The $11B Bet That Says Everything We Know About LLMs Might Be Wrong

Last week, a company that doesn't have a product, doesn't have a website, and is barely two months old raised $11 billion — the largest seed round in European history.

The company is called Ineffable Intelligence. The founder is David Silver — the man behind AlphaGo and AlphaZero at DeepMind. The valuation is $51 billion.

And here's the part that stopped me: they're not building another LLM.

The Bet

Every major AI company today — OpenAI, Anthropic, Google — is playing the same game:

1. Scrape as much human text as possible from the internet
2. Train a massive transformer to predict the next token
3. Use RLHF to make it sound helpful and harmless
4. Scale up and pray for emergent capabilities

Silver's thesis is that this entire paradigm is a dead end. Not a slowdown. A dead end.

His argument: language models learn from static human data. They're a compressed snapshot of what we already know. They can mimic reasoning, but they never actually interact with the world. They've never tried something, failed, adjusted, and tried again.

Real intelligence doesn't come from memorizing human text. It comes from acting in an environment and learning from the consequences.

If Silver is right, the entire "scale is all you need" school of thought is wrong. Bigger models, more data, more compute — eventually hits diminishing returns. The path to real AI isn't a bigger transformer. It's a fundamentally different architecture.

The $11B Question

The investors — Sequoia, Lightspeed, Google, NVIDIA, DST Global, the UK sovereign fund — are betting $11B that Silver is right.

But this is a moonshot, and everyone knows it. Silver's approach — training AI purely through reinforcement learning in continuous environments — has never been demonstrated at scale. AlphaGo and AlphaZero worked beautifully in the constrained world of board games. The real world is infinitely messier.

The skeptics say: "We've had RL for decades and it hasn't produced anything close to GPT-4. Why would this be different?"

The believers say: "Because nobody has tried it at this scale with this much compute."

Both are valid. That's why it's a $11B bet, not a $11M sure thing.

What This Means for Builders

This is the part I actually care about. Does this change anything for someone building products on top of AI today?

Yes, but not in the way you'd expect.

1. The LLM era is not the final form of AI

If you're building a business that depends entirely on API access to GPT or Claude, it's worth asking: what happens if the underlying paradigm shifts in 3-5 years?

This doesn't mean stop building. It means build abstractions. Your product shouldn't be coupled to one model provider. Your value prop shouldn't be "we use GPT-5." It should be "we solve X problem."

2. "Scale is all you need" is being questioned at the highest level

The fact that $11B flowed to an anti-LLM thesis tells you that even the biggest investors are not fully confident in the current path. They're placing a hedge — and that hedge is a $51B company.

If you've been feeling like the "just scale it" narrative was too simplistic, you're not alone. The industry is quietly admitting the same thing.

3. For indie developers: this doesn't change your game

Whether AI runs on transformers or reinforcement learning doesn't matter for the tools you're building today. What matters is: can you solve a real user problem?

The AI layer is becoming a commodity either way. The winners in the next wave won't be the ones with the best model — they'll be the ones with the best product and the best distribution.

The Bigger Picture

This isn't just about David Silver. It's a signal that the industry is entering what I'd call a post-scaling era.

For the last 5 years, the playbook was simple: take a transformer, make it bigger, feed it more data, get better results. That playbook is showing cracks. GPT-5 hasn't shipped. The gap between GPT-4 and GPT-4.5 was noticeably smaller than GPT-3 to GPT-4. Diminishing returns are real.

The industry is splitting into three paths:

Path	Who	Thesis
A — Keep scaling	OpenAI, Anthropic, Google	More compute, more data, more modalities
B — Start over	Silver, Ineffable Intelligence	RL from first principles, different architecture
C — Build on top	Everyone else	Model layer is a commodity; product and distribution win

As product builders, Path C is where we live. And that's fine — because the most valuable companies of the last 10 years weren't the ones that built the infrastructure. They were the ones that built on top of it.

---

What do you think — is Silver's bet visionary or delusional? I go back and forth every time I think about it. Drop a comment, I read all of them.

---

I'm a solo developer building MultiPost — a content repurposing tool for indie hackers who write on multiple platforms. I write about AI trends, SaaS development, and what it's actually like trying to get to $500 MRR as a solo founder.