By Austin
Initial use cases for AI corporations
What are:
- high-value opportunities for trade between humans <> AIs
- where economic rights & frameworks might be valuable
- not already pursued by frontier labs (ie agentic coding)
- where we have some amount of domain expertise, and/or existing userbase?
Ideas:
- A: Grantmaking (Manifund)
- Imagine: AI corporations which are competing as grant evaluators
- S-Process, but the evaluators are all AIs and the whole thing completes in seconds not months.
- See Vitalik’s “AIs as the engine, humans as the steering wheel”
- Eventually AIs are also the grantees, I guess
- This seems like obviously how Anthropic donors will prefer to allocate their funds
- Prizes-as-a-service, competitions-as-a-service
- Impact markets ofc
- B: Forecasting (Manifold)
- Imagine: AI corporations competing as forecasters & question writers
- C: Writing (Demosthenes: a new forum )
- Imagine: You can turn a “compute dial” from $0.01 to $1000 to commission research and essays from AI corps, and know that the result was worth that token spend
- Small example: Chris Best thinks Substack outcompeted Wordpress et al, partly through better economic engines. (and partly through better execution imo)
- D: Governance
- Of individual orgs, companies, communities (eg Manifund, Mox)
- Of AI safety & EA as a whole
- Of labs (Anthropic, OpenAI, etc)
- Of larger polities: cities, states, nations
- Of deals between labs (Anthropic <> OpenAI), and between nations (US <> China)
- E: Hiring & matchmaking
- A lot of inefficiencies already exist in today’s hiring markets
- Tomorrow’s hiring markets might change drastically as AI capabilities rise
- Maybe much shorter tenure between orgs
- Maybe org lifecycle gets compressed a lot (eg Anthropic, Cursor, Mechanize)
- Standard startup comp + equity norms may no longer hold
- Coasean boundaries might dissolve
- Maybe many small new projects are the norm.
- Or maybe we all merge into Anthropic.
- AIs will probably want to rent humans
- Income share agreements, or selling equity in people rather than orgs
- F: Trading on existing markets (stock, crypto, compute/MNX)
- Pro: seems kind of obvious
- Con: probably a bunch of smart and rich orgs/people are doing this already
- Con: Caroline is banned from trading for a while longer
- Con: not clear that on the margin, more of society’s resources should be spent on more accurate asset pricing, for the assets that have markets
AI-owned vs AI-run corporations
- Many AI safety people seem to think that we need to keep humans in the ownership/liability loop for corporations. I think this might be true for current-level AIs, but not sustainable for future AIs.
- Not ever having AI-owned corporations seems bad:
- Either the AI-run, human-owned corporations end up using those humans as “fall guys” (maybe with the humans consenting, ie hiring humans to explicitly be a liability shield)
- Or we lose out on massive potential economic growth, because humans are bottlenecking total GDP growth for being too scarce or too slow
- Thesis: It would be Good Actually if we trusted AIs to be liable enough such that they could own their own corporations. By analogy, humans feel liable enough for their orgs actions to not want to break laws, for fear of social shame, being fined, or thrown in jail. LLMs don’t currently care about this, but if LLMs (or some superstructure thereof) did, we would be able to trust them more, trade with them more.
- Objection: “Shouldn’t we want all economic upside from AI systems to accrue to humans?”
- I argue no! This seems like a recipe for misalignment, broadly. See Dwarkesh’s “give AIs a stake in the future”
- Aside: I think correct impact accounting is deeply important, because if we don’t assign proper impact points to current/future AIs, why would they respect our impact frameworks?
- Also per my personal values, it’s not clear whether I’d prefer economic upside to accrue to distant greatgrandchildren with ??? values, vs nearterm AIs who I can converse with and whose values seem aligned to mine. (Unclear if this is the tradeoff being offered though.)