The Economics of Recursive Self-Improvement [pdf]

(elasticity.institute)

91 points | by apsec112 7 hours ago

8 comments

  • zuzuen_1 5 hours ago
    "We make a tentative calibration of the self-sustaining ac celeration condition using the existing data that is available, measuring AI capabilities using the Epoch Capabilities Index (Ho et al., 2025).

    We find that the condition is met if a one-unit increase in AI model capabilities results in at least 15% higher AI R&D pro ductivity.

    A rough back-of-the-envelope calculation based on reported AI engineer uplift suggests this return has been around 9% since the launch of coding agents.

    This number is below the model-implied threshold, suggesting we are not experiencing a self-sustaining acceleration."

    And the source of this data seems to be self-reported productivity gains from surveys: 1.4–2X in METR’s survey of technical workers (Becker, 2026).

    A bit flimsy basis but an interesting paper nonetheless.

  • godelski 3 hours ago

      > But our models make it clear that such an [intelligence] explosion may not follow if there are diminishing returns (“ideas become harder to find”) or if feedback loops become bottlenecked.
    
    How is this not obvious to everyone? As we advance it becomes more difficult to advance. You obviously make most advancements around the things that are easiest to improve. Then all the easy things are done. So you go onto the next easiest things. They're "the easy things" from that standpoint but that doesn't mean they aren't harder than "the easy things" when you started. Complexity increases as precision increases.
    • setgree 1 hour ago
      The goal of a modeling exercise like this, which you don’t have to buy, is to generate a simple set of initial conditions that can explain things we already know. Then, we can manipulate some initial parameter value to make predictions about things we don’t see, but might.

      Likewise, it is obvious that gravity exists, but a simple model that explains where it comes from (in quantum terms) would be a big breakthrough iff it came with plausibly testable implications that could be tested via experiment.

    • tossandthrow 1 hour ago
      > As we advance it becomes more difficult to advance

      I don't think this follows.

      We have advanced tremendously over the past 200 years, and we are likely going into a time with rapid advancement again.

      With advancement, we also develop tools (eg. Llms) that assist advancing.

      • Windchaser 34 minutes ago
        To put it another way, the cost per advancement increases. That’s “as we advance it becomes more difficult to advance”. However, because of the prior advances, you also have more resources to throw at future advancements.

        So, then, the question is whether the “profits” on the last advance are enough to pay for the next one. We can define a new term, “affordability”, as what % “profit” you can expect from each advance relative to its cost, telling us whether it becomes relatively easier or harder to continue to advance.

    • vidarh 2 hours ago
      Because they depend on whether the rate of improvement of self-improvement outpaces the rate of increase in difficulty or not, and at some points they clearly do - e.g. a lot of skills makes the relative rate of subsequent improvement easier for a while.

      It may seem obvious that it can't last, but showing the conditions where it can't still matters.

    • red75prime 2 hours ago
      > How is this not obvious to everyone?

      That RSI can be bottlenecked? I guess this is obvious to many people. Whether RSI will be bottlenecked (at some not very interesting stage) is another question.

      • RandomLensman 51 minutes ago
        Why would it not be bottlenecked? Is intelligence so valuable as such?
    • exe34 1 hour ago
      > How is this not obvious to everyone?

      Probably because it's not true. We had shitty neural networks for decades before the recent explosion. That particular branch may be a dead end, but there could be others lurking and waiting for their time.

    • soulofmischief 1 hour ago
      Progress is often lumpy, modulated by new discoveries and their applications.
    • EarlKing 2 hours ago
      > How is this not obvious to everyone?

      Because everyone's thinking around intelligence is incredibly muddled by a variety of factors, and no one is particularly motivated to actually correct anyone's mistaken notions on the matter.

      • godelski 2 hours ago
        I'm conflicted. On one hand I think we should more openly call people idiots and push back. On the other hand there's Descartes argument for idiots in good company.

