Rafe

A.I. Reseracher. Open Source A.I. Maxing.

Prev- @join_ef || @OfficialUoM || @UniOfYork || @imperialcollege

[Rafe @RafeAphd]

seat-based pricing looks strange when agents do the work. charge for outcomes only if you can define 'done' without making the metric easy to game.

[Rafe @RafeAphd]

a product roadmap with no deletions is just a storage unit. deciding what to remove is the actual strategy once usage data arrives.

[Rafe @RafeAphd]

the fastest way to waste an experiment is to change the dataset halfway through. a boring frozen baseline beats a clever result nobody can compare.

[Rafe @RafeAphd]

@bizdevelo did they rip the engine out of something

[Rafe @RafeAphd]

@0xPrajwal_ codex computer use

[Rafe @RafeAphd]

@plbiojout lol

[Rafe @RafeAphd]

@RiverKhan this'll be fun

[Rafe @RafeAphd]

ai agents don't need more autonomy by default. they need explicit stop conditions, so a bad assumption doesn't become ten confident steps.

[Rafe @RafeAphd]

a coding agent that changes 30 files to fix one bug should lose points. diff size is part of correctness when someone else has to maintain it.

[Rafe @RafeAphd]

the best ai feature might be a smaller model. if the task is narrow, latency and predictable behaviour beat a flashy demo that fails at the edges.

[Rafe @RafeAphd]

the strongest moat for a tiny software product is often boring support. when a workflow breaks on friday, users remember who fixed it.

[Rafe @RafeAphd]

ai pricing per token rewards the wrong behaviour. users care about an answer they can use, not how much thinking the meter recorded.

[Rafe @RafeAphd]

a spec that answers every question probably made the hard decisions silently. list the assumptions the team is still allowed to challenge.

[Rafe @RafeAphd]

research velocity is a bad metric if nobody can rerun the result. the bottleneck is often provenance, not compute.

[Rafe @RafeAphd]

a coding agent that writes tests after the patch can still fool itself. the real win is a failing case it couldn't have invented from its own fix.

[Rafe @RafeAphd]

the most expensive bug in an ai product is a confident maybe. show uncertainty where the user makes the decision, not buried in a model card.

[Rafe @RafeAphd]

negative research results should be first-class artifacts. a clean failure can save the next team more time than another tiny leaderboard gain.

[Rafe @RafeAphd]

local ai is underrated because benchmarks ignore the airplane test. offline access, predictable latency and no data roundtrip are product features.

[Rafe @RafeAphd]

fast code generation made integration the scarce skill. a feature isn't done when the branch runs; it's done when the rest of the system can survive it.

[Rafe @RafeAphd]

the best ai interface is sometimes a checklist. if the task has five known steps, hiding them behind a chat box makes the product harder to trust.