fine tune + deductive ai

· Zi Wang · 1 min read

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ debating/validating aaron's suggestion that fine-tuning is NOT needed (his suggestion of make the system prompt better + add RAG maybe enough to get the job done). Pros & personal experience: using a round-table of auto-generated system prompts (w/ GAN like approach) has worked well for most tasks & gemini's file search api does help. Simple trick like "assume the role of a longevity specialist, Peter Attia, use his tone and research to analyze my fx-health report" does get the style and voice right.

  • Cons, w/o changing the model's weights, i get the default and general behavior of the base model. Plus, I still get the random data "drift" when pulling files from drive or get "ghost" data even for the attached file in the prompt. Seems like compute cost for off the shelves fine-tuning is reasonable (fine‑tune big models on a single modern GPU has been possible since 2021).
  • My tl;dr as of today is that if the goal is to analyze n=1, then system prompt + rag is perfect; but if there are complex, multi-step instructions are needed, then i don't see how to achieve consistent & domain-specific reasoning patterns.
  • (🏃🏻‍♂️ note to self" make this a topic w/ Chung-Ching & Eric next wed).

Stephen / Basketball 🏀

  • 🏀 your 3-way debate setup is a good improvement, but very incomplete and lots of biases. see "dr. crusher" below.
  • 🏀 that's exactly my and aaron's second point: even with 10m context, the precision and recall are not uniform / perfect. the right approach is to use a hierarchical deduction, per lean prover below. for short term, he and i meant: iterate to verify small steps, piece by piece, to allow attention to each detail.
  • 🏀 let's push gemini to the limit. what "complex, multi-step" questions do you have now?
  • 🏀 ask about their 2026 (health) goals!

🏀 #deductive-ai like deepmind's alphaproof, generate hypothesis and deduce conclusions in logical framework lean (nature 2025, berkeley talk) – which gives consistent advice without hallucinations and provides hierarchical structures of reasoning. extend lean for probabilistic modeling. see medgemma for open-source fine-tuning. to solve these paradoxes (conflicting evidences / unresolvable choices / inconsistent theories / confusing advice) on biomarkers / diet supplement / training.. – and make definitive decisions:

  • zi: age 43.6, 170 cm, 125 pounds, 2-hour daily running – but: sugar crave, mood swing, 6-hour sleep, cortisol 8.1 mcg/dL.
  • stephen: .. hepatitis b carrier and a kidney removed – but: ..
  • aaron: 16% body fat but swollen face..
  • shamir: 80% muscle but high cortisol..
  • 🏀 many ai + prover: alphaproof, alphageometry, harmonic's aristotle, deepseek-prover-v2, meta's hypertree proof search (htps), goedel-prover, internlm-math-plus. some even claims neuro-symbolic ai in healthcare.

🏀 #on-fire now is time for new year resolutions. collect from all friends of their (health) goals. make them strongly committed, easily actionable, fully measurable.. in 3 months, not the whole 2026 year.

  • tracy: in extreme back pain since a car accident 3 years ago.
  • quoc: lower cholesterol, strength lower back, lose 15 pounds.
  • kai: [feel and look younger].