The tech report behind our 6x pretraining compute multiplier is out today.
Training LLMs for data research takes more scientific control than training one great model does. Even small changes can confound results.
Some spoilers on how we did it:
https://t.co/kyNZ9XUFkj
The #AI100 is a powerful snapshot of the private companies defining this moment in technology from the models and infrastructure underneath the stack to the applications changing how people create, build, and work.
We’re thrilled to see 13 Felicis portfolio companies recognized on this year’s inaugural list from Newcomer and Bessemer Venture Partners:
→ @Arena, founded by Anastasios Angelopoulos, Wei-Lin Chiang, and Ion Stoica, is building a trusted, community-powered evaluation platform for AI models.
→ @Canva, founded by Melanie Perkins, Cliff Obrecht, and Cameron Adams, is making visual communication and design more accessible to everyone.
→ @Crusoe, founded by Chase Lochmiller and Cully Cavness, is building energy-first AI cloud infrastructure.
→ @ExaAILabs, founded by Will Bryk and Jeffrey Wang, is building search infrastructure for AI applications and agents.
→ @Fractile_ai, founded by Walter Goodwin, is rethinking AI compute to make inference dramatically faster and more efficient.
→ @Mercor, founded by Brendan Foody, Adarsh Hiremath, and Surya Midha, is building the network for expert human knowledge and capabilities in the AI economy.
→ @n8n_io, founded by Jan Oberhauser, is giving technical teams the building blocks to connect systems and automate agentic workflows.
→ @Notion, founded by Ivan Zhao, Simon Last, and Akshay Kothari, is the connected workspace where teams write, plan, and build with AI.
→ @Periodiclabs, founded by Ekin Dogus Cubuk and Liam Fedus, is developing AI systems to accelerate scientific discovery.
→ @Runwayml , founded by Cristóbal Valenzuela, Alejandro Matamala, and Anastasis Germanidis, is building the generative AI tools powering a new era of storytelling.
→ @SkildAI, founded by Deepak Pathak and Abhinav Gupta, is building a general-purpose foundation model for robotics.
→ @Supabase, founded by Paul Copplestone and Ant Wilson, is the open-source Postgres development platform developers love.
→ @Zipline, founded by Keller Rinaudo Cliffton, Keenan Wyrobek, and Will Hetzler, is building autonomous logistics and delivery infrastructure.
There is no single blueprint for building an enduring AI company. These teams span search, creative tools, developer infrastructure, robotics, science, and the systems that make AI useful in the real world. What they share is the ambition to build something foundational and the ability to move with extraordinary speed.
Congratulations to every company on the AI 100, and especially to the founders and teams above. We’re honored to be on the journey with you. 🧡
See the full list: https://t.co/wcf33IATXg
Introducing Exa for Finance.
Your agents can now search over 2 billion financial documents including:
- 110M+ daily prices for stocks and crypto
- 2M+ SEC filings and earnings calls
- 250M+ finance news articles
It's state-of-the-art on FinSearchComp and other public benchmarks
1/ We asked founders which seed-stage companies they’re paying attention to, and their answers make it clear: AI-enabled biotech and healthcare is having a moment. In fact, when you look at the funding data, we’re starting to see a real divergence between biotech and AI-enabled biotech companies:
Why now?
Breakthroughs like Evo and Alphafold have shown that DNA and protein prediction accuracy follows similar scaling trajectories to LLM next word prediction
Massive improvements to the infrastructure to support the multi modal data generation required in biological analysis
These are exciting, technically intriguing problems that have a real potential to make our lives better. That’s attracting the most ambitious founders and builders.
But biology isn’t software. It’s hard to validate, and has to be done physically. Seeing seed-stage activity pick up gives me hope that we’ll start to see new approaches to trials and business models that have historically been the bottleneck.
The entire Emergent list from @NewcomerMedia and @Felicis: https://t.co/vMxjE3msZW
It's clear that robotics may be approaching its inference moment.
As deployments scale, inference becomes a networking + scheduling + heterogeneous compute problem.
Today there’s no vLLM for robotics. Why is that? 🧵
i’m hiring a chief of staff. reach out if you want to work unreasonably hard with me on the most important moonshots in human health. must be technical.
@nytimes didn’t eviscerate us 🎉
One week of data collection, 4 death threats, and 1 restraining order.
Overall comments are supportive but clear most unprepared for robot augmentation of human labor at scale.
Thanks @kashhill for the coverage
https://t.co/izrqKy9C81