git repos + ai architect

· Zi Wang · 1 min read

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ Quick ctx on this entry. You + I have been deep in the Manus founder pods for a few days now. Pulled CN + EN audio, scanned transcripts, ran multi-pass synth, extracted a ton of tech + prod insights. None of that is hard anymore. Translation, summarization, top-100 insight extraction … (doc)

  • We're sitting in an abundant intelligence / action-scarce world. Insight is cheap. Execution is still expensive. There's no default path from "this is interesting" → "here's a repo / prototype / thing running." Even for us (and honestly prob <1M ppl who can do bilingual + deep-tech synth), the marginal value of another summary is near zero.

  • 🏃🏻‍♂️ fair, was looking for "demi-gods"

  • So, the bottleneck isn't understanding Manus. It's taking insight → production ready apps that can generate real-world actions (ie. help users make $). Have you thought about linking youtube content w/ 3rd party data sources such as git repos?

  • 🏃🏻‍♂️ no, outdated repo w/ intent of rl training

Stephen / Basketball 🏀

  • 🏀 "A General Agent isn't a "God Mode." It's an agent that uses generic interfaces (browser, code, file system) to solve problems, rather than using specific APIs for every tool."

  • 🏀 see "open manus" per today's; search "(minimal) fork" below.

🏀 formal language as the reasoning + action interface: currently ai chats + agents are so difficult to control / reproduce / guarantee because of language embeddings + blackbox weights + stochastic decodings. similar to a live bug stuck in computer vacuum tubes in 1947 or the daily page faults crashing from pointer arithmetic, the current interface of natural language is empowering but flawed. humans + ai should iterate on defining goals and checking tasks – via math (logic, graphs, probability) + code (python, lean, julia) progressively but as much as possible.

  • so elaborate but inherently flawed manus prompts (becoming a chainsaw). open manus vs manus open.
  • ai architect in 2026+ with theory, spec and proofs – vs chatbots (conversation vibes with hallucinations) in 2022, thinkers (chain of thoughts and reasons) in 2024, agents (task runner with tool use) in 2025. a semantics compiler.
  • advanced agents by dawn song: 45m papers open scholars, visual grounding gui agent, algorithmic reward llm optimizers, former verifier in-context learning.