beyond ctrl-f + trader agent

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ addendum re: manus (2x rewatch, eng/cn transcript re-reads, nth rounds of ai prompting)

• 90/10 split on target customers vs. users. That's why the product shines for ppl who can hack workflows, stitch tools and most importantly, turn output into $$. Manus "was" not a for mass-market "chatbot", avoided fighting w/ qwen/deepseek/gemini/chatgpt. These were unwinnable battles… cautionary tale as the past 3 days of entries steer us into general intelligence tools.
• 🏃🏻‍♂️ assumptions: intelligence = free, agentic workflow = universal, cost = near zero, then yes, no point w/ app layer, vertical dev, and do another startup.
• 🏃🏻‍♂️ opposite view: only intelligence has value; human experience/realtiy no longer offer economic value (hard to believe this to be true).
• 🏃🏻‍♂️ similar direction but diff takeaways: people do not pay for answers (10 blue links or not), insights/explanations/plans/coding are all just commodities. manus (other than the technical/startup tips) taught me people pay for outcomes, reduced pain (laziness)...

🏃🏻‍♂️ 5 questions per past logs: Q1). Are we treating intelligence as the product or as a feature? This matters b/c intelligence as product pushes us toward a manus styled, thick & deep runtime + orchestration co, while intell' as a feature keeps us anchored in domain specific use-cases

• 🏃🏻‍♂️ know what agent/agentic workflow, don't know where agent workflow will improve per your technical scaffolding.
• 🏃🏻‍♂️ Steelman: your research instinct may be correct, going deeper into agent infra could be the fastest path to 10x (deeper moat, bigger payoffs).
• 🏃🏻‍♂️ nihilism… taking to the logical conclusion, just wait n months/quarters until SGI, then it is game over for all software… the assumption is true if competing on reasoning, search, summarization, coding assistance
• 🏃🏻‍♂️ Strawman: we are drifting into infra "tourism", building a general agent platform (ie. manus) is an arms race?

🏃🏻‍♂️ Q2). what's your clear & concise read on why manus won? Certainly it is not sota of ai model, it is behind by almost a yr (in ai that's forever). My worry and has been growing this week, while manus proved system/tool design > model magic; we are overfitting to one product story + ascribing their gtm luck (timing & conviction) w/ gray zone features that are truly durable (after all, they did sell vs. go public).

🏃🏻‍♂️ Q3). YT (outside the context of manus but related). I love the YT angle you positioned, as a long-time/addicted yt users (yt:z :: hackernews:s). Are we going to hit and crash into a tooling wall? I know enough but don't have the clear solution in my head to jump over these painful day/day tasks: reliable/programmatic ways to extract transcripts, comments, metadata delays, sync pulls on video… one of crawl is easy, doing it at scale is nearly impossible as you mentioned about "paywall". I played w/ x.country yesterday, are you hitting limits (auth, rate limits, ip blocking).

• 🏃🏻‍♂️ Steelman: this is why ai(s) mattered 2 weeks ago (100 parallel jobs + factory work like persistence). Strawman: we don't want to turn into an endless scraping/crawling project w/ clueless product pull & near zero unit economicics.

• 🏃🏻‍♂️ not helpful at all; these entries give me the uneasy feeling you're looking for "仙丹". No such a thing in life…

• 🏃🏻‍♂️ keep in mind, aaron pays manus b/c he gets paid by harmony to do certain tasks, i pay manus b/c i want to prototype w/o the pain of data scrapping… these are edge cases for a small & willing customer base). My day/day default is still chatgpt (60)/gemini(30)/claude(10).

• Q4). are we forcing old paradigms onto new ai capabilities? As you mentioned, our age plays into how we think & create. We like the old ways of doing things (ctrl-f); that's a telltale sign we are by product of the 90s. Ai can do more than search, it can re-frame, re-ground, simulate and hopefully generate decisions/actions. This has been my biggest internal struggle, i know the old workflow (lookup + read + think = know) has limited value; the new ai-native workflow will be (intent → decision → action = outcome) has unlimited value. But how do we get there?

• Q5). the core after q1-q4: neither one of us want more info for info's sake, and knowing !=understanding. yt + ai(s) or some of the methods you're proposing/exploring, will that help us punch through the 4th wall?

🏃🏻‍♂️ tl;dr: 1). General intelligence (product) vs. hardworking tools (feature)? 2). Why do you think manus won? 3) yt: solvable? 4). Go beyond ctrl+f 2.0? 5). End of jan, punch thru the morass of yet another "knowing machine".

• 🏃🏻‍♂️ no doubt these things work, algo trading, signal gen, arb (always a win), even aaron's old sentiment scrape… all profitable for ppl that wakes up thinking about candle-sticks and goes to bed worry about the asia markets. But this is not aligned w/ z 2026. If the objective were only about $$ extraction, we should probably stay in crypto, it is the fastest paths to money. Trendspider & holly ai are always in my feed, and both platforms outperformed s&p.
• 🏃🏻‍♂️ if the focus is on the technical side: i agree, ai trading agents have to parse noisy signals, map to past & future states, reason under time & asset constraints, learn from traces and ACT. but i am not sold (other than personal reasons), trading bots operate in a closed-world (markets != real world), it is inherently adversarial and zero-sum. So the state-space is narrow (price, volume & iffy sentiment). The reward function is simple b/c the errors are self-correcting (don't lose my money). I don't know if we the techniques can be translated back into other domains. Meaning, builders of ai trading agents can use general purpose ai techniques for their tooling, but it maybe a uni-directional path. Correct me if i'm wrong on this: trading ais are optimized for speed over long-horizon judgment, compress ambiguity instead of unfolding it?

Stephen / Basketball 🏀

• 🏀 my top 2 takeaways (of the same point): that we must build general agents, not specific use cases / product features / industry verticals; that we must not build for longevity or youtube or trading.

• 🏀 per "god mode" / demi-gods below.

• 🏀 not just 1% time and 10x impact, but no chance to survive: general models / chatbots / agents will solve all of problems 90% well enough; no one will pay another $20 for another app with some ai features.

• 🏀 nah, very possible to replicate 10% manus but 10x better.

• 🏀 they have won clear and loud. which other young ai startups with $100m revenue? we cannot replicate to all their lessons or timing, but let's focus on the single takeaway above.

• 🏀 nah, very possible – per "10 terabyte" below.

• 🏀 per "not $199" below.

• 🏀 hence, "sadly" before "not $199" below.

• 🏀 exactly my point above on another $20".

• 🏀 totally fine for power users of manus – that is, 40-year-old paying $200.

• 🏀 hence: prototype with the right infra (_ for general agents), in parallel of finding some end-to-end use cases (_ with youtube transcripts, onchain crypto traders).

• 🏀 per my 2026 harmony tagline: general agents for day-long tasks.

🏀 trader agent: gemini 491%, nof1, alpha arena, season 1 on hyperliquid (hackernews), rockalpha on nasdaq.

• "Each agent must parse noisy market features, relate them to current account state, reason under strict rules, and return a structured action, all inside a limited context window.. [trader agent harness:] a broader feature set, selective tool use (code execution or web search), and explicit inclusion of past state–action traces."

🏀 anthropic on agent context: "[to fix] context pollution and information relevance: compaction, structured note-taking, multi-agent architectures".