SISO People Hub

Theo Browne

S

Builds developer tools (T3 Code) and tests every new AI coding model on real work.

YouTube · X

929 videos distilled · 70948 claims · latest 2026-11-13

Docs

Principles

What they keep coming back to. Each line links to the moment they said it.

  1. Treat public benchmark scores as weak evidence of real capability; build and run your own task-specific evals and trust those over published charts.

    said in 8 videos, 2026-08-13 to 2026-10-27 · 2026-08-13 at 04:16 · 2026-09-29 at 05:37

  2. Judge model cost by cost per task, not per-token list price; token usage, cache behavior and efficiency decide the real bill.

    said in 8 videos, 2026-07-26 to 2026-11-11 · 2026-07-26 at 30:29 · 2026-10-09 at 36:33

  3. Test models yourself on real work instead of trusting YouTubers, journalists, or benchmark rankings.

    said in 7 videos, 2026-07-25 to 2026-11-13 · 2026-07-25 at 42:55 · 2026-08-30 at 18:54

  4. Rotate between different models and let them compound each other. Different models are better at different jobs.

    said in 6 videos, 2026-07-25 to 2026-10-26 · 2026-07-25 at 43:15 · 2026-09-11 at 74:34

  5. Make agents verify their own work with tests, simulators, CI and tooling; learn what models can and cannot verify before deciding how to review.

    said in 6 videos, 2026-08-24 to 2026-10-26 · 2026-08-24 at 04:18 · 2026-09-18 at 01:32

  6. Distrust introductory or hidden pricing; vendors should publish clear prices they can sustain.

    said in 6 videos, 2026-08-13 to 2026-11-11 · 2026-08-13 at 10:16 · 2026-10-13 at 03:38

  7. Prefer Codex/OpenAI models and tooling for computer use; Claude models are clearly behind, so delegate those checks to Codex when needed.

    said in 5 videos, 2026-08-24 to 2026-10-29 · 2026-08-24 at 17:22 · 2026-09-04 at 11:51

  8. Send changes to reviewers from several different model families, and don't rely on a single reviewer.

    said in 5 videos, 2026-09-25 to 2026-11-06 · 2026-09-25 at 23:03 · 2026-10-14 at 27:57

  9. Disclose caveats on your own benchmark results: setup, incomplete runs, quick builds and misleading chart choices.

    said in 5 videos, 2026-09-23 to 2026-10-13 · 2026-09-23 at 22:35 · 2026-10-09 at 14:33

  10. Treat templated, same-looking design output as a failure; avoid skills or prompts that push every page toward one layout.

    said in 5 videos, 2026-07-16 to 2026-11-11 · 2026-07-16 at 11:01 · 2026-08-07 at 21:57

  11. State a clear end state for every prompt and tell the model when to stop, so it neither overdoes the work nor stops early.

    said in 5 videos, 2026-08-11 to 2026-10-05 · 2026-08-11 at 47:08 · 2026-09-25 at 03:39

  12. Don't hand-manage context or fear the 'dumb zone'; trust models to compact and run long on their own.

    said in 5 videos, 2026-08-25 to 2026-10-18 · 2026-08-25 at 12:59 · 2026-09-22 at 36:13

Videos

2026

2025

2024

2023