Ineffable’s David Silver bets on simulated worlds, not bigger language models

David Silver says his new AI company, Ineffable Intelligence, is pursuing reinforcement learning in complex simulated environments rather than relying on vast collections of human-written data. He predicts “extremely strong evidence” the approach is on track within a year, and superintelligence within three. Those are ambitious forecasts, not demonstrated results, but the proposed route is concrete enough to examine.

Ineffable Intelligence Watch analysis

What happened

In an interview discussed by BigGo Finance, Silver describes building “crucibles”: simulated worlds where agents learn through experience, with simple objectives rather than a detailed human-written path to a solution. He argues that systems trained this way might discover methods people have not already supplied. The environments could range from fast simulations to realistic physics, which he acknowledges would be computationally expensive.

Silver says Ineffable plans to study how powerful agents interact safely in simulation before allowing real-world deployment. He also describes possible monitoring methods, including “shadow actions”, where an agent’s proposed actions are observed without being carried out. The company’s co-founder Lasse Espeholt says the approach is designed to keep simulation, training and inference on GPUs.

Why it matters

This is a different wager from scaling language models: build environments where AI systems can learn by doing, then see whether that learning travels beyond the environment. Silver’s case draws on reinforcement-learning successes such as AlphaZero, while acknowledging that Go is narrow and that it remains an open question whether simulated worlds can support the complexity he wants.

If the method works, it could offer another route to systems that learn without being spoon-fed human examples. If it does not, an impressive simulation can still be a very expensive place to discover that the real world has different rules.

Our read

The interesting part is not the three-year superintelligence forecast. It is the testable proposition underneath it: that richer simulations and reinforcement learning can produce useful capabilities that language-model training will not. Silver’s timeline gives readers a clear claim to revisit; it does not yet give us evidence that the claim will come true.

What to watch

• Whether Ineffable’s planned large GPU cluster is fully available by the end of 2026, as the report says.
• Whether the company publishes results from its simulated environments within Silver’s one-year milestone.
• What evidence Ineffable will require before moving a system from simulation into the real world.

Discussion spark: Should frontier AI labs prioritise learning in simulated environments over scaling language models, or is that a promising detour until it produces evidence beyond games?

not affiliated with or endorsed by Ineffable Intelligence