Essential weekend reading. The Scaling Era: an oral history of AI by @dwarkesh_sp and thank you @stripepress!

this is great

Link to the full episode:

Highly recommend tuning into @dwarkesh_sp's episode today with @karpathy. They dive deep into why RL is so information-sparse and what that means for realizing the decade of agents. A few highlights that stood out: - “RL is terrible; it’s just that everything else is much

Get in touch to learn more about our latest work in RL here:

Thrilled to be featured in Dwarkesh’s latest episode with Richard Sutton, widely regarded as the father of reinforcement learning and 2024 Turing Award winner. As Richard explains, we’re entering the Era of Experience, where training AI means creating environments that capture

See more of Dwarkesh’s visit and get in touch to learn how Labelbox delivers large-scale, high-fidelity data collection to advance next-gen robotics.

As his latest guest, @svlevine (co-founder of @physical_int) predicts, robots could be running households entirely autonomously by 2030.

We recently invited @dwarkesh_sp to stop by our SF robotics lab. World-class podcaster, rookie robotics intern.

.@svlevine is one of the world's leading robotics researchers (and co-founder of @physical_int). He thinks fully autonomous robots are much closer than people realize - when I pushed him on a prediction, he said 5 years to robots that can autonomously run your household). The

If you’re a Dwarkesh fan, check out the landing page and follow along, this is just the beginning of something special.

We’ve always admired how @dwarkesh_sp sparks conversations with top thinkers in AI, academia, and tech. Now we’re teaming up to connect with a community that shares our mission of pushing the limits of what’s possible in AI. The first episode together with one of his most

We’ll continue evaluating frontier models on more constraint domains and reporting as the gap between leading AI capabilities closes. Check out our blog post for more info!

Lessons learned: Constraint interactions, not just rules, limit performance, and success on synthetic tasks doesn’t always transfer to real-world cases. We observe that high constraint densities tend to also expose weaknesses, and analyzing failures helps guide targeted

Our initial findings show that no current model maintains consistent feasibility under real-world, high-complexity scenarios. On synthetic stress tests, o3 demonstrates the highest feasibility, closely followed by GPT-5. In a domain-grounded data center migration benchmark, GPT-5

We tested whether leading models could generate schedules on RCPSP (resource-constrained project scheduling problems) that meet all constraints and remain consistent as complexity increases. To do this, we varied task difficulty across hundreds of levels and applied realistic

Introducing ConstraintBench: a new benchmark for evaluating LLM reasoning on realistic resource-constrained project scheduling problems (RCPSP), a well-known NP-complete challenge. It tackles some of the toughest planning challenges (such as project management, construction,

As AI advances, so do the human skills required to shape and align it. Full report: https://t.co/pmiGSriNU6

Grok-4 just landed on our Complex Reasoning leaderboard, and it’s impressive💥 - Math: 81.8% - Pure Math: 84.8% - Applied Math: 79.9% - CS: 75.4% - Reasoning: 77.8% - Aggregate: 80.7% See how it stacks up:

obligatory AI company SF billboard