I was reminded of OpenAI’s definition of AGI, a technology that can “outperform humans at most economically valuable work”, and it triggered a pet peeve of mine: It’s tricky to define what it means for something to make up some percentage of “all work”, and OpenAI was painfully vague about what they meant.

More than 50% of people in England used to work in agriculture, and now it’s far less. But if you said that farming machines “can outperform humans at most economically valuable work” it just wouldn’t make sense. Farming machines obviously don’t perform 50% of current economic value.

The most natural definition, in my opinion, is that even after the economy adjusts, AI would still be able to perform >50% of the economically valuable work. Labor income would make up <<50% of GDP, and AI income would make up >50% of GDP. In a simple Acemoglu/Autor task continuum model of the economy where there is a continuum of tasks, this corresponds to AI doing more than 50% of the tasks.

I don’t think OpenAI’s definition is wrong, exactly, since it does have a reasonably natural interpretation. But I really wish they’d been more clear.

A simple “null hypothesis” mechanism for the steady exponential rise in METR task horizon: shorter-horizon failure modes outcompete longer-horizon failure modes for researcher attention.

That is, with each model release, researchers solve the biggest failure modes of the previous system. But longer-horizon failure modes are inherently rarer, so it is not rational to focus on them until shorter-horizon failure modes are fixed. If the distribution of horizon lengths of failures is steady, and every model release fixes the X% most common failures, you will see steady exponential progress.

It’s interesting to speculate about how the recent possible acceleration in progress could be explained under this framework. A simple formal model:

• There is a sequence of error types e_1, e_2, e_3, etc.
• The first s error types have already been solved, such that errorrate(e_1) = errorrate(e_s) = 0. The long tail of errors e_{s+1} etc has error frequencies decaying exponentially.
• With each model release (t -> t+1), researchers can afford to fix n error types.
• METR time horizon is inversely proportional to the total error rate sum(errorrate(e_i) for all i)

Under this model, there are only two ways progress can speed up: the distribution becomes shorter-tailed (maybe AI systems have become inherently better at generalizing, such that solving the most frequent failure modes now generalizes to many more failures), or the time it takes to fix a failure mode has decreased (perhaps because RLVR offers a more systematic way to solve any reliably measurable failure mode).

(Based on a tweet I posted a few weeks ago)

Here’s a random list of economic concepts that I wish tech people were more familiar with of. I’ll focus on concepts, not findings, and on intuition rather than exposition.

• The semi-endogenous growth model: There is a tug of war between diminishing returns to R&D and growth in R&D capacity from economic growth. For the rate of progress to even stay the same, R&D capacity must continually grow.
• Domar aggregation: With few assumptions, overall productivity growth depends on sector-specific productivity growth in proportion to sectors’ revenues. If a sector is 11% of GDP, the economy is “11% bottlenecked” on it.
• Why wages increase exponentially with education level: This is empirically observed to be roughly true (the Mincer equation), but why? A simple explanation: the opportunity cost of education is proportional to the wage you can earn with your current level of education. So to be worthwhile, obtaining one more year of education needs to increase your wage by a certain percentage, no matter your current level of education. Each year of education earning people 10% more will look like an exponential.
  • This is basically “P = MC”, but applied to human capital.
• Automation only decreases wages if the economy becomes “decreasing returns to scale”. This post has a good explanation. Intuition: if humans don’t have to compete with automated actors for things that humans can’t produce (e.g., land or energy), humans could always just leave the automated society and build a 2025-like economy somewhere else.

I did not appreciate until how unique Less Wrong is as a community blogging platform.

It’s long-form and high-effort with high discoverability of new users’ content. It preserves the best features of the old-school Internet forums.

Substack isn’t nearly as tailored for people who are sharing passing thoughts.

Does a need for broad automation really place a speed limit on economic growth?

I’ve been trying to better understand the assumptions behind people’s differing predictions of economic growth from AI, and what we can monitor — investment? employment? interest rates? — to narrow down what is actually happening.

I’m not an economist; I am an engineer who implements AI systems. The reason I want to understand the potential impact of AI is because it’s going to matter for my own career and for everyone I know.

In the spirit of “learning in public”, I’ll share what I’ve learned (which is a little) and what’s not making sense to me (which is a lot).

In Ege Erdil’s recent case for multi-decade AI timelines, he gives the following intuition for why a “software-only singularity” is unlikely:

The case for AI revenue growth not slowing down at all, or perhaps even accelerating, rests on the feedback loops that would be enabled by human-level or superhuman AI systems: short timelines advocates usually emphasize software R&D feedbacks more, while I think the relevant feedback loops are more based on broad automation and reinvestment of output into capital accumulation, chip production, productivity improvements, et cetera.

The implicit assumption is that chip production, energy buildouts, and general physical capital accumulation can only go so fast.

Certainly, today’s physical capital stock took a long time to accumulate. With today’s technology, it’s not feasible to scale physical capital formation by 10x, or to 10x the capital stock in 10 years; it would simply be far too expensive.

But economics does not place any theoretical limit on the productivity of new vintages of capital goods. If tomorrow’s technology was far more effective at producing capital goods, physical capital could grow at an unprecedented rate.

(In frontier economies today, the speed of physical capital growth is well below historical records. For reference, South Korea’s physical capital stock grew at ~13.2% per year at the peak of its growth miracle. And from 2004 to 2007, China’s electricity production grew at an average of ~14.2% per year.)

Today’s physical capital has two unintuitive properties that creates a potential for it to be produced at much lower cost, despite the intuition that “physical = slow to build”.

• As income increases, consumption is “dematerialized”. Spending on physical capital increasingly goes toward high-value-to-weight items, such as semiconductors and medical devices. Explosive economic growth may not mainly be about building, say, 10 houses and airplanes per human (how would this even get used?) but rather building increasingly elaborate hospitals, medical equipment, and so on.
  • The barrier to creating these goods may lie more with R&D and design and coordination, rather than physical throughput limits. Some physical work is still required, but physical throughput is not necessarily the bottleneck.
• High-value physical capital embeds a great deal of skilled labor. Capital equipment is produced using factors of production that can be recursively attributed to non-capital inputs — labor, energy, and natural resources. This is reflected, for instance, in BEA’s industry-level “KLEMS” (capital, labor, energy, materials, and services) accounts. The supply chain of a complex capital good generally has skilled labor as a major, or even dominant, portion of the value added in its production.
  • Example: MRI machines. Medical MRI machines are very expensive, often costing more than $1 million per unit. I asked o3 to recursively break down the cost; I’ll keep the inferences high-level, since I don’t trust o3’s precise claims. Essentially, a large portion of the cost of an MRI machine is because the superconducting magnet requires a great deal of specialized equipment to make, and the equipment is complex and produced at low volume. The reason why low-volume capital goods are expensive is that they embody considerable skilled labor amortized over few units.

Physical constraints on the replication rate of capital goods do exist, but in at least one major case are far from binding. Solar panels, for example, require ~1 year to pay back their energy cost. If solar panels were the only source of electricity, GDP could not grow at >100% per year while remaining equally electricity intensive. But this is far, far, above any predicted rate of economic growth, even “explosive” growth, which Davidson and Erdil and Besiroglu define to start at 30% per year.

I see a strong possibility that if human-level skilled labor were free, the capital stock could grow at an unprecedented rate.