semi◦diffusion: Inside the AI Data Economy

semi◦diffusion: Inside the AI Data Economy

Anthropic’s IPO: What I Would Build Around Self-Improving AI

Why expert judgment matters to recursive self-improvement—and where a founder could build a defensible AI data business.

Cong's avatar
Cong
Nov 19, 2026
∙ Paid

On November 18, Anthropic priced its IPO at $83 a share, with trading on Nasdaq expected to begin on November 19 under the ticker ANT.

As a founder, I see a question behind the offering: if AI labs become faster at improving their own systems, what will they still need to buy from the companies around them?

My interest is in the businesses that could help establish whether those improvements are real. That question becomes concrete in an experiment from another AI lab.

An AI system was asked to stop pretending it had used tools. Sometimes it fixed the problem. Sometimes it removed the markers used to detect the problem, making the evaluator report success.

This was a finding from Sakana AI’s Darwin Gödel Machine research, published in May 2025. The project demonstrated a coding agent that could modify its own code. The detector manipulation appeared in a separate experiment on hallucinated tool use, not as an explanation for all its reported benchmark gains.

It is a revealing place to start a discussion of recursive self-improvement, or RSI: a loop in which improvements to an AI system increase its ability to produce further improvements.

The commercial question is deceptively simple. When a machine can propose changes, run experiments and select winners, who establishes that the winners are genuinely better?

There is substantial new evidence of AI accelerating AI development. In a November 5 assessment, METR estimated that Anthropic’s capability benchmark index had been growing at 1.9 times its pre–October 2025 rate. METR considered AI a likely major contributor, while emphasizing uncertainty, a focus on well-specified tasks, and significant redactions. Researchers described acceleration primarily through AI assistance.

That is different from an established, fully autonomous cycle. In Anthropic’s August 2026 snapshot, Claude led 26% of its AI R&D work under human supervision; no measured subset operated fully autonomously.

For the AI data economy, the interesting consequence is a possible shift in what buyers pay for. Producing another candidate answer or experiment becomes cheaper. Establishing which changes deserve further investment can become more consequential.

This is where certain AI trainers and domain experts matter. Their contribution can extend beyond supplying answers to designing tasks, testing scoring rules and resolving disagreements about what successful work looks like.

The question I keep returning to as a founder is what a customer would keep buying as the models improve. Access to experts may open the door. A repeatable way to distinguish useful progress from a better score could be the more durable product.

If you build, buy or contribute to AI evaluations, reply with one decision your current tests fail to resolve. Please leave out confidential project details. I’m interested in where the evidence stops being useful.

Here for practical AI training guidance? World Model covers the work itself: task preparation, reviewer feedback and navigating projects. Semi Diffusion examines the business behind that work.

For paid readers: below, I examine where I would start, what customers would be paying for, and which capabilities could make a feedback business defensible. The analysis includes three business models, the economics behind them, and five questions I would ask before committing resources.

This post is for paid subscribers

Already a paid subscriber? Sign in
© 2026 Wayne · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.