Claude, On Average
▶ The full series playlist — Brutalist https://www.youtube.com/playlist?list=PLG9H-C6rp5RU
Claude, On Average
▶ The full series playlist — Brutalist https://www.youtube.com/playlist?list=PLG9H-C6rp5RU
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
[0:00] Every time I ask Claude the same question, I can get a different answer. Sometimes slightly different, sometimes dramatically different. Same prompt, different draw. That's not a bug. This episode, what that actually means, and why everything else in this series is a consequence of it. Three runs, same prompt, three answers, all plausible, none identical. A database returns one row. Claude returns a draw from a distribution. These are not the same operation. The alternative, treating Claude like a lookup, breaks immediately. You ask a question, you get an answer, you run it again, different answer. You can't cache the right one. There is no right one. There's only the distribution.
[0:46] What it actually does, next token prediction. At every step, it estimates the most likely continuation of everything before it, the prompt, the system message, everything it has already generated. It's not retrieving a stored answer. It's generating, token by token, conditioned on all of that. Temperature zero makes the most probable token always win, but probable is conditioned on everything in the context. Add a different persona and you shift what's most probable, even at temperature zero. Steering the distribution is not the same as specifying an output. Frozen is not pinned. This is also why it's useful. The variance is the creativity. You want different takes on the same brief. You
[1:31] Want the unexpected riff. The tails of the distribution are where the interesting answers live. You're not sampling noise. You're sampling possibility. And this is what it costs you. One draw is not the answer. One draw is a sample. If exactness matters, if two runs must agree, you don't prompt. You script. You run the gate. Never certify a single draw as truth. The distribution rules, same prompt, different draw, that's the definition. Temperature zero steers toward likely. It doesn't specify an exact answer. Variance is the creativity. Certify it with a gate, not with faith. Scripts pin the exact. Skills bias the draw. Gates catch the bad draws. The whole system is
[2:19] Consequence of one fact, Claude is a distribution. If this reframes how you think about prompting, subscribe. Nick Bear Brown.
