Reactions from ranked influencers
12 posts
@emollick It’s about open models reducing margins, so fewer profits to invest in r&d, so AGI comes later. I see the idea but given the US is already so ahead in investment, seems totally fine.
@jachiam0 I made a similar clarifying distinction here
It's worthwhile to be more precise in (3) and claim something like "open models are scale-up decelerating". That argument is reasonable, though not airtight or quantitative, e.g. there is a maximum amount of capital that will flow to scale-up at any moment, a maximum velocity to scale-up buildout, etc., and we seem rather close to some of these despite the continual flow of open models. Reminder that the production of intelligence is a self-commoditizing process, insofar as the models are generally available and not restricted in ways to prevent this. Despite these and other good counterarguments, at least in the abstract, I basically agree that open models probably decelerate scale-up on the margin. That open models are scale-out decelerating is harder to show because it's not clear how all the various factors add up. For example, in worlds where models are closed or safeguarded in ways that block lawful economic work from being done with them outside of labs, the diffusion rate of that intelligence level is by definition much slower unless you're in a world where lab capital is sufficiently large to be able to literally capture the total utility of their models, which is likely a meaningful fraction of GDP. Mythos/Fable being a closed model is the obvious example of scale-out deceleration and it seems like some labs might want to continue on this path. Fable blocking AI research, for example, is probably net decelerating, but it's still hard to count in part because, as many have pointed out before, diffusion that's too rapid or too ungoverned can at least in principle cause overreaction and deceleration in expectation. Overall I am tired of these categories
@natolambert If the arrival of open models increases the market size or diffusion rate then quite plausibly it will be accelerationist because that will be a lot more capital chasing the frontier. Just like we now have exponentially more capital chasing AI since we know demand is insatiable.
@jachiam0 > This seems like a genuinely plausible argument, though it feels like the kind of claim that is best made with a mathematical model I agree but I think the sensible prior is that this at most "decelerates" us by some low double digit percent. AI capability growth is exponential
There's a lot riding on the claim "reducing frontier margins -> CAPEX appetite decreases". CAPEX appetite may just grow with capabilities, regardless of where it comes from, regardless of whether it's trailing by 6-12 months or not. Memory shortages + supply chain price hikes have squeezed margins, and yet the buildout accelerates
A model might be: Q(P,C)=αC^η−βP where Q is Total usage, P price, C capability. Capability boost and price sensitivity is the rest. If high demand elasticity, or strong capability response, accelerationist dominates on volume. If high fixed costs relative to margins, decelerationist force stronger on frontier. My tldr; if the OS world is pure substitute for frontier, then it compresses investment incentive, and if it's not and expands market, then it's ok. A dirty chart here.
There has been a lot of discussion about Point 3 here, the claim that open weight models or open source AI is decelerationist (with much subsequent discussion about related implications in Point 4). Most of the discussion is noise, rage, and honestly quite useless. There is a substantive question embedded in this point and we would all do well to try and address it by careful reasoning and modeling, instead of giving knee-jerk reactions and impugning motives. For what it's worth: personally I am in favor of a healthy open weight / open source ecosystem that lags the frontier by a bit, and I think we're collectively *much* better off if there are strong American open weight models available. Nonetheless, questions about the dynamics of AI capex are tangible and worthwhile. Can we at least firm up how the question is posed before we rage? In my head, it goes like this. Open weight models are plausibly substitutes for closed-source models. This has at least two effects on prices for tokens. 1) The availability of open weight models drops the cost of tokens to the cost of inference (as long as they are adequate substitutes for closed-source models). A decrease in the cost of tokens drives an increase in demand for AI: the total volume of tokens getting used goes up. Adoption increases at a faster pace. Because the demand is there for AI inference compute, spend on AI inference compute may increase - as long as the price per token, which depends on the capability level of the models, is such that an investment in a GPU for inference recoups the cost of the investment in a reasonable timeframe. (This effect would not have been possible five years ago, when models were definitely not good enough to drive consumer compute demand.) This effect is somewhat accelerationist: it potentially leads to *inference compute buildout.* 2) The availability of a substitute for closed source models means that the big AI companies cannot make as much of a margin on the frontier models they sell access to. It becomes harder for big AI companies to recoup the cost of their investment; a lower return on investment may depress the amount of investment. Even introducing a little bit of uncertainty can trigger cascading reactions that slow down the waves of investment in AI by a lot. This is a decelerationist effect. But there is a subtlety here: if there's an increased demand for tokens, does that compensate? *Maybe.* We have to consider one other factor: the increased demand for tokens may increase demand for *inference compute* but not necessarily *training compute,* and these are a bit different. They are not perfectly fungible for each other. So the return on building training compute may (emphasis: may! not a guarantee!) go down even if the return on inference compute goes up. This means even if AI adoption becomes more widespread, the frontier may move forward a little bit more slowly. My understanding of Dean's point (and I welcome corrections from Dean on this) is that if the state intervenes by providing subsidies for training open weight models or otherwise distributing them, that may have the net effect of locking in the frontier where it currently is, by making the kind of capex required for advancing the frontier less attractive to investors. (People who wish for a pause or slowdown: take note, this is actually a strategy you should seriously consider.) This seems like a genuinely plausible argument, though it feels like the kind of claim that is best made with a mathematical model that is a function of variable assumptions, rather than a claim one can make by fiat.