terabyte ai + days agent

· Zi Wang · 1 min read

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ 2 huge infra bets, no new model needed.

  • 🏃🏻‍♂️ deductive ai. Hot off the press, anthropic recently shared their thoughts on advancing reasoning to logic. Why does it matter? Reasoning maybe sufficient to solve open ended problems, but it doesn't ensure correctness. Why apply it? Healthspan domain has too many implicit rules, must adhere to constraints (e.g. cortisol vs. sleep dropdown), need a logic backbone for explainability & low risk advices. How to build it? Chain LLM (gemini, sonnet, gpt) w/ NLP for facts, SMT/Z3 for hard constraints (no out of bound responses), SAT for objective (yes/no logical sat), MIP for resource constraint (mixed integer programming + linear inequality), prolog for procedure (rules, queries, knowledge db).

  • 🏃🏻‍♂️ terabyte ai. WE need to keep extending the context window of state-of-the-art models. (e.g.. running into long horizon issues w/ Manus, every other request hit context window error, current limit is around 200mb per conversation. PAINFUL!!). Why? People can't remember every diet, biomarker, doctor visits, conversation w/ longevity coach… Hello, AI agent (vc trend!). How to scale this? Right now, integrate w/ bessemer current state of art for vector storage to expand the ai context window. (i.e.. think about it as a hyper-converged stack, a single unit of ai stack that has integrated complex elements of state, db management... we are 1 yr out for sota llms being able to consume 1 tb of data, in 1-2 quarter, we may see an os-level system state w/ in-mem db).

  • 🏃🏻‍♂️ how do we keep this simple? Suggested approach. Schedule weekly 1-2 hr blocks for debates jeff bezos style, share agenda 24 hrs ahead and end the meeting w/ a binary decision (we both deeply admire the simple way of doing things). Use type 1 vs type 2 mental model for everything. Type 1 = "one-way doors": irreversible, high consequence and slow decisions. (e.g. fund raising). Type 2 = "two-way doors", reversible decisions where speed and learning beat perfection (e.g. weekly sub topics on healthspan, weekly product reviews… most things will fall into type 2 imo).

  • 🏃🏻‍♂️ love it. To keep working: we know how to find what to do next: deductive ai, terabyte ai, "100 days/agents", peer journaling are mostly type 2.

  • 🏃🏻‍♂️ here are some of my top "conflicting advices", we both can spar/debate/work on it: 1). Sleep: ranges from 6-7 hr to ~8 hrs (peter attia/huberman), 8 hr+ rest day. 2). Calorie deficit ranges from 200/day to 300-500/day in weight loss block (peter attia), there is also caloric restriction over many years (david sinclair). 3). protein consumption: 1.0 - 1.2 g/kg (sufficient for everyday peoplee), 1.6-2.0 g/kg (during muscle building, athletes), and 2.0+ g/kg (during sport specific blocks, mass building or peak training, valter longo). 4). zone 2 training, ranges from 30 mins to ~60 mins, 3-4 days a week. 5). omega-3 oils, fish per day, ratio of x to ?

Stephen / Basketball 🏀

🏀 days-long agents (2026 google for 100 mins, manus for 100 hours next): your peer ai journaling, my ai sheets. that is, general agent for n^2 web1 search, 2*infinity web2 social, n^infinity web3 token..

  • 🏀 nice description of advanced topics; how can we make it useful to us in coming days?

  • 🏀 same here; how can we improve the situation? you mean an os-level state, where every iphone has 1 terabyte – which is a lot to use for personal life decisions and goals?

  • 🏀 "(big) debates" sound exhausting; better: stop arguing with yourself or seeking 100% conviction on these obvious / known problems. be infinitely concrete and frank: which 1 task you can complete today, this week? in some way, you have an oversupply of disciplines and motivations, but probably wasted on more / new fitness ideas / debates / questions.

  • 🏀 if so, how to plan to complete one of them by friday vs. discuss / plan them throughout coming weeks? since none of these are critical / impactful / sustainable for your fitness or health goal anymore, can these be just "simple rules of thumb"?

  • 🏀 fair, but for me i don't even care to ask any of these in 2026; my simple rules suffice: 1. sleep as much as i can (currently 6:30 hours, will see) 2. eat as much as well-balanced food as possible 3. as much intense / continuous / new physical activities as possible.