From DeepMind Paris to Llama 3: Michal Valko shares his three hard problems at Prague AI meetup
misovalko · x · 2026-11-07
What started as a 1:1 coffee turned into a 70-person AI meetup in Prague. Speaker Michal Valko previously helped set up DeepMind's Paris lab and later worked on Llama 3 at Meta.
His three hard problems: creativity, confidence (a model knowing when to trust its own answers), and collective recursive self-improvement—agents working together to improve how the next generation learns. He picked the last one for his new company Isara Labs, calling it the easiest of the three. He's also experimenting with a counterintuitive question: if training only rewards small improvements, what happens to ideas that need a few bad steps before working—like touch-typing, where speed drops before it rises.
More from Companies & People
- AI weekly: unreleased OpenAI model finds 948 Linux bugs, Anthropic posts $17B quarter, chips fall 12% — eyishazyer · 2026-11-07
- The temple device isn't the bet — Deepinder Goyal himself is — DrDatta_AIIMS · 2026-11-07
- CuspAI forms AI Materials Foundry with 45 partners including Meta and Nvidia — jonfildes · 2026-11-07
- Indie hacker tibo_maker rethinks 'quantity over quality' in the AI era: don't ship slop — tibo_maker · 2026-11-07
- Andrew Ng's AI Dev NYC conference set for Nov 30–Dec 1 with Yann LeCun — DeepLearningAI · 2026-11-07
- Ex-Googler reveals Google UK contract bans any criticism forever — austinc3301 · 2026-11-07