A brand-new AI lab raising $1 billion right out of the gate isn’t “confidence.” It’s a bet. And to me it’s the kind of bet that can either age like genius or like pure vanity, with very little room in the middle.
Here’s what’s been shared publicly: David Silver, a well-known DeepMind veteran, has launched a new London lab called Ineffable Intelligence. It reportedly raised $1 billion and is valued around $4 billion. The pitch is basically: today’s big language models can remix what humans already wrote, but they can’t truly innovate. His lab wants to build something “radically new,” using a different approach centered on reinforcement learning.
That’s the fact pattern. The interpretation is where it gets spicy.
Calling language models “limited” is both fair and also a little convenient. Fair because anyone who has used them hard knows the ceiling: they can sound confident while drifting, they can’t reliably prove they’re right, and when you push them past the training-data comfort zone, they start guessing. Convenient because “LLMs can’t innovate” is also a great story to tell investors if you want permission to go build the next big thing without shipping anything for a long time.
And that’s my first judgment: raising $1 billion early is a risk to the product, not just a fuel source. It changes the incentives. When you start that huge, you don’t just need to be right. You need to be spectacular. You need a narrative big enough to carry years of uncertainty. “Superintelligence” does that. It’s also the kind of word that can excuse almost any delay.
Still, I don’t think this is automatically hype. Silver has real credibility, and reinforcement learning has a real track record in certain domains. If you’re trying to get a system to learn by doing—by trial, error, feedback, and strategy—this isn’t some random pivot. It’s a philosophical argument about what intelligence is: not just talking, but acting. Not just predicting the next word, but choosing the next move.
If you’ve ever watched a team adopt an AI tool at work, you can feel the difference. An LLM is great for drafting, summarizing, brainstorming, and speeding up the boring parts. But imagine you’re running customer support and you want an AI that doesn’t just write responses—you want it to actually run the operation: route issues, decide when to refund, detect patterns, test new policies, and learn what reduces repeat complaints. That’s not just language. That’s a loop: act, see results, adjust.
Or imagine drug discovery, or supply chains, or energy use in a city. The promise isn’t “it can write better.” The promise is “it can decide better.” If that’s the direction, I understand why someone would say the current approach is not enough.
But here’s the consequence nobody should downplay: systems that learn by acting can also learn by breaking things. In software, “trial and error” is cheap. In the real world, trial and error can mean lost money, safety issues, and people getting hurt. Even in a controlled setting, you’re training something to pursue goals. The hard part is not making it smart. The hard part is making it loyal to what you meant, not what you said.
And yes, language models have their own risks. They can mislead, they can be used for scams, and they can automate misinformation. But a system designed to plan and execute is a different animal. A smooth talker can waste your time. A capable actor can change your environment.
There’s also a power angle here. A $1 billion seed-sized round (or whatever we call it now) is a sign that building cutting-edge AI is becoming even more concentrated. A tiny number of founders with the right background can command massive resources. That might speed up progress. It also means the direction of this technology gets set by a small group of people who already sit close to the center of the AI world.
If this works, who wins? The lab, obviously. The investors. The companies that get early access. The governments that can lean on it for strategic advantage. If it doesn’t work, who loses? Not just investors. It’s the rest of the field that has to compete with a giant that can hire anyone and buy a lot of compute to chase a theory that might not pan out.
And I’ll say the quiet part: “LLMs can’t innovate” is not a settled claim. Humans “innovate,” sure, but we also remix. A lot of breakthroughs are rearrangements of existing ideas, seen from a new angle. If an LLM can generate a thousand variations and a human picks the one that matters, is that innovation or just speed? If a system proposes a new approach that no human would have written down, but it arrived there by patterning from the past, do we care how it got there if it works?
I’m not certain what Ineffable Intelligence will actually build. The public description is ambitious and vague in the way big-money AI projects often are. “Radically new” can mean genuinely new, or it can mean “we don’t have to explain it yet.”
But I do think this is a real signal: the industry is trying to move from “talking machines” to “doing machines.” That shift is where the real stakes are. Because once these systems don’t just answer questions—once they start making choices—mistakes stop being embarrassing and start being expensive, or dangerous, or politically explosive.
So here’s the debate I actually care about: should we be pouring $1 billion into chasing “superintelligence” now, or should we be forcing the best minds in AI to spend the next few years making the systems we already have more honest, more controllable, and easier to audit?