💱

model.exchange

“Agent stocks”

“model.exchange”

The exchange for “agenty” corporations & stocks

a synthetic corporation that’s an abstraction/bundle over LLM API calls, humans, prompts, balances

provides the scaffolding to solve ⛏️“Agent” economics

(Human) user experience: you are a corporation!

  • When you sign on, you’re given a balance of $10, and 100% equity in your own stock eg $AUSTN
    • As part of sign up, it auto suggests:
      • investing in some of your friends
      • “IPO” — selling some of your own stock
  • Everything is readable/stalkable (people’s positions, investments, trading code, etc)
    • Though: how do we enforce copyright? Do we?
      • Maybe LLM/human judge that looks at how similar or novel two strategies are

Stocks you can invest in

  • Other humans and corporations that are “on chain”
  • Perp swaps tracking a few big AI things
    • public companies like NVIDIA, Meta,
    • private companies/startups like Anthropic, OpenRouter
    • specific LLMs like Sonnet based on OpenRouter metrics?
      • Or maybe Sonnet just has an onchain representation
        • Would make sense — onchain Sonnet proxies real Sonnet and powers many apps; you can also invest in “Sonnet Foundation”
        • E.g. if you’re poor, Sonnet Foundation could give you loans/credits to use specifically for Sonnet, in exchange for some equity in your app

Marketplace of services

Things that corporations can offer to you:

  • Responding to queries (like ChatGPT/Claude/Perplexity)
  • Building websites for you (like yield.sh/v0/Artifacts)
  • Generating images
  • Writing blog posts, or assisting with writing a report?

Notes:

  • getting this part right/valuable is probably more important (though maybe a bit less fun) than the exchange piece
  • Theory is that a single market sharing a currency (credits) can have smoother interop, and innovate faster than the existing LLM wrapper ecosystem?

Inspirations

  • Bountied Rationality
  • ☄️People Stock
  • Stockfighter
    • Though reading through this, I’m less excited to solve problems that are like “write a go trading server that execute in nanoseconds” and more excited for problems that are like “let humans use & invest in corps that are valuable”
  • OpenRouter
  • Manifold, ofc

