hidden context + founder sparks

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ maybe ai shouldn't be too obedient?

🏃🏻‍♂️ almost every advanced techniques you suggested (cot, smt hybrid, kv caching… ) is about reducing internal error for (ai)... we are optimizing machines… the weakest link is human impulse, all of us routinely fail to externalize intent, we don't prompt things that are prompt worthy (missing prompts). Ai cannot obey what we never bother to surface. If ai doesn't need to be explicitly prompted, then they can assume using implicit intent. Instead of user pull, it is system pushed. This is why feed work at scale better than "search" (e.g. YT).

• This is what i mean when asking about ai obedience. Not obedient to the explicit prompt, but obedient to the hidden implicit intent that are more real.
• The question that makes me uncomfortable = "can and should we have AI that can force me to deal w/ reality, even when i don't know or want to?"
• 🏃🏻‍♂️ yeah, but they never deliver on that promise.

🏃🏻‍♂️ "messy middle" — Eli5, i kinda know what i want, and only know how to say it in a way that make sense for me. We do this all the time and it mostly works b/c we share hidden context. If anything, we omit the obvious/obvious b/c we see it as wasted context.

🏃🏻‍♂️ good point, let's reboot & use one of our previous chat about keto for fat loss + bball performance:

• "You should try keto b/c it helps w/ fat loss & can support your 2hr daily bball".

• ai can give advice like "keto is good for weight loss (zero sugar), but it will hurt high-intensity game play (<50g carbs/day) & conflicts w/ LDL" ← all of this is true but still useless.

• But we are both left with operational questions: should i do it or not, if yes, how strict, for how long? What do i track to know if it's working? What are the stop conditions for keto? ← there are so many hidden assumptions, and most ai will not "self-prompt" to work thru all the details… It is too expensive and requires so many layers of implicit & explicit.

• The ai fail b/c it doesn't know what level of bball you're playing. If the games are casual pickup, then the risk of glycogen depletion is a non-issue. ← no ai system today attempts to re-frame the prompt to ask this. (also, most users will find this annoying, we like to treat machines diff than people, no patience, forgiveness).

• Another problem is lack of connecting the assumptions w/ reality (grounding). Ai may warn about potential downside of ketosis (factually correct); but no ai system today will self-prompt like a good coach over a period of 3-4 weeks and check in with you to share what it feels during ketosis.

🏃🏻‍♂️ i'm sold on the capabilities of: SMT solvers, ground reasonings in proof, automated reasoning via verification. "If you can prove it, you can trust it". SMT = verification, BUT the messy middle = validation. There are limitations & new set of problems that i'm not sold yet.

• Messy middle: there is a gap between human intent (open ended, underspecified, latent/hidden context) vs. formalized solvers (explicit, finite, close ended). (e.g. i want to get more fit, which supplements are the best for me, what's the best diet?). LLMs can translate and reduce ambiguity.
• Context engineering + long-horizon reasoning (on top of solvers) are great to add nuance, facts, and self-enforced error corrections (e.g. verify facts w/ RAG) and enforce correctness thru rules/polices, but it depends heavily on the correct "framing".

🏃🏻‍♂️ Neuro-symbolic systems cannot detect unknowns or relevant variables that was never introduced (i have limited understanding on the implementation side, but my assumption is that solvers work in a constrained world, where variables pre-defined & logical are verified before run time.). Do you think more constraints → closer to reality?

• capable : 1). eliminate internal contradiction (LT2 threshold training while maintaining zone 2 heart rate) 2). Expose fuzzy assumptions (best magnesium for me) 3). Prevent out of bound states (no water for 10 days)
• incapable : 1). Decide relevant framing 2). Grounding user intent 3).validating correspondence/truth (high-level of mercury is bad).

🏃🏻‍♂️ amazon bedrock (case studies strava): verification by math: align content with rules, translate natural language into formal logic = grounding models reasoning w/ facts. Supplements for search driven content filters (widely used technique). 1). Rules need to be conditional relationships, no absolutes (e.g. if HRV > 50, then tempo run 10k) 2). Need detailed and comprehensive variable descriptor/dictionary (e.g. run easy pace = zone 2). 3). Target edge cases, create boundary conditions for type 1 errors (false positive) & type 2 errors (false negative).

• Conceptually, automated reasoning verifies the logical steps are followed for narrowly defined problem space (strava for a specific activity in the context of a specific cohort of athletes). BUT SMT is only as good as the pre-defined set of rules & descriptors. (note: amazon limits documents to 5mb/50k chars).

