Pure RL vs post-trained LLMs is the new Python vs SQL. Let the games begin.

introducing @_hex_tech's Notebook Agent Act II build beautiful interactive analyses, easily organize & tidy your projects, and debug with ease all with the help of the agent 💜 already used by thousands of teams including @figma @NotionHQ & @tryramp

Have you integrated APOL1 genetic testing into your practice? Discover the No-Cost APOL1 Genotyping Program for eligible patients sponsored by Vertex Pharmaceuticals—helping you deliver precision care without added cost. Learn more today!

Sometimes, data analysis IS complex; but its presentation should always be elegant. Love how Notebook agents allows users to do hard things and make it look easy.

Datalane's $22.5M Series A 🚀 As AI models get better, durable value shifts to proprietary data. We back the infrastructure that makes data usable at scale. Datalane goes straight to the core: building the data layer for the local economy. https://t.co/nJ4vvtiOyk

NeurIPS 25 was less of a conference and more holy pilgrimage. The sacred scriptures didn't dissapoint, ranging from automating environment design to search over scaffolds, long-context architectures and making continual learning/multi-turn RL robust. Read on for our takeaways!

AI researchers aren’t short on ideas. The real bottleneck is running better experiments, faster. @sarahcat21 on why experiment design, execution, and analysis will unlock the next phase of AI progress: https://t.co/abOKFJnElA

Let’s repeat: continual learning is the next frontier.

Apply once. Work on the most important AI projects. We connect experts across software, data, healthcare, finance, and more with leading AI labs. Upload your resume, complete a short AI interview, and get instantly matched to flexible, remote projects paying up to $500/hour.

FoundationDB built a production-ready distributed database with ACID guarantees in just a few years. Everyone said it takes a decade! Their secret: they figured out how to test well. Full technical breakdown: https://t.co/thUO5pbO76

We've been working closely with NVIDIA to build the next generation of frontier models, which we debut today: our family of General World Models. We are thrilled to deepen our partnership with Jensen and his team. To celebrate, Jensen sent us a special holiday message. Hope you

We are honored to have received the Physical AI Pioneer Award, alongside @physical_int, @DynaRobotics, and @agilityrobotics, at the Nebius Robotics and Physical AI Awards & Summit. As @adrianmacneil mentioned in his panel, “Physical AI is moving from promise to reality, and that

(made with @runwayml)

Always curious and full of energy, Beagles turn every walk into an adventure 🐾🎉.

https://t.co/ddd5HDvtMW

Scaling compute and environments buys you more shots on goal, but without the right learning methods this will not scale RL efficiently or even at all. Why autocurricula—not crude RL maximalism—might unlock reliable RL:

It’s always been the data.

When talking to artists, two words come up: "process" and "control." AI engineers use different vocabulary: "automation" and "generation." This fundamental mismatch explains why most GenAI tools see explosive growth then gradual disenchantment. Full essay:

We don't actually have but one internet — there's a massive amount of alpha left in simply redefining how we look at and use the data we already have. The distributed systems work behind @datologyai's data curation: https://t.co/bdTkmIrZ22

Founding engineers at @LangChainAI, @perplexity_ai, and @stytchauth share their experiences: • Speed matters more than perfection. • Don't over-engineer for scale before you have users. • Be ready to spend a lot of time on recruiting. Full panel:

Congrats to @AntithesisHQ on their $105M Series A from @JaneStreetGroup ❤️ We wrote all about them and what makes DST so powerful a few weeks ago:

@itunpredictable good read!

SKYX $SKYX expands its retail footprint with the launch of its patented SKYFAN & TURBO HEATER at Target, marking another step in scaling its plug-and-play smart home platform. Management expects the winter rollout to support revenue growth through fiscal 2026 and advance its path

With the release of Deepseek 3.2 yesterday + Olmo 3 two weeks ago, people are talking about scaling RL again. One under appreciated hardware detail: How do you maximizing hardware efficiency during RL training between generators and trainers? For many top labs, as much as 90%