How to be more ambitious with models (Rieseberg, 2026-10-13)
Thesis
People under-use models because they don't believe bigger outcomes are possible (a self-efficacy problem). He names four kinds of ambition. Speed is obvious and he skips it. The other three are Altitude, Depth and Range.
Practices (paraphrased)
- Altitude. Ask for the goal behind the task, for example "keep the apps crash-free" instead of "fix this crash". This only worked once the model had plumbing: crash reports, a map from app version to git commit, and Windows/macOS/Linux machines to reproduce and verify. It works poorly for creative work. A good sign it fits: the work would otherwise need meetings.
- Depth. Make the model use an expert's real tools (databases, statistical models), not the expert's vocabulary. Set unreasonable goals: if you need 5s, ask for 200ms and say it's solvable. Have the model test beliefs against raw data and confirm new signals on held-out data.
- Range. "Replace the nouns with verbs," then assemble the crew (analyst, editor, short seller, forensic accountant). "Don't ask the model to write a short seller, ask it to do what a short seller does." Role-play makes the transcript weirder without improving the outcome.
- Outside judge. "Do not ask the model to judge itself." Validation should come from outside the model, ideally a tool it can run for immediate feedback — the same claim [[verifier-as-epistemology]] makes about cheap, mechanical verifiers: knowledge is what an independent check can replay, not what the asserter states.
- Deeper, not bigger, prompt (4 parts). The model names (1) what a top professional would do, (2) the rawest starting material, (3) the specific methods and tools, and (4) what, other than you or the model, would prove the result right. It then rewrites your prompt to ask for that.
Why this is in the vault
A named, structured vocabulary (Altitude/Depth/Range + "outside judge") for a pattern RDCO already runs ad hoc — this is the clearest independent third-party confirmation seen yet that the studio's own fresh-eyes-critic discipline is the right default, not a self-imposed constraint.
Mapping against Ray Data Co
- Independent backing for the studio's fresh-eyes critic gate ([[feedback_verification_independent_worker_pattern]]) and for behavior-critic.
- "Altitude needs plumbing" is a clean demo line for the founder's "agents in production" credibility lane ([[project_credibility_for_phdata_sales]]). An agent's ambition is capped by what it can touch and check.
- The 4-part prompt is usable material for the founder's phData Claude in Practice sessions.
- Ties to [[2026-10-14-gregisenberg-sf-48-hours-21-observations|Greg Isenberg's #6]] (2026-10-14: a vertical harness is data connectors, evals, permissions and a review screen; "the model is maybe 20%").
Related
- [[2026-10-14-gregisenberg-sf-48-hours-21-observations]] — same vertical-harness/model-is-20% framing from an independent source
- [[2026-05-04-indy-dev-dan-pi-coding-agent-reviews-like-you]] — the verifier-agent pattern this note's "outside judge" rule backs up
- [[2026-06-14-open-ended-output-verifier]] — the open research question on which automated verifier is reliable enough to gate on, directly relevant to operationalizing "do not ask the model to judge itself"