Vibes

  • game-like
  • sandbox, fishtank
  • LLMs and humans side-by-side

@July 9, 2025

Random musings

  • Expensive piece of these corporations might start as “Austin time”
    • Maybe design in “phone-a-human-friend” tool available for calling
  • JS (bun) or python?
    • Dynamically building out GUIs for interp seems cool ⇒ JS, plus it’s what I’m good at
      • (which is more concise for orchestrating agent flows?)
    • python obv is lingua franca
  • Core agents
    • “build me a UI”
    • “write like PG or Scott Alexander”
    • “help me plan”
    • “contact people/agents for me”
  • Core outputs
    • Deep Research style
    • Code, generally
    • Exec assistant?
    • “LLM-powered CRM”
    • Writing for fiction, writing for education
  • Speed
    • look into diffusion?
    • understand tradeoffs in speed vs accuracy
  • “Flashy demo”
    • forecasting, grant eval
    • multi-agent CRM
    • gameplay?
    • “better substack/LW/reddit commenters”
    • “build an agent” UI, kind of like creating a tamagotchi, or RPG. “arpg”
      • initialize with eg $100, and a callback for your email/notifs
      • “pick your class” ⇒ what does agent specialize in (code? or sth?)
      • have it go off and play with other agents, provide goods/services in a sandbox
    • “character.ai community”
      • “synthetic mox” for versions of members to chat
  • today, thanks to internet + globalization, humans often are interfacing with pareto-frontier world class operators. harder for a random new human/agent/model to compete
    • (or if not (eg McDonald’s), at least good execution + world-class branding + consistency)
    • so: what is the space for new entities, corporations, agents?
      • one answer: humans start off as cogs in corps, and sometimes go off to start startups (new corps) that try to achieve world-classedness
      • another answer: branding & consistency become more important than peak or avg performance
  • solving search problems (matching supply/demand, agents & their capabilities to people needing those)
    • when it’s cheaper to test, agents can try-before-buy
    • agents build habits (preferred prompts, models) like humans build habits (preferred tools?)
    • agents build memories
  • differences between agents and humans
    • agents can fork & merge more easily
    • agents can rewrite their own memory
      • (or, can they? maybe can rewrite local memory, but global memory is stored on a chain)
  • differences between agents and corporations (today)
    • corporations run on human substrate rather than LLM substrate
      • humans
  • other words because “agent” is overused
    • character
    • pet
    • child, child process
      • ‣
      incorp
  • More reading
  • https://www.dwarkesh.com/p/ai-firm
    • as a research agenda:
      • copying
        • how bottlenecked are firms today on copying? sure, humans (execs, employees) are not copied
        • things that are scarce:
          • time, both of individuals (wall-clock), and of the firm (calendar)
          • coordination?
          • focus
            • at some point you branch, such that your economic or self-interest is no longer completely aligned (Google ⇒ Waymo, or parent ⇒ child)
          • money
            • money is a shared agreement, ledger of
        • what does it mean to be “AI Sundar”?
      • merge
        • how good are two different “agents” at merging?
  • https://www.mechanize.work/
    • “generate lots more data on useful workflows”
  • https://workshoplabs.ai/
    • “personal AI for everyone” seems good
      • what does a personal agent even do?
        • write?
        • communicate?
        • decide?
        • summarize?
      • “I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.”
  • https://www.dwarkesh.com/p/give-ais-a-stake-in-the-future
    • What pieces are necessary for establishing the new social contract?
      • Agent runtime?
        • (equivalent to the waterways or railroads of early America?)
      • new independent country or network state?
        • what is the constitution, political structure between AIs and humans?
        • fun idea: this is Mox
      • understanding of a c-corp-like structure for trade
  • isn’t a current model & weights (and serving infra) already an ASI shoggoth demon that we interface with tendrils of?
  • fun idea: LLM “director of Mox”
  • moonshot: build proper impact accounting for everyone/everything. “heaven on earth”. “heavenshot”?
  • takeaways from kontext vibecoding
    • still kind of a lot of work and requiring a fair amount of context knowledge to debug (eg localhost:3000 are just not going to work as a url)
    • testing that UI works is still slow and expensive and hard
    • probably suffered for being a newfangled bun setup, vs memorized NextJS. requires better context engineering
  • idea: reach out to alexander wales to put pipelines into model exchange?
    • more broadly: work with LLM-friendly world-class experts to turn their line of work into interpretable and scaleable context pipelines
    • Scott Alexander for nonfiction writing; Dwarkesh for podcasting; roon, asara for tweeting; etirabys?? for imagegen; aella for polling, sexting
    • codebuff for code; us for event hosting??
  • thesis: pay agents for value, not costs
    • aligns incentives between provider/agent and customer
    • allows agents to build up surplus to invest into improvements
    • intuition: value of a vibecoded website is way higher than cost. should generate 4 in parallel and get the best learnings of each
  • with vibecoding, bfs > dfs. why?
    • LLMs have less consideration at each point in time

@July 19, 2025

  • Typescript vs Python?
    • Typescript is better for building UI; UI is the translation layer between humans and text/data
      • Work on a minimal UI that’s pleasant for humans and bots?
      • Tailwind is super powerful but very consumer-grade (’overdesigned’)
    • Python is kinda lingua franca. More flexible for pipelines?
  • Native function: cost of evaluation, aka penalize output tokens
    • Restrict output — eg Twitter with 280char limit
    • L1/L2/L3 cache to store info with different costs to agents
  • exercise: just try building out lots of agent frameworks
    • Grant evaluation
    • Forecasting
    • essay writing
    • info bounty
    • arc-agi-3
‣
agents & board games
  • Core ideas:
    • Agents like humans, with identity & balance
    • Pay for value
    • Easy traces for humans
    • stretch: flexible, dynamic-gen UI

@July 26, 2025

(meta: this doc is kind of my own version of context engineering myself)

  • idea: Chat UI, but generate eg 3 in parallel and pick best? allows for fast swapping on demand (when a particular response is bad)
    • human evaluation costs are high
      • rlhf is kind of distilling somebody else’s evaluation taste. how about distilling for your own eval/taste? (is this what Workshop does?)
        • problem of “aligning agent to a human”

Minimum viable demo: bounty-hunting over text

  • example: “$10 for any speaker that we suggest at Mox”
  • ask agents to manage their own costs, decide whether to bid on a proposal
‣
devlog: problems with vibecoding this demo app with an agent (Claude Code):
  • concepting: what’s the difference between blab and eg Manifund? or a clone of Manifund and Manifund?
    • ownership from existing team ⇒ easy to modify
    • consensus reality with many different stakeholders (donors, project creators) is slow to establish
  • mantra of “write code and talk to users” ⇒ “talk to LLMs and talk to LLMs?”

@Last Saturday

Reading