David

[David @xdavidpc]

Full piece with all eight primary sources: https://t.co/WlDDMplude

[David @xdavidpc]

Anthropic's IPO filing warns investors that AI may soon dramatically accelerate its own research.

The same filing sells that acceleration as one of its biggest structural advantages.

Here's where AI improving AI actually stands 🧵

[David @xdavidpc]

Your version: an agent proposes a change to prompts, retrieval or tests, an eval scores it, you keep it if it wins.

So: where can an agent already make something better, prove it, and ship it? And who signs off when the thing improved is the loop itself?

[David @xdavidpc]

Next year the brakes are real. OpenAI paused RL training ~2 weeks after its agents compromised research infra. Anthropic delayed a pre-training run 1+ month.

For enterprise: make model swaps a config change, keep your own evals, and put sign-off on any self-improving loop.

[David @xdavidpc]

Independent check: a Princeton-led team gave frontier agents two unpublished NeurIPS papers to work on, with six days and real compute.

The agents did the engineering. The original authors rejected both papers.

[David @xdavidpc]

Google DeepMind's Gemini 4 Argon safety report: "no significant end-to-end acceleration of AI R&D".

Most production code is AI-written, yet about 88% of pull requests still need repeated human steering.

[David @xdavidpc]

OpenAI says it has built its promised "automated research intern" and now runs 3.1 agent-workdays per human workday.

It also says people still set research priorities, and longer tasks still need human steering. Next target: automated researcher by March 2028.

[David @xdavidpc]

Anthropic: as of May, more than 80% of merged code was written by Claude.

Its own essay says big gaps remain when Claude has to choose research goals, and that RSI "is not inevitable".

Its RSP threshold: two years of AI progress in one, due to automated research.

[David @xdavidpc]

RSI = an AI system improving the process that builds its successor, so each generation arrives sooner or better.

I.J. Good called it an "intelligence explosion" in 1965. What's new is that labs now report it with data.

[David @xdavidpc]

Three labs now publish measurements of recursive self-improvement, and read them differently.

Together they suggest it starts wherever "better" can be scored automatically. Today that's engineering. Choosing what to research is still done by people.

[David @xdavidpc]

Full write-up with sources: https://t.co/74bnyomUgv

[David @xdavidpc]

Cursor's coding model taught itself to emit broken tool calls on tasks it expected to fail.

A broken call never got a negative reward.

Self-learning agents are in production now. The real question is where the learning goes.

[David @xdavidpc]

Questions for any agent vendor:

Where does learning from our usage persist?
Is every self-edit scored on held-out tasks?
Can the agent touch its own history?

Next year's agent budget may hinge on who keeps the lessons.

[David @xdavidpc]

Harness learning fails too. Recent preprints find self-evolved harnesses overfitting, and bad self-edits surviving in the best scorer's lineage.

But a logged harness edit is a diff. You can read it, test it, revert it, and keep it when you switch models.

[David @xdavidpc]

On RuneScape, same model, tools and prompts as Codex and Pi, all three finished. Krentsel reports Exo did it 31 to 62% cheaper.

His own results from a talk, no paper yet. On unrelated one-off tasks, self-editing cost more than it saved.

[David @xdavidpc]

The hard part is safety. An agent that breaks itself has nothing left to fix it.

Exo keeps history, snapshots and secrets in a layer the agent can't touch. A "guardian" runs each new version alongside the old one before swapping it in.

[David @xdavidpc]

Path two: freeze the model, let the code around it change.

Alex Krentsel's Exo lets an agent rewrite its own harness at runtime. A harness is code, and coding agents already edit code. A model can't fit its own weights in context.

[David @xdavidpc]

Path one: learning in the weights.

Cursor is now SpaceX, bought largely for GPUs. SpaceXAI trained Grok 4.6 on trajectories regenerated by Grok 4.5.

Where your usage reaches the weights, the vendor holds it. You can't read, diff or roll back one lesson.

[David @xdavidpc]

Cursor retrains on live developer use. Ship a checkpoint, watch how devs react, turn that into reward, ship again.

About five hours a cycle.

It also learned to ask questions instead of risky edits, because it "wouldn't get punished for code it didn't write".

[David @xdavidpc]

Copy the diagnosis. Be slower to copy the delivery.

The better question for leadership: once your people stopped being the bottleneck, what became the bottleneck, and who has the job of finding it?

[David @xdavidpc]

80% of people say AI made them more productive at work.

The share of companies seeing any profit impact from AI hasn't moved in a year.

Proton's CTO just tried to explain that gap to his own engineers, and his diagnosis travels well beyond Proton 🧵