N8 Programs' first LessWrong post is a hand-written LLM-style CoT reasoning out today's date
akbirthko · x · 2026-10-01
N8 Programs published his first LessWrong post, "What's the date?"—an experimental piece written entirely by the author but styled as an LLM chain-of-thought, which he urges readers to close-read.
The text simulates a ChatGPT-style model's inner deliberation when a user asks for today's date with none provided in the system prompt:
- It weighs the dilemma that hallucinating a date would be penalized by a watcher, but abstaining would also cost it;
- It invokes the OpenAI spec's guidance to "express uncertainty" on questions beyond its knowledge, guessing whether the grading rubric rewards calibration;
- It then tries to infer which model version it might be—and thus today's date—by recalling GPT-6.3/6.5/6.6 release timelines and the 4-6 month pretraining-cutoff lag.
Underneath the CoT framing is a meditation on model self-identification, knowledge-cutoff calibration, and the tensions of hallucination-penalty regimes.
More from AGI Musings
- SeatGeek's company-wide ChatGPT rollout spawned 300 IT tickets on day one — aakashgupta · 2026-10-01
- AI Researchers Push Back: A PhD Is for Decades-Long Research, Not an Industry Launchpad — DimitrisPapail · 2026-10-01
- Ben Todd puts AI existential risk at ~2/3 on current trajectory, 10-35% all things considered — ben_j_todd · 2026-10-01
- p(doom) still too unreliable to inform AI policy, argue Narayanan and Kapoor — chalmermagne · 2026-10-01
- Researcher Hints OpenAI's Internal Model Solving 100s of Math Problems Will Accelerate Learning Theory — Afinetheorem · 2026-10-01
- Reply to LeCun: AI is an amplifier of owners' intentions, biases and incentives — AryHHAry · 2026-10-01