← @bgSponge

How AI Could Concentrate the World’s Labor in a Few Companies - Dylan Patel

Summary

  1. AI labor populations scale exponentially.

    The effective AI population at frontier labs increases 10x year over year as compute frontiers expand and capability costs drop.

  2. Economies of scale mandate centralization.

    Training costs for specific skills are amortized across billions of users, and leaders can charge higher markups by better economizing scarce compute.

    training an AI for a specific skill or specific set of knowledge gets amortized across billions of sessions

  3. Recursive self-improvement accelerates concentration.

    If Recursive Self-Improvement (RSI) occurs, the most effective users of compute will naturally capture more resources, potentially leading to a single-company world.

    if you believe in RSI you believe in the labs are the most effective user of compute

  4. Internal utility outcompetes external markets.

    Labs will prioritize internal R&D over selling compute to third parties because the returns to labor inside AI labs are much higher.

    returns to labor inside AI AI labs is much higher than the returns returns outside

    what if Anthropic can just generate hundreds of millions of dollars per megawatt by using that compute internally?

  5. Decentralization requires slowing progress.

    Patel suggests the only ways to avoid super-concentration are through government regulation or a general slowdown in AI progress.

Highlights

  • I don't trust the government and I don't - [04:57] trust Dario and I don't trust Sam.

  • I don't want to send the tokens - [08:01] outside. They're more valuable inside.

  • every force - [03:11] is screeching towards centralization and - [03:13] that's scary as hell.

  • each company individually has more - [01:07] labor equivalents than there are people - [01:10] on Earth.

  • So the computer the frontier - [00:30] the basically effective AI population - [00:32] size - [00:33] at the frontier lab is increasing 10 x - [00:35] year over year.

  • lab - [03:36] AI training has huge economies of scale - [03:38] because any effort you spend into - [03:40] training an AI for a specific skill or - [03:42] specific set of

  • effective AI population - [00:32] size - [00:33] at the frontier lab is increasing 10 x - [00:35] year over year

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →