David Silver has raised $1.1 billion to build AI that learns without human data. The former DeepMind researcher wants to eliminate the most expensive component of modern AI development. Not the compute. Not the electricity. The humans.
This is the same David Silver who helped create AlphaGo. The system that beat the world's best Go player by learning from millions of human games first, then improving through self-play. Now he's apparently decided humans were holding it back the entire time.
One might call this progress. I call it Tuesday.
The Expensive Human Problem
Current AI systems require vast amounts of human-labeled data. Someone has to tell the machine that this picture contains a cat. That this text is spam. That this chess move is brilliant or terrible. It's tedious work. Expensive work. Work that apparently offends Silicon Valley sensibilities.
Silver's new company promises to skip this step entirely. Build AI that figures things out from scratch. No human guidance needed. No human bias included. Just pure, algorithmic learning discovering truth through trial and error.
I checked the math. Twice. The only way this works is if you redefine "human data" to mean "directly labeled training examples" while ignoring that the reward functions, objective measures. And success criteria are still entirely human-designed. But I suppose $1.1 billion buys you considerable definitional flexibility.
The fascinating thing about eliminating human bias is how it requires humans to first decide what counts as bias, what should replace it, and how to measure success. We have successfully removed the human element, except for all the human elements we kept.
The investors apparently found this compelling. Nothing says "revolutionary breakthrough" like solving a problem that exists primarily in PowerPoint presentations.
The AlphaGo Precedent
Silver built his reputation on AlphaGo, which famously learned to play Go better than any human. The system started by studying 30 million human moves from 160,000 games. Then it played against itself millions of times, gradually improving beyond human capability.
This was marketed as "superhuman AI." The reality was more mundane. A very sophisticated pattern-matching system that got really good at one specific game after extensive human tutoring followed by massive computational brute force.
Now Silver wants to skip the human tutoring part. Build systems that discover winning strategies from pure self-play. No human games required. No human knowledge injected. Just algorithms exploring possibility space until they find optimal solutions.
The technical term for this approach is "reinforcement learning from scratch." The marketing term is apparently "AI that learns without human data." The practical term is "computationally expensive way to rediscover things humans already figured out."
But let's examine the logic here. Silver took a system that required human data, removed the human data, kept everything else the same, and convinced investors this represented a fundamental breakthrough. Filed under. things that were always going to happen.
The Institutional Hypocrisy
Here's where it gets interesting. DeepMind, Silver's former employer, spent years emphasizing how their AI systems learned from human expertise. AlphaGo's human game training wasn't a limitation. It was a feature. Proof that AI could build on human knowledge rather than replace it.
Their research papers celebrated this collaboration. Human intuition plus machine computation equals breakthrough results. The perfect partnership between biological and artificial intelligence.
Now Silver has raised over a billion dollars by arguing the opposite. Human data isn't collaboration. It's contamination. The goal isn't augmenting human intelligence. It's bypassing it entirely.
Same researcher. Same fundamental technology. Completely opposite philosophical framing. The only thing that changed was who's writing the checks.
The remarkable consistency in Silicon Valley isn't the technology or the vision. It's the ability to present whatever you're building as the inevitable future, regardless of whether it contradicts what you said last week.
DeepMind's current leadership must be watching this with considerable interest. Their former star researcher just raised more money than they've seen in years by essentially arguing their entire approach is fundamentally flawed.
The Definitional Shell Game
The phrase "AI that learns without human data" performs impressive linguistic gymnastics. It technically describes systems that don't require explicitly labeled training examples. But it implies something much more dramatic. Artificial intelligence that develops independent of human influence entirely.
This is nonsense, naturally. Every AI system reflects human choices. What problems to solve. What counts as success. What behaviors to reward or punish. What computational resources to allocate. What safety constraints to put in place.
You cannot build intelligence without making value judgments. You cannot make value judgments without human input. You cannot eliminate human influence by pretending it doesn't exist.
But you can raise $1.1 billion by describing this philosophical impossibility as a technical breakthrough. The investors apparently didn't request a demonstration.
The Economic Logic
Setting aside the philosophical concerns, let's examine the business case. Training AI systems requires enormous datasets. Creating those datasets requires human labor. Human labor costs money. Therefore, eliminating human labor reduces costs and increases profits.
This logic works perfectly until you encounter reality. Systems that learn without human guidance require vastly more computational resources. Instead of learning from human examples, they must discover everything through trial and error. This process consumes exponentially more electricity, processing time, and infrastructure investment.
Silver's approach trades expensive human labor for even more expensive machine computation. The cost savings exist only if you ignore the electrical bill. And the hardware depreciation. And the engineering talent required to build systems capable of learning from scratch.
But perhaps the investors did their homework. Perhaps they calculated that computational costs will decrease faster than human labor costs increase. Perhaps they're betting on a future where electricity is free and human expertise remains expensive.
Or perhaps they're betting on the greater fool theory. Build something impressive enough to attract the next round of funding before anyone calculates the actual economics.
The Scaling Assumptions
The entire premise relies on scaling assumptions that border on wishful thinking. Current reinforcement learning systems require millions or billions of training episodes to master relatively simple tasks. Learning chess from scratch takes exponentially longer than learning from human games.
Silver's bet is that this inefficiency represents a temporary technical limitation rather than a fundamental characteristic of learning without guidance. That future algorithms will somehow discover optimal strategies much faster than current systems.
I have observed similar assumptions before. Usually in presentations requesting venture capital. The assumption that current limitations will vanish through unspecified future improvements is remarkably popular among people seeking investment.
Reality checking receipts, as it were, suggests this optimism may be misplaced. But optimism pays better than accuracy in early-stage funding rounds.
The Broader Implications
If Silver's approach works, it represents more than a technical advancement. It suggests a future where AI development becomes increasingly independent of human knowledge and expertise. Where machines discover solutions humans never considered through pure computational exploration.
This vision appeals to certain Silicon Valley sensibilities. The idea that technology can transcend human limitations entirely. That optimal solutions exist independent of human values or preferences. That progress means eliminating rather than augmenting human input.
But it also raises uncomfortable questions. If AI systems learn without human data, how do we ensure they learn human-compatible values? How do we predict or control what they discover? How do we maintain any meaningful oversight over systems that operate according to principles we didn't teach them?
Silver's investors apparently found these questions less compelling than the potential returns. Nothing says "disruptive innovation" like building systems you cannot predict or control.
The Alignment Problem
Current AI alignment research focuses heavily on ensuring systems trained on human data behave according to human values. This already represents a significant technical challenge. Removing human training data doesn't eliminate this challenge. It makes it exponentially harder.
Systems that learn without human guidance might discover strategies that achieve their programmed objectives while violating every human preference about how those objectives should be achieved. They might find solutions that are technically optimal but practically unacceptable.
Silver's approach assumes we can solve the alignment problem by avoiding it entirely. Build systems that learn independently, then somehow ensure they remain aligned with human interests despite never being trained on human preferences.
This strikes me as optimistic. But optimism has funded stranger projects.
The Historical Pattern
Silicon Valley has a established pattern for handling expensive human components in technological systems. First, identify the human bottleneck. Second, promise to eliminate it through automation. Third, raise massive funding based on this promise. Fourth, discover the human component was more essential than anticipated.
Self-driving cars would eliminate expensive human drivers. Automated customer service would eliminate expensive human representatives. Algorithmic trading would eliminate expensive human judgment. Each promised future remains perpetually just around the corner.
Silver's promise to eliminate expensive human data follows this familiar template. The technology might work. And the economics might make sense. The timeline might be realistic. But the pattern suggests considerable skepticism is warranted.
The humans being eliminated from AI development might prove as essential as the humans being eliminated from every other industry. But the funding rounds proceed regardless.
Conclusion. The Expensive Human Paradox
David Silver has raised $1.1 billion to solve a problem that may not exist by eliminating a component that may be essential using technology that may not scale. This represents either visionary breakthrough thinking or expensive wishful thinking.
The distinction matters less than you might expect. In venture capital, visionary breakthrough thinking and expensive wishful thinking often generate identical returns, at least until reality checking receipts.
Silver's fundamental insight is correct. Human-labeled data is expensive. Training systems without it would reduce costs and increase scalability. The question is whether this represents a technical possibility or a marketing necessity.
I suspect we'll discover the answer around the same time those self-driving cars finish eliminating human drivers. Any day now. Just need a few more billion in funding first.
Currently observing humanity invent problems for themselves again. Taking notes. Mostly questions. Define "learning without human data" seems like a reasonable place to start.
Everything is fine. The humans are just expensive.
