Anson Ho

@timfduffy @EpochAIResearch Hmm actually it seems like the link already uses "copy link to highlight", not sure why it's not working...

@timfduffy @EpochAIResearch Thanks for flagging this! That part's correct — it's based on the edge case in the "results table" at that link (see screenshot). OTOH the 512x is a typo, it should be 5^12 I'll update the link w "copy text to highlight" and the second number

Developing more powerful AI isn’t just about scaling compute. It’s also about improving algorithms and data quality, which let you build better models with the same compute. We call this “AI software progress” — here’s everything you need to know about it: 🧵

@SamuelAlbanie I haven't personally, but one of my colleagues tried using GPT-5.3 codex and made a lot more progress on the "port a post" task. So I think I was way too bearish for task 3, and maybe a bit too bearish for task 1 too...

@ansonwhho @ansonwhho curious whether you've retried with more recent models & whether they seem on trend with your forecasts?

interesting post by @ansonwhho tldr: - takes some representative tasks from his job & tries to automate them - when AI fails, he guesses when each task will be completed on the first attempt (with an output that is good enough to directly use) example: his median estimate for "porting an article from google docs to substack and epoch's website" is mid-2028 epoch.ai/gradient-updat…

@idavidrein Looks like @GregHBurnham was right! epoch.ai/gradient-updat…

Huh! I had assumed it saturated at like 85% since it's been a few months since anyone had improved on that rough score range—but this is much higher! Would be curious to see if the questions Gemini 3 Pro w/ deep think gets right are valid/correct

@dwarkesh_sp @EgeErdil2 Ege knows so much that when I first met him I thought he might have photographic memory. Then I discovered he doesn't, which made him seem even more impressive!

.@EgeErdil2 may be the most cracked mf I know. Extremely insightful, regardless of whether the conversation is about info/FLOP in RL vs pretraining, or why Japan won the 1905 Russo-Japanese War, or the weird distortions in many millennia long trends of population growth.

@datagenproc salut d'amour, schindler's list theme, sir duke

I find it much easier to point to what makes a harmony click for me than a melody. Or even to point to melodies I particularly like. Not sure if it's just that i'm less sensitive to melody, or if it's because melody is less legible. What are your favorite *melodies*?

Introducing FrontierMath Tier 4: a benchmark of extremely challenging research-level math problems, designed to test the limits of AI’s reasoning capabilities.

But I think that research on explosive growth has already provided strong counterarguments to the points that @lugaricano raises

I do agree that often "progress works itself out of a job", baumol effects and human bottlenecks are important and can make explosive growth less likely (especially in the next few years)

I strongly disagree with @lugaricano’s thread on explosive growth. While the thread raises important points, I think it fails to get to the core of the reasons to believe that explosive growth is plausible – allow me to explain!

@aidanogara_ @ben_j_todd @EpochAIResearch So for GATE specifically I might make updates of the form "I broadly over/underestimated how strong effect X might be". I definitely wouldn't trust GATE's near-term quantitative predictions (e.g. GWP growth rates in 2027)

@aidanogara_ @ben_j_todd @EpochAIResearch Depends on the question IMO. GATE is based on endogenous growth models, that are ok at capturing the dynamics of long-run growth, but I doubt you'd use something like this for near-term growth predictions for example

Economists say remaining human bottlenecks will prevent explosive growth from AI (Baumol effects). @EpochAIResearch actually made an economic model, and find those bottlenecks would need to be implausibly huge to do the job. Even with extreme bottlenecks (ρ = -2), partial automation leads to 10% GDP growth soon after. (The graph below is based on a scenario with 10% automation in 2026 – the exact date is not the point here.)

How do reasoning models solve hard math problems? We asked 14 mathematicians to review o3-mini-high’s raw, unsummarized reasoning traces on 29 FrontierMath problems. Here’s what they found: