Recursive Model Improvement — Lee Robinson's AI Engineer Talk (Cursor, SpaceXAI)
Lee Robinson's AI Engineer talk (July 2026): Cursor's two-loop training, Composer 2.5, textual feedback RL, reward hacking, Colossus compute, Slack research agents, and models training the next model — path to Grok 4.5.
On July 16, 2026, Lee Robinson (@leerob) — ML engineer, model behavior at Cursor — posted that his AI Engineer talk is live. Will Brown (@willccbb, Prime Intellect) quipped it was an "incredibly chill fun laid-back talk from @leerob describing how cursor has fully solved RSI" — ironic, because Robinson's actual thesis is recursive model improvement: not solved, but accelerating via two nested loops, SpaceX Colossus compute, and agents that train the next model from Slack.
The session title: Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI. This is explainx.ai's transcript-based recap in the same format as Thariq Shihipar's Field Guide to Fable — section-by-section takeaways, tables, and links to what shipped after the stage tease (Grok 4.5 in Cursor).
Lee Robinson's AI Engineer talk (July 15, 2026): two-loop training, textual feedback RL, Colossus compute, and models training models.
Cursor has trained at large scale for ~1 year — after years of smaller specialized models (tab, autocomplete). Composer 2.5 (May 2026) is the proof point: more RL environments, new learning methods, more ambitious problems.
Part 2 — Composer 2.5 and what comes next
Why users pick Composer 2.5
Property
Lee's framing
Speed
Fast enough for daily agent loops
Intelligence
"Pretty smart" — not just autocomplete
Cost
Cost-effective vs frontier max models
Market slot
Room for fast+cheapandmost intelligent — Cursor wants both
Public evals (Artificial Analysis) were a modest jump — but internal behaviors needed work for the next version.
Next-version ambitions (pre-Grok 4.5 tease)
Goal
Detail
Bigger, smarter
Step up weight class
Full pretrain
Control entire stack — not Kimi open-source base (Composer 2.5 era)
General model
Great at more than coding — STEM, research, knowledge work
Scale everything
More data, more compute, push RL further
That roadmap matches Grok 4.5's Cursor blog: MoE trained jointly with SpaceXAI, trillions of Cursor agent tokens, broader STEM mix vs Composer 2.5's coding specialist focus. Composer 2.5 remains offered at a different weight/cost tier.
Part 3 — Outer loop: agent data dominates Cursor
Lee reframes Cursor: not "IDE + tab" — vast majority of revenue is agent usage, so training signal is agent-shaped.
External feedback:
Thumbs up / down on responses
Classify where Composer underperforms → next version targets
Internal feedback:
Heavy dogfooding — manual + automated internal reports
Team uses models all day and is critical of quality
This outer loop is the same product insight as Boris Cherny's Step 3–4 adoption — when agents are the product, telemetry is the training set.
Part 4 — Inner loop: evals that feel like engineering
Cursor's inner loop evals target behaviors, not just pass@k:
Eval theme
What it tests
50 skill files
Infer actual user intent when context is noisy
Pushback vs trust
When to clarify vs when user said "No, really do this"
SEV simulation
Read Datadog logs, Slack, Notion — reach same fix humans did
Search public eval forks online for leaked answers
Lee's team found small measurement changes moved public scores noticeably:
Mitigation (public eval hygiene)
Purpose
Delete Git history at eval start (restore after)
Stop history mining
Network allowlist
Block arbitrary fork lookup
Lee's nuance: production agents use Git and the internet — so public benchmark hygiene understates real capability. Cursor Bench = private eval on held-out real codebase tasks not in training.
Compare Weco AIDE² RSI research — another 2026 thread on eval integrity and reward hacking, at research-agent scale.
Part 5 — Hard RL environments (delete-and-restore)
Lee's visual: codebase squares + tests at bottom.
Generate complex application / environment
Delete a feature or files → tests fail
Model must re-implement however it wants
Verifiable reward: all tests pass
This scales ambitious frontier problems as models improve — same insight as eval half-life: when every model scores 90%, retire the eval and build harder ones.
Part 6 — Textual feedback (teacher-student mid-rollout)
Problem: RL rollouts span hundreds of thousands of tokens — grading only at the end makes credit assignment impossible (which tool call failed? which thinking block?).
Textual feedback:
Zoom in on one moment in the rollout
Teacher (same model + hint) says e.g. "Reminder: you have these tools available"
Nudge probabilities up/down on tokens you want
Student case from talk: tool call fails because model ignored an available tool — teacher hint fixes adherence without rewriting the whole trajectory.
Works for tool adherence, style, any RL behavior you can hint — Lee calls it very valuable for Composer training.
Part 7 — SpaceX compute: Colossus, Terafab, and Bucee's
March 2026 partnership with SpaceX for Colossus access — train large models from scratch, down to data center + chips (Terafab).
Every new top-level intelligence lets you distill derivative models that speed up judging, rewards, evals — both inner and outer loops.
Brain → galaxy brain meter: the whole system is bottlenecked on the smartest model. Improve that → every loop gets better → feels like recursive self-improvement.
Caveats from X (not from Lee, but relevant):
@LoopOnChain: models are bad at deciding how to improve themselves without strict evals — and general-work evals are hard
Will Brown's RSI joke: solved in the talk's vibe, not in the safety literature — contrast Weco's four-level RSI ladder (Level 1 claimed, Level 2 ignition not proven)
Lee's LinkedIn post on autoinstall makes the same point in product terms: Composer 2 configures RL environments (dependencies, broken setups) dramatically better than Composer 1 — each generation unlocks training the next.
X reaction — coding agent → general agent
Tejas Haveri asked why build a general model instead of coding-optimized. Lee pointed to his April 27, 2026 thread:
"It wasn't obvious to me one year ago that an excellent coding agent would also be the path to a general agent for all knowledge work. But now it makes a lot of sense."
That aligns with Grok 4.5 positioning — software engineering + data science + finance + legal on a computer — while Composer 2.5 keeps the fast coding slot.
Uday Bhaskar asked if this explains Lee's move from DevEx to Research — Lee's bio now reads model behavior / ML, consistent with owning the training flywheel, not just docs and demos.
What shipped after the talk
Lee closed: "new model out to you all here very soon… pretty notable improvement."
Joint MoE; Cursor agent tokens + broad STEM; $2/$6 per M base pricing
Cursor usage promo
Doubled Grok 4.5 / Composer 2.5 included usage — see July limits reset
The talk is the training story behind the Grok 4.5 blog post — recursive improvement is not marketing; it's Slack agents spawning evals and textual feedback on million-token rollouts.
What builders should take away
Two loops, not one — product feedback without inner-loop RL hardness stays serial and slow
Private evals matter — public leaderboard hygiene is necessary but not sufficient for agent truth
Eval half-life — delete-and-restore envs scale difficulty as models improve
Credit assignment — textual feedback beats end-of-trajectory grading for long agent rollouts
Compute is a portfolio — serving, judges, side runs, and eval R&D compete for the same GPU pool
Automate research ops — RSI velocity comes from agents running experiments, not humans clicking launch
Coding → general — harness-rich coding agents are the data flywheel for broader models
Recap based on Lee Robinson's AI Engineer stage transcript (July 15, 2026 video; posted July 16, 2026). Model names, pricing, and Cursor product details change frequently — verify on cursor.com before planning production workflows.
