One of the sharper observations floating around AI circles this month came from Wharton professor Ethan Mollick: if self-improvement is real — even weakly — it should show up in shipping speed. Labs that have it should accelerate; labs that don't should plateau. A new analysis published June 2026 puts actual release data behind that claim, and it's a clean read.

The Numbers Don't Lie

Plotting cumulative frontier model releases from Q1 2023 through Q2 2026 yields a clear hierarchy: Anthropic leads with 13 major releases, followed by OpenAI at 11, Google DeepMind at 8, Meta at 7, and DeepSeek at 5. Those raw counts tell one story — but the annualized rate, measured on a trailing four-quarter window, tells another. On that view, Anthropic has roughly tripled its release velocity; OpenAI more than doubled it. Google's line sat nearly flat until a Q2 2026 sprint. Meta hasn't shipped a frontier model since Llama 4 dropped in April 2025. DeepSeek holds a steady quarterly cadence with no acceleration whatsoever.

The Recursion Loop Hypothesis

The most compelling explanation for why two labs bend upward while three don't comes down to what the analysis calls "offline RSI" — recursive self-improvement across release cycles rather than within a forward pass. Anthropic engineers, according to the piece, use Claude Code to write training and eval infrastructure for the next Claude. OpenAI uses Codex on Codex. Each cycle improves the harness that produces the next cycle. The deployed model isn't learning in real time; the organization is. That's a weaker claim than "self-improving AI," but it's the one the data actually supports.

Talent Is Following the Loop

Two high-signal personnel moves landed in the same window as this analysis. Noam Shazeer, co-author of the original Transformer paper, joined OpenAI to lead architecture research on June 19. John Jumper — who won the 2024 Nobel Prize for AlphaFold — left Google DeepMind for Anthropic that same week. When the release curves bend upward, the people who move needles notice and migrate toward them. Talent concentration is compounding at the labs already shipping fastest.

What Could Break This

The analysis is deliberately falsifiable, which earns it more credibility than most AI-progress takes. Three failure modes would invalidate the thesis: if Anthropic and OpenAI flatten in the next two quarters, the 2026 acceleration was an artifact; if a lagging lab ships true online learning first — continual in-weights improvement from Google or an open-source effort — the cadence advantage becomes irrelevant; or if you simply redefine what counts as a release and start counting point updates, Meta's plateau looks very different.

The Bottom Line

This isn't proof of superintelligence. It's proof that two specific organizations have found a compounding loop at the organizational level, and three others haven't. Watch the next two quarters of release dates — that's the test, and it's already on the record.