it might surprise some that I am often quite supportive of my friends, colleagues, mentees, and mentors taking roles at frontier labs. I think these roles often are high-impact, rewarding, and exciting. With that said, today's events made me want to comment on the one consequence I personally find most saddening. Many of those deeply enmeshed in the frontier AI space might feel this in a way most farther away probably do not. It is the sadness of no longer seeing these folks contribute to public discourse on AI. First, as a point of clarification, I want to distinguish this from publishing open research. A well-understood cost in today's environment of joining a frontier lab is sacrificing the ability to openly publish research. There are some exceptions based on topic, company, and circumstance, but this is largely true. This cost is also something to be sad about, but what I am referring to is broader. To be glib, it is their contributions via Twitter, not arXiv, that I am talking about here. Dan Selsam's essay today reminded me of this. Dan and I were both Stanford PhDs advised by @percyliang. Our PhDs did not overlap, but I certainly knew of Dan when I was a student, and Percy occasionally referenced him to me as a north star in a certain sense. I have met Dan on a few occasions over the years. As many who know him better than I have already said, he is brilliant and prescient. My favorite example of this is his work on AI for math via the IMO grand challenge: quantamagazine.org/at-the-in… This challenge, initiated around the time of GPT-3, seemed absolutely ludicrous to me at the time. Because Dan was a former student of Percy, I said nothing at the time, but I mean it seemed entirely outlandish. As many have commented with the recent advances in AI for math, it seemed ridiculous to think AI would be able to get an IMO gold at a time when the most promising models of the day (GPT-3) could barely add at the level of a first grader. Everyone should read what he has to say. But what I want to comment is on the broader trend. As I experience it, when folks go to join the labs, they disappear from public discourse. Sure, we might see occasional glimmers like today's discourse about pacing AI, or past moments like the Anthropic supply chain risk designation or the OpenAI board coup or so on. And it is really important that lab employees speak in these moments. Society benefits greatly from their courage. And their courage should not be taken for granted. Yet I truly miss their participation in the day-to-day public discourse on what is happening with AI and how it will affect our society. There are so many incredible people at these labs, many of whom I would consider the world-best experts on certain topics, and I miss seeing what they have to say.