Weekly Links: LLM user panels, math proofs as tree search, and Stripe's eval benchmark
njyx · x · 2026-11-22
This week's Steampunk AI link roundup covers:
- LLMs for user research: asking LLMs directly for product opinions yields middling averages; prompting personas, collecting unfiltered reactions and aggregating them works far better — useful for surfacing spikey reactions and your most valuable fans.
- Stripe's eval benchmark: Stripe released an eval for how well models integrate with its own APIs, effectively defining the hill models/agents should climb — a pattern that could spread to many high-profile APIs.
- Lean & Mathlib communities certify the Claude-assisted Riemann proof: the 17M-line Lean proof announced by Anthropic's Levent Alpöge is gaining credibility, potentially the next Millennium Prize problem to fall to an AI+human team; the proof-as-tree-search mechanics are the more fascinating part.
Also noted: an Ironman completed in GTA6, Anthropic's IPO, and possible changes to America's AI safety rules.
More from Companies & People
- Matt Turck: AI ads have taken over NYC, making it look like SF — mattturck · 2026-11-22
- Anthropic now has more hundred-millionaires than all of Los Angeles — peterwildeford · 2026-11-22
- Nat Eliason shares insider look at Alpha School's incentive system: 5-year-old buys sewing kit with earned credits — nateliason · 2026-11-22
- From 400 applications to 1 hire: why the pick came from the reject pile — brandongalang · 2026-11-22
- Yacine: Neon Is Literally Better Than Anything Anthropic Has — yacineMTB · 2026-11-22
- Not just a harness: startup clarifies it's a Meta-Solver Platform for hard problems — Xianbao_QIAN · 2026-11-22