Partnering with Ineffable Intelligence: A Superlearner for the Era of Experience

TLDR

• Sequoia is co-leading the first round of Ineffable Intelligence, David Silver’s London lab building a pure-RL “superlearner” with no pre-training on human data.

Key Takeaways

• Ineffable Intelligence’s core bet: an agent that learns all knowledge from environmental consequences alone, skipping pre-training and human-data imitation entirely.
• David Silver led AlphaGo Zero, which hit 5,000+ ELO by removing human pre-training and training purely through self-play, up from ~3,700 with human data.
• Self-play produced the ~800 ELO leap that led to AlphaGo defeating Lee Sedol in March 2016, a milestone long considered impossible given Go’s ~10^170 legal board positions.
• Silver also drove AlphaZero, AlphaStar, and AlphaProof at DeepMind, making him one of the few researchers who sustained RL as a dominant paradigm through the LLM era.
• The stated goal is a system capable of rediscovering language, mathematics, and physics from first principles, then going beyond what human training data can express.

Why It Matters

• LLM-trained systems inherit the ceiling of human-generated data; a pure-RL agent trained from scratch removes that ceiling by construction, not by hypothesis.
• Silver’s track record shows the approach already produced decisively superhuman, non-human-style play in Go; the question is whether the same scaling logic extends beyond bounded games.
• Sequoia explicitly frames this as a contrarian bet on an uncertain timeline, signaling the round is stage-zero research capital, not product-stage growth funding.

Sequoia Capital · 2026-04-27 · Read the original