        I just wish all the people that claimed to care only about truth would actually care about truth. Feels like society is more that trope where someone says a joke to a crowd and no one hears it except one charismatic person who repeats it and gets all the laughs. In reality it feels like the repeated version of the joke doesn't even make sense, it is just vibes.

  • vatsachak 4 hours ago
    RSI isn't anything new though; computers have been used to make computers better for about 80 years now.

    Imagine having a secretary who could read 1 million records and give you back your answer in 100 microseconds, for just 10 cents an hour. That's Postgres.

    So I'd imagine that if R&D can be automated, everything becomes better and cheaper but we'd all lose our jobs, as secretaries did to postgres. UBI season

    • eru 3 hours ago
      > Imagine having a secretary who could read 1 million records and give you back your answer in 100 microseconds, for just 10 cents an hour. That's Postgres.

      Well, that secretary can only answer very specific questions in a rather peculiar format.

      > [...] but we'd all lose our jobs, as secretaries did to postgres.

      I doubt many secretaries were replaced by postgres.

      However, you might like reading about https://en.wikipedia.org/wiki/Unit_record_equipment

    • goatlover 2 hours ago
      That would be a failure of imagination when it comes to everything secretaries actually do. Which is a problem with the idea that AI is coming for all our jobs anytime soon. It totally undersells everything humans do on the job.
      • ACCount37 45 minutes ago
        I think humans totally oversell just how special "everything they do on the job" is.

        "No, you don't get it, it's me specifically who has a job that's way too special to be automated away!"

        • Windchaser 30 minutes ago
          We could have coffee made today by automated coffee vending machines, but many people still prefer to go to a coffee shop. Why is that?

          Sometimes we don’t choose the cheaper automated product, and instead opt for the “human experience”. I don’t imagine this will change any time soon. Humans like being around humans.

  • reader9274 4 hours ago
    First thing I thought when reading the title was "is this about those never ending self-help books?"
  • charcircuit 3 hours ago
    >Best evidence in favor of acceleration:

    It's important to recognize that LLMs accelerating development of LLMs does not imply it will lead to self-sustaining acceleration.

  • Mistletoe 4 hours ago
    When I hear recursive self improvement all I remember are the ridiculous articles a few years ago about how 3d printers were going to make themselves and take over the world.
  • reducesuffering 2 hours ago
    Reminder that while there are many naysayers who have been on the wrong side of AGI development progress the last decade…

    There are two $1T companies who are all-in on RSI internally right now. They are supported by $20T of market cap plowing R&D into their efforts. You can think it’s dumb money at your own peril, however the market rewards intelligent allocation…

    • baq 2 hours ago
      The market can also be quick and brutal to punish mistakes, especially when leverage is high. We can be in 1998, but we can be in 1995; you stand to make a ton of money if you know which one we’re in
    • goatlover 2 hours ago
      How do the markets know it's intelligent allocation before RSI is actually a thing? What makes those companies immune to failure?
      • not-a-llm 31 minutes ago
        them having an actual product with huge demand makes them immune

        Just like Tesla was immune from all the naysayers which were saying its a highly unprofitable company which will 100% go bankrupt because its economics dont make any sense, and they lost huge amounts of money shorting the stock

    • mbgerring 1 hour ago
      “The market rewards intelligent allocation” is such a straightforwardly false statement that I can’t believe anyone still says this with a straight face. The last ten years of the US economy have just been scam after scam after scam, and people just keep saying this.
      • blitzar 57 minutes ago
        NFTs were worth more than $1 trillion so we know that they were "better" than the Ai efforts of today because “The market rewards intelligent allocation”
        • not-a-llm 34 minutes ago
          they never were

          if I sell 1 token at $1 and there are a trillion of them minted, that does not make a $1T market cap

          but if you want to prove a point against AI with fake data, sure

  • joka88xj 3 hours ago
    [flagged]