▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=iBrAWpjXNxs
• Original title :: Please stop using stupid models

Theo answers a claim by Sentry founder David Cramer that switching back from frontier models (Fable, Astra) to older high-reasoning ones (Opus, Soul) makes no practical difference. His counter: people who don't notice a difference are prompting narrowly. Frontier models matter not because their ceiling is higher but because their floor is higher — they are stupid less often, which compounds exponentially over long autonomous runs.

Ceiling vs floor

Theo's mental model is a quality curve per model over many responses. Astra's peaks exceed anything Fable can do, but its troughs drop to Gemini-Flash levels — it has been confused about what "revert" means and has failed to center an icon in a div. That spikiness makes it hard to calibrate how ambitiously to prompt. Fable's appeal is the opposite: the gap between best and worst narrows and the floor rises.

The common error is judging models by their peak. Novel math or a 3D Blender scene is irrelevant if the model misreads a plain instruction. "I don't like Fable because it's way smarter. I like Fable because it's less dumb" — and less dumb and smart are nearly opposite ends of a spectrum.

Width of prompts, not depth

The axis that exposes the difference is prompt width — how far the model runs unattended from start to finish — not task difficulty. A Jira ticket that names the files and the wrong behavior clears everyone's bar; Cramer is right about that case. Theo's target case is different: screenshot a user-reported bug from a DM, tell the agent to fix it, test it, record a video proving it works, open the PR, and babysit review comments — with no intervention.

As the window widens, the chance of hitting one of the model's low points grows. He works the math: 5% failure per 10 minutes gives a ~73% failure rate at one hour and worse at four. Cutting that per-window failure rate from 5% to 3% turns a ~70% four-hour failure rate into ~50%. Small floor improvements translate into exponentially longer viable runs.

His own logs back this: median prompt duration went from 53 seconds to 2:20, P95 from just under 7 minutes to over 16 minutes, with the May–June jump (Fable and Soul) nearly doubling P95 month over month.

Practical advice

• Stop watching the agent work. Watch only if the first output is badly wrong, then ask the model what signals led it astray.
• Learn from bad runs, and adjust the codebase far more than the prompts. If agents keep tripping on the same speed bump, a new human dev would too — that's on the codebase.
• Overnight autonomous runs are valuable mainly as failure-discovery: they surface the rough edges you then smooth out.
• Stop naming files in prompts. Describe the problem and let the agent find the files.
• Push the boundaries outward on both ends: have the agent run the simulator, push the build, attach a verification video to the PR — so a human is pulled in only when the work is likely already correct.

Who this misses

Theo's diagnosis of Cramer and David Khourshid is that they're prompting like it's February and undervaluing their own time. Talented leaders who already gave up day-to-day coding are the most resistant: they surrendered the code-writing middle but still hold onto the prep before and the vetting after, so they test models only on the part they had already delegated — where the models have been adequate for six months. The two metrics he wants people to optimize are how long a model can run without input, and how likely the work is correct by the time a human is involved.

He explicitly exempts people using cheaper models for cost reasons, noting only that a cheaper model can end up more expensive.

Sponsor

Greptile's T-Rex review agent, which spins up sandboxes and sub-agents to actually run and test changes, returning images and video as evidence rather than a bare approval.