13 Comments
User's avatar
Yupanqui's avatar

a one time step instead of a new slope is the most reassuring read i've seen this week. hope the next point agrees

Mike Boysen's avatar

This is the most useful framing of the "explosion" question I've seen, because it asks what the progress cost and not just how steep the line is.

I'd push on the numerator. A benchmark index measures what a model can do on a test. Buyers pay for what they can now get done that they couldn't before, and how reliably. Those two don't move in lockstep. A few index points can take a model past the threshold where an entire job (or a step in one) becomes something you can hand off. Then nothing happens for a while until the next threshold. So output per dollar can fall on the index while value per dollar jumps.

Your "Organizing Growth" point maps onto this neatly. The common problems get solved first. What's left are rarer, harder outcomes that need more specialized effort. That's diminishing returns on capability, but not necessarily on demand.

The question I'd love to see estimated is how many more customer jobs each doubling of spend turns into ones that can be reliably delegated. That's the number that decides whether the reinvestment loop keeps paying for itself.

Luis Garicano's avatar

Thanks. Your question is of course the right one and I don't have an answer in that metric. Would love to see someone digging into it.

Simon Holzapfel's avatar

Love this post. Thanks for breaking it down this way. Did you contemplate any potential impacts from quantum compute as a bottleneck breaker or from total factor productivity positive feedback loops [maybe I missed that]. In particular if you look at flow time [I can explain offline if this is unfamiliar] for work units in organizations now, it’s over 90% inefficient, so tons of room to grow. @Brian Potter just put out a stack about search space and innovation that might be relevant here, too.

Luis Garicano's avatar

Thanks! No we have not considered quantum computing impact on organizations, would love to hear a bit on this.

Michael Wiebe's avatar

>Hence AI capabilities improved 1.91 times faster after October 2025, but R&D spending grew 6.19 times and the technical workforce 3.10 times.

This is comparing a change in the capability growth rate to a change in the input level. Shouldn't we compare to the change in the input growth rate?

Luis Garicano's avatar

Thanks, Michael. That paragraph reports progress per dollar or per researcher-year. The numerator is capability-index points gained per year; the spending denominator is dollars spent per year. Dividing gives:

index points gained/year /dollars spent/year=index points gained/dollar.

So if annual progress is 1.91x while annual spending is 6.19x, progress per dollar goes down to 1.91/6.19 ≈ 31% of its previous value. The analogous calculation gives a 38% reduction in progress per technical employee-year. Comparing the rate of progress with the number of researchers is also the accounting logic used by Bloom et al.

I hope this is clearer

Michael Wiebe's avatar

But isn't that still comparing f'' to f'? I get a 1.9x speedup in ECI growth compared to a 1.6x speedup in R&D spending growth.

https://x.com/michael_wiebe/status/2120540887327629396

Marshall Yount's avatar

This research is intriguing!

Followup question: Did you consider the impact of purchased training data (from Mercor, et al) with respect to the October 2025 hinge date?

FY2024 $196M

FY2025 $361M

2026 H1 $430M

2026 Q3 $597M

(each numbers is +/- $30M, the range is broad because the S1 doesn't show the 2023 starting value)

These costs are growing very, very fast and there is an inflection point very near the October 2025 "hinge date".

In some sense, these training dates represent an outsourced set of labor costs. If I understand the filings correctly, these vendor costs are a subset of the R&D expenditures described in Note 8 of the S1.

Daniel Florian's avatar

Thanks for this insightful article - one critical comment though: You write that ideas are harder to find and yet finding only one idea can change the whole field of AI research. Take mathematics: LLMs used to be pretty bad at maths until reasoning made them solve the trickiest math problems. What makes you confident that we are not also a single idea a way of recursive self-improvement?

Kazimierz Stanczak's avatar

Extremely interesting and thought-provoking. Thank you! Also, for making me think quite hard:)

I have one comment.

Consider your research production function in Section 2:

dA/dt = δX^λ A^ϕ.

For ϕ < 1, a balanced growth path with g_A > 0 requires:

g_X = [(1 − ϕ)/λ]g_A.

But if research expenditure X is a constant fraction of aggregate output Y along a balanced growth path, then g_X = g_Y. This imposes an additional restriction linking the growth rates of knowledge and aggregate output.

In Section 3, allowing AI to supply its own research inputs introduces another issue. Your explosion condition,

ϕ + λψ > 1,

implies accelerating rather than balanced growth in the simplified autonomous research system.

So my question is: can these technologies be embedded in a general equilibrium admitting a balanced growth path? I am not sure, given the restrictions discussed by King, Plosser, and Rebelo (1988).

More importantly, would your conclusions survive once research inputs, factor prices, and aggregate resource constraints are determined endogenously?

Luis Garicano's avatar

Wow, thanks Kazimierz for reading carefully, thanks for the questions! First, a general comment for your (correct) comment on section 2 and the comment on section 3. Balanced growth imposes additional restrictions, but we are not looking for balanced growth but for an intelligence explosion, hence an equilibrium *without* balanced growth.

1. You are right that a balanced growth path imposes an additional restriction.

2. On Section 3, you are right, if the exponent of the research equation ϕ + λψ is larger than 1, then there is no balanced growth. An an equilibrium need not follow a balanced growth path. Aghion, Jones and Jones explicitly examine both balanced and explosive growth, as well as bottlenecks that can prevent an explosion.

3. On your final question, AJJ already incorporate equilibrium factor prices, capital accumulation and resource constraints. We attempt to estimate the research technology within such a framework.

Kazimierz Stanczak's avatar

Thank you, Luis. It's all clear.:)