mixture-of-experts
(4)
Noam Shazeer and the Art of Making Machines Talk
He helped give language models their modern architecture, then left the lab to see whether people would actually talk to them. His route from Google to Character.AI, back to Gemini, and on to OpenAI began with a crayon chart on a wall.
Reflection AI
Reflection AI is a New York-based artificial intelligence lab founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. It is building autonomous coding agents and open frontier language models, with the stated ambition of reaching superintelligence by first solving software engineering. Its flagship product, Asimov, is a code research and comprehension agent for large codebases. After raising $2 billion at an $8 billion valuation in October 2025, the company has positioned itself as an open, Western alternative to closed labs like OpenAI and Anthropic and to Chinese open-weight models like DeepSeek.
Zyphra
Zyphra is a superintelligence research and product company building efficient, open foundation models and a full-stack AI cloud. Best known for its SSM-hybrid Zamba models, the ZAYA1 reasoning family trained entirely on AMD hardware, and the Zonos text-to-speech system, Zyphra argues that frontier intelligence should run cheaply, locally, and outside the walled gardens of a handful of labs. In June 2025 it raised a $100M Series A at a $1B valuation led by Jaan Tallinn.
Azalia Mirhoseini
Azalia Mirhoseini is an Iranian-born AI researcher, Stanford professor, and co-founder of Ricursive Intelligence - a frontier AI lab valued at $4 billion that uses AI to design better chips, which in turn train stronger AI. Best known for AlphaChip, the deep reinforcement learning system that now designs Google's TPUs and has compressed chip floorplanning from months to hours, she also co-invented the Mixture-of-Experts architecture underpinning GPT, Claude, and Gemini. With 20,000+ citations and a $335M-funded startup launched in under four months, she is closing the recursive loop between artificial intelligence and the hardware it runs on.