Position

They Found the Moat

The 2023 Google memo everyone quotes for “we have no moat” also said exactly where the moat was. It named two things. Google could take neither.

On the most-quoted document in AI

In May 2023 an anonymous Google researcher posted an internal memo to a Discord server. It argued that neither Google nor OpenAI had a defensible position in AI, because open models were closing the gap faster than either could widen it. SemiAnalysis published it, verified its authenticity, and noted the obvious caveat: one engineer’s opinion, not company policy.

It became the most-quoted document in the field. Almost always for the title.

The headline reading is correct as far as it goes. Weights leaked. LoRA made fine-tuning cost about a hundred dollars and a few hours on consumer hardware. Small models iterated on repeatedly beat large models retrained occasionally. Three years on, that all held.

But the memo said two more things, and those are the ones worth reading now.

The first thing: curation beats scale

Buried in the argument about why open models were catching up is a claim about data: that carefully curated datasets outperform enormous unlabeled ones, and that the good curated sets were increasingly open.

Sit with that. In 2023, inside the company with the largest web index ever assembled, an engineer wrote down that size had stopped being the advantage. Not as a prediction — as an observation about what was already beating them.

Everyone read that paragraph as being about training corpora. It is equally true of retrieval corpora, and nobody acted on it, because acting on it means throwing pages away. Every incentive in a page-count-measured, ad-monetised business runs the other way. You do not get promoted for shrinking the index.

The second thing: the only advantage left requires letting go

The memo gestures at what might actually be defensible, and it isn’t a model or a dataset. It’s owning the ecosystem — being the thing a community builds on. It cites Meta’s accidental version of this: leak your weights, wake up as the platform.

And it attaches a price. Community leadership requires relinquishing tight control — which is a strange sentence to find in a memo from a company whose entire economic engine is control of the ranked list.

“Open source has some significant advantages that we cannot replicate.”

the memo, May 2023

Cannot. Not “have not yet.” The author understood that the barrier wasn’t engineering. Google could have built any of this in a quarter. What it could not do was give up the control that makes the revenue work.

So: two answers, both unreachable from inside

Take them together and the memo is more interesting than its reputation. It says the moat is not in the model — correct. It says quality of corpus beats quantity — correct, and Google can’t act on it. It says durable advantage comes from a community you must cede control to — correct, and Google really can’t act on that.

It is a document that locates the treasure and then explains, without quite saying so, why the people who found it will never dig.

Which is the whole product

We named this thing MoatGoat before that reading of the memo was obvious to us, and the joke has aged into a thesis. There is no moat in the model — we don’t train one, and the small open model we ship is replaceable by design. There is no moat in the pages either; ours come from Common Crawl and open dumps, and anyone can download the same bytes tomorrow.

The two things that aren’t copyable are the two things the memo named:

• Curation. Not which pages we have — which we threw away, and why, published as a ledger. An index built by exclusion is a set of decisions, and decisions don’t copy off a disk.
• The herd. The ranking signal is produced continuously by people who chose to be here. Steal the corpus and you get a snapshot going stale from the moment you take it; the part that keeps it alive isn’t in the file.

That’s also why giving up control isn’t a sacrifice here. There’s no ad auction to protect. The thing the memo said you’d have to relinquish is the thing we never had.

We might be wrong in the same shape

The memo’s real lesson isn’t its conclusion, it’s its posture: an honest inventory of your own position written by someone with no incentive to flatter it. So here is ours.

A curation signal is only a moat while it stays ahead of what a model can infer for free. If quality classification gets cheap enough that anyone can re-derive our judgements from the raw corpus in an afternoon, the curation half is gone — and the herd is the only thing left. If the herd stays small, there is no signal to be ahead with, and we are a worse Common Crawl with a nicer interface.

Both are testable, and neither is settled. What we won’t do is the thing the memo was written against — assume a moat exists because it would be convenient.