AI engineers discuss why examples lock models down and how few-shot lists fix it
On October 24, Netflix AI lead Chris Albon kicked off a discussion on prompting techniques on social media, drawing in Margot Mitchell, Mitchell Hashimoto, and others. The core question: why does giving an AI model examples make it stop thinking, and how to avoid that.
Confirmed
- Chris Albon described a key pain point in how he communicates with AI: he tends to illustrate intent with examples (e.g., "suggest dinner recipes for serious cyclists—think people who eat oats and blueberries"), but once the model sees an example it "locks its jaw," abandons generalization, and just copies the example.
- Mitchell Hashimoto explained the mechanism: when the input contains ingredient examples rather than recipes, the model can't find a token pattern matching "recipe," so it tends to continue the example itself instead of completing the requested task.
- Margot Mitchell offered an actionable fix: structure the examples as a list for the system to complete, rather than having the model imitate full samples—that is, one-shot / few-shot prompting; for instance, have the model first list ingredients a cyclist would eat, then generate recipes. Hashimoto proposed the same approach in the same thread.
- Chris Albon agreed that "this is indeed the right way" and quipped "the prompt engineer is dead, long live prompt engineering," emphasizing that structured prompting still works.
Why it matters
- The discussion directly connects a common prompting pitfall (examples derailing output) with an established technique (few-shot list-style completion) and adds a mechanism-level explanation—offering immediately practical guidance for developers working with large models daily.
2026-10-24 ~ 2026-10-24 · 6 related posts
Primary sources
- [source] Netflix's Chris Albon: LLMs lockjaw onto your examples and stop thinking — chrisalbon · 2026-10-24
- Mitchell Hashimoto explains why models get locked onto your examples instead of the task — mmitchell_ai · 2026-10-24
- [source] The fix: restructure your examples as a list to complete — one-shot / few-shot prompting — mmitchell_ai · 2026-10-24
- [source] Few-shot prompting trick: turn examples into a list the system completes — mmitchell_ai · 2026-10-24
- Prompt engineering is 'dead', long live prompt engineering: a cyclist-recipe prompt shows the craft persists — chrisalbon · 2026-10-24
- Chris Albon endorses the list-completion few-shot prompting approach — chrisalbon · 2026-10-24