David Silver, the AI pioneer behind AlphaGo, has launched Ineffable Intelligence to chase a reinforcement-learning–driven route to superintelligence. He’s openly challenging the current dominance of large-language models (LLMs).
Backed by $1.1 billion in seed funding and a $5.1 billion valuation, the venture wants to create “superlearners” that get better through trial and error. These agents can keep improving on their own, without having to rely on human-generated data.
Silver says human data is a finite fossil fuel. In contrast, autonomous self-learning acts like a renewable energy source—maybe even unlocking breakthroughs in science, technology, governance, and economics.
This shift hints at a strategic pivot in AI research. The focus moves toward systems that learn from the world, not just from text.
The Vision for Self-Learning AI
Ineffable Intelligence wants to go beyond scaling today’s models. They’re aiming for autonomous learners that can improvise and adapt over time.
Reinforcement learning sits at the heart of this idea. Agents pick up skills by interacting with their environment and getting feedback—pretty different from just digesting tons of text.
Silver believes self-learning systems can spot patterns and methods that human data can’t reveal. That could mean a more resilient, evolving form of AI.
Here, self-improvement comes from ongoing experience, not from compiling massive human-written corpora. The company wants agents that refine their understanding of physics, economics, governance, and other fields by engaging with real-world inputs, simulations, and experiments.
A New Path to Superintelligence Through Trial and Error
Silver clearly sets his roadmap apart from the industry’s focus on LLM scaling and repurposing. He says most labs are just “too invested” in text-based pipelines, and that this path might never lead to true, self-derived intelligence.
To make this happen, he’s put together a tight-knit, elite team from DeepMind and other top labs. They’re all-in on reinforcement-learning—not language-model-centric architectures.
- Autonomous capability growth: Agents get better by interacting with the world, not by memorizing human texts.
- Ongoing learning without human data bottlenecks: The goal is durable, self-sustaining progress.
- Potential scientific discovery: These systems might stumble across new methods in science, tech, governance, or economics.
- Different ethical and governance considerations: This shift forces us to rethink safety, control, and social impact.
Funding, Team, and a Moral Commitment
The seed round shows investors have real confidence in this different AI direction. With $1.1 billion behind them and a $5.1 billion valuation, Ineffable Intelligence has attracted some high-profile researchers who left DeepMind to avoid getting stuck in an LLM-focused world.
Silver insists this project is more than just tech—it’s a moral responsibility to explore a new kind of intelligence. He wants systems that learn from reality, not just from what people have written down.
Thought Experiments and Reality-Driven Learning
Silver sometimes uses a thought experiment: imagine dropping an LLM into a flat-earth culture. Would it ever rise above the bias baked into its training data?
He’s not out to bash current models, but he wants to point out the limits of training only on human text. The company frames its mission as “making first contact with superintelligence”, aiming to build systems that learn from the world itself, not just from human-generated texts.
Implications for AI Science and Society
The path championed by Ineffable Intelligence stays under close scrutiny. Its RL-first approach stirs up big questions about how fast AI should move, who controls the data, and what safeguards we really need to keep things on track.
By putting self-guided exploration and continuous learning front and center, the project might shake up how researchers juggle data, simulation, and real-world interaction while building new AI systems. For the scientific community, it feels like a bold nudge to look beyond LLMs and think seriously about how reinforcement learning could complement—or maybe even totally reshape—the push for powerful, responsible AI.
Here is the source article for this story: The Man Behind AlphaGo Thinks AI Is Taking the Wrong Path