David Silver, the DeepMind scientist who architected AlphaGo, has founded a startup called Ineffable to build what he calls an AI "super learner" based on reinforcement learning. The approach is explicitly framed as an alternative path to large language models, the dominant architecture behind today's AI boom. Silver discussed the venture in an interview with Madhumita Murgia of the Financial Times.
Silver said he feels a moral obligation to advance the technology. That's a notable framing from someone who spent years inside one of the world's top AI labs and chose to leave rather than continue scaling LLMs from the inside.
Why reinforcement learning, and why now
LLMs learn by predicting the next token across vast text corpora. Reinforcement learning works differently — an agent takes actions, receives feedback, and adjusts. It's the technique that made AlphaGo beat the world's best Go players, and it's the same family of methods behind many of the AI systems that play games or control robots.
📊 Market Data Snapshot
Silver's bet is that RL can produce capable systems without the enormous compute bills that come with training frontier language models. If that holds up, it changes what kind of infrastructure AI companies need — and which ones get funded.
Ineffable is early. There's no product, no benchmark, and no published research from the company yet. What exists is a founder with a track record and a thesis that runs against the current fashion in AI.
The talent signal matters more than the tech right now
Silver isn't the first senior researcher to leave a big lab for a startup. He is one of the more credible ones to do it with a non-LLM approach. That matters for where elite talent and venture money flow over the next few years.
For crypto specifically, this is a story with no direct price impact. Bitcoin doesn't care. Ethereum's roadmap doesn't change. No token's fundamentals move because a former DeepMind scientist started a company.
But the AI-crypto sector has spent the last few years pitching itself as the decentralized answer to centralized AI. Most of those pitches lean on LLMs, federated learning, or GPU compute markets. If RL-based systems turn out to be more efficient and less compute-hungry, the value proposition shifts. Not overnight. Over a five-to-ten year horizon, possibly.
What crypto AI projects are actually exposed to
The immediate read-through is sentiment, not fundamentals. AI-themed tokens trade on narrative, and a high-profile researcher validating RL as a serious alternative to LLMs could eventually pull attention toward projects building on that architecture. There aren't many of them yet.
There's also a less comfortable angle. Crypto AI projects often tout hires from major labs as proof of credibility. If researchers of Silver's caliber keep choosing traditional AI startups over crypto ventures, that pipeline looks thinner than the marketing suggests. The talent war isn't just between AI labs — it's between AI and crypto, and crypto isn't obviously winning the top of the funnel.
Compute markets face a similar question. RL models can in principle be trained more efficiently than frontier LLMs, which would reduce demand for the massive GPU clusters that underpin several crypto compute narratives. That's not a prediction. It's a scenario worth watching, because the sector's pitch depends heavily on compute scarcity persisting.
What to watch from here
Ineffable hasn't published results. The next concrete checkpoint is whether the company releases research, benchmarks, or a technical paper showing RL competing with LLMs at meaningful scale. Until that happens, this is a founder story, not a technology story.
Silver's own words set the bar: he says he feels a moral obligation to advance the work. The market will eventually ask whether Ineffable advanced anything at all.