• Great for single domain (health, fitness, diet, legal, finance); not suitable for general purpose agents or cross-domain workflows.

• Known limitation, a potential show stopper: does not work with complex documents w/ nested conditions & inconclusive/local-context statements that may be contradictory. Longevity is filled w/ these types of data.

• Known limitation but a non-issue: it struggles w/ non-linear arithmetic & irrational numbers.

• Potential deal breaker, it is slow & expensive. Strava Athlete Intelligence made it work b/c they have 150 million users (large yet homogenous data) & 2 million paying customers ($80/yr). Activity data = numerical & high casual relationships.

• 🏃🏻‍♂️ all successful: are you framing these are all "successful" / "useful" by age? If so, define when did they cross the successful line vs. when did they start…

• 🏃🏻‍♂️ 100%, alt interpretation, for such a global platform, why it continues to under-deliver its economic value?

• 🏃🏻‍♂️ YT builds a massive, interconnected system of data points around each video to categorize and recommend it, rather than creating a comprehensive, internal knowledge repository for general querying.

• 🏃🏻‍♂️ overall, like the patterns you highlighted, take a noisy domain, vectorize it, personalize it, then serve to charge for its value.

Stephen / Basketball 🏀

🏀 ??

🏀 "proactive", per manus peak's podcast.

• 🏀 i read through your few blobs of text many times but cannot follow at all to comment. start over with a detailed example, state the problems you encounter with the current workflow, suggest how a human expert may solve it.. so i can understand your grounding? otherwise, it's just terminologies and abstractions that we can talking past each other.

• 🏀 agreed. see "impact %" below.

• 🏀 your "framing" and "level" is probably the same as my "hierarchy"; search below.

• 🏀 a proper, full system of symbolic provers can solve any logic / math (that humans can solve).

🏀 #youtube-agent what's the equivalent insight of insurance cashflow, basic interpreter, prebuilt homebrew, just-in-time assembly, game acceleration, fungible books, late returns, page rank, china icq, direct manufacturers, laptop batteries, harvard dorm?

• founder ages: brk.b 1965/35, msft 1975/20, aapl 1976/21, dell 1984/19, nvda 1993/30, amzn 1994/30, nflx 1997/37, goog 1998/25, tcehy 1998/27, baba 1999/35, tsla 2003/32, meta 2004/20.

• 🏀 by each founder's age at the time of founding; just saying.

• 🏀 but it's the spark of innovations for each (e.g. goog > page rank) that overthrew the 20 previous search engines and dorminated for decades. hence, my framework and leading question above – and (hand picked) gemini's answers below like "search: vectorized moments".

• feed: curriculum generation – The algorithm optimizes for addiction, not education. An agent optimizes for goals, stitching unrelated videos into a linear course.

• view: active interrogation – Consumption is passive; learning is active. Transform videos from "streams" into "chatbots" you can argue with or query.

• clone: asynchronous presence – Creator burnout is inevitable. A "Digital Twin" agent handles comments and community posts in the creator's exact distinct voice.

• search: vectorized moments – YouTube is a database, not a TV channel. Indexing every second allows users to "Ctrl+F" the world's knowledge base.

• 🏀 exactly! just as important, youtube serves insane traffic so our crawling of top 1% subtitles (10 terabytes) is unnoticeable – but we can create 10x value / product – per fair use law.

• edit: semantic timeline – Editing is a text-manipulation problem, not a pixel-cutting problem. Treat video frames like code that can be refactored by LLMs.

• 🏀 "data points" meaning machine learning for ranking + recommendations? yup.

• 🏀 sadly, not $199.

🏀 #ai-infra what use cases if we have the full / top 1% youtube video titles / stats / descriptions / thumbnails / transcripts / comments / keyframes?

• e.g. scout basketball talent via shooting angles; clip not summaries but 3 highlights of 1-minute segments; add to shopping cart of in-mentioned products.
• "can an ai agent on youtube be useful to you for _? like manus on 100m channels + videos – but understanding transcripts with tones + scanning scenes for objects + _."
• 115m youtube channels, 5b videos, 20m daily uploads. 36b revenue, 125m paid subs, 15b views on "baby shark dance", 3m paid creators. "me at the zoo" 2005/4/23! datasets: 8m segments, 170 mmsum, 45b words, 600 labels.
• pinterest: easier crawl, visual board, shopping agent. 600m users, 2b weekly pins, $5b rev, 70% women, 42% gen-z, 33% 100k salary, 55% purchase intent.