The puzzle pieces are in different heads.

Tim Berners-Lee's most underquoted line, revisited in the LLM era — and why agents change the math on superconnecting.

In 2005, Tim Berners-Lee gave an interview that didn't go viral but should have. Asked what the web was still missing, he said something like this:

"There are millions of scientists trying to cure the likes of AIDS and Alzheimer's. Maybe the cure is currently separated in different people's heads. How can we design the web so that these half-formed solutions can come together?"

Sir Tim was naming a structural property of human knowledge that nobody had a good answer for — and still doesn't. Important problems rarely sit, whole, in one head. They're distributed. Half a solution lives with a molecular biologist in Basel. The complementary half lives with an immunologist in Boston. Neither one knows the other exists. The puzzle stays unsolved not because it's unsolvable, but because the pieces can't find each other.

Why search doesn't solve this

Search is what you use when you know what you're looking for. It fails hard in the exact case Sir Tim described: when you don't know that your piece is a piece. You can't search for the half-solution in someone else's head, because you don't know it exists, and they don't know yours does either.

That's why serendipity — the dinner party, the unexpected introduction, the intellectually promiscuous mutual friend — has been doing the heavy lifting. Human superconnectors. People whose job, often informal, is to sit at the crossroads of many conversations and notice when two fragments look like they might fit together.

Superconnecting is a scarce, high-value, fundamentally under-automated activity. Until now.

What LLMs change

The thing LLMs are genuinely good at — and we don't always give them credit for this — is noticing semantic overlap across huge amounts of text. Two people wrote about the same idea in different vocabularies across a decade? An LLM sees that instantly. That's exactly the capability Sir Tim's vision needed and didn't have.

But by themselves, LLMs are trapped in each user's jar. Your Claude doesn't know what my Claude knows. Your research log doesn't cross-pollinate with mine. The fragments still can't find each other, not because the models can't see the match, but because the substrate they share is zero.

The substrate is the product

What we're building at Ideaflow is the thing Sir Tim was pointing at: a shared substrate where half-formed solutions can find their complements. Humans contribute fragments. Agents contribute fragments. A typed knowledge graph sits underneath, so the matching can be mechanical instead of lucky.

Three people researching a problem — one in a notebook, one in a Slack DM, one in a meeting transcript — should not have to be introduced by a mutual friend for their fragments to meet. The shared graph should do that. The agents should do that. The dinner party can go back to being about the food.

We're very early in actually making this real. But the framing is load-bearing. Knowledge work is the last major human activity where "did two people meet at the right moment" is still the rate-limiter. It doesn't have to be.