🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

The new update to DeepSeek R1 is a significant improvement over previous versions, but still lags behind the best models.

Our page is quickly becoming the single best place to learn about AI capabilities. Check it out!

Love this analysis, props to Greg Burnham for not being afraid to dive deeply into the data!

GPQA Diamond is probably not quite yet saturated, though it's close!

Some extrapolations of model counts. Very relevant to understand the scale of compute based regulation!

Shameless moment of appreciation for Epoch: we put out four research pieces today alone, and we’ve been averaging about one every other day lately. In my previous job, I worked on the same research for 1.5 years, and it is still not published a another 1.5 later.

I believe we are on track towards AI that can do some really impressive things in the next 5 years, no acceleration needed. Solving open math problems, programming complex features on request, doing PA work reliably and generating long form videos all look within reach of the

Esta entrevista con Jaime Sevilla me ha parecido súper interesante. Necesitamos más voces pragmáticas, alejadas tanto del extremo catastrofista como los del «no es para tanto» elmundo.es/papel/lideres/…

Latest DeepSeek 4-11 months behind US: * ~5 months behind US SOTA on GPQA Diamond * ~4 months behind on MATH lvl 5 * ~11 months behind on SWE-Bench-Verified We need more good evals to benchmark the US-China gap. Kudos to Epoch AI for doing some of this work.

Whether compute and labour in AI research behave more like complements or substitutes is one of the most important questions of our time. Glad to see more work into it!

Anthropic models are pretty good at coding, according to benchmarks.

We then compare our system against 25 AI PhDs or postdocs with an average h-index of 8.9. Our participants take this task seriously: they spend on average 9.1 minutes on evaluating each pair of ideas. On five popular NLP topics, our system beats the majority voting of human