Our Take
Another massive bet on pure RL when the field's biggest wins still required human game rules and reward functions.
Why it matters
London's AI funding surge suggests DeepMind's decade of talent development is creating a new geographic center for AI research outside Silicon Valley.
Do this week
AI teams: audit your current RL projects before Friday to identify where you're still depending on human-labeled data.
Silver's Ineffable raises $1.1B at $5.1B valuation
David Silver, former DeepMind reinforcement learning lead, raised $1.1 billion for Ineffable Intelligence at a $5.1 billion valuation just months after founding the company. Sequoia Capital and Lightspeed Venture Partners led the round, with participation from Index Ventures, Google, Nvidia, the British Business Bank, and the U.K.'s Sovereign AI fund (per Wired reporting).
The company aims to build a "superlearner" that discovers knowledge and skills without human data, relying purely on reinforcement learning. Silver spent over a decade at DeepMind developing systems like AlphaZero, which beat professional players at chess and Go by learning from self-play rather than studying human games.
Ineffable joins a wave of researcher-led "coconut rounds" (seed rounds exceeding $1 billion). AMI Labs, co-founded by Yann LeCun, raised $1.03 billion at $3.5 billion last month. Recursive Superintelligence, founded by another former DeepMind scientist, reportedly raised $500 million with demand for $1 billion.
London emerges as AI funding hub
Three major AI ventures with DeepMind connections have now secured massive funding rounds in London. DeepMind's continued presence since Google's 2014 acquisition created a talent pipeline that's generating multiple billion-dollar startups. Jeff Bezos's Project Prometheus is reportedly seeking office space near Google's AI hub, signaling sustained momentum.
The technical bet remains ambitious. Silver claims success would represent "a scientific breakthrough of comparable magnitude to Darwin" where their "law will explain and build all Intelligence." However, even AlphaZero required human-defined game rules and reward structures. Pure self-learning without any human scaffolding remains unproven at scale.
Watch the human data dependencies
Teams building RL systems should map exactly where human knowledge enters their training pipelines. Even systems marketed as "human-free" typically rely on human-designed reward functions, environment constraints, or safety bounds. Silver's track record suggests meaningful progress is possible, but the complete elimination of human input faces fundamental challenges around goal specification and safety constraints.
The funding environment for researcher-led AI startups has clearly shifted toward massive initial rounds. Technical teams with strong publication records should expect accelerated fundraising timelines, but also heightened expectations for breakthrough results rather than incremental improvements.