Basic argument

In short —

1. There exists an adaptive limit representing the maximum rate at the economy can adapt to new technology.
2. Exceeding this limit yields no socio-economic benefit, in the general case.
3. Exceeding this limit imposes significant risks and costs.
4. AI growth-rate currently exceeds the adaptive limit.
5. Therefore, we should slow rate of AI growth-rate.

Is this argument sound?

(1) There exists an adaptive limit.

The economy is a complex adaptive system — it tries to maintains a general equilibrium by responding to exogenous changes. These responses are internal homeostatic processes such as the price mechanism, labour movements, and the stock exchange.

However, these homeostatic processes are non-instantaneous. Prices are sticky, employees have contracts, and the NASDAQ closes at 8:00 PM. The internal homeostatic processes which equilbriate the economy operate on a timescale of years, not weeks.

Imagine if OpenAI had released ChatGPT-3.5 but hadn't released ChatGPT-4. The release of ChatGPT-3.5 would've had a profound impact on the economy. Here's the standard neo–Keynesian story —

Firstly, people start using ChatGPT-3.5 to automate part of their jobs, increasing their productivity. Some employees use ChatGPT-3.5 to complete more tasks within the same working hours, whereas other employees use ChatGPT-3.5 to complete the same tasks within fewer working hours. In both cases, value is created.

Because employees are completing more tasks, the companies start shipping more products and services. Supply goes up, and quality goes up too. So investors and shareholders receive more money. Because employees take more leisure time, their jobs looks more attractive to job-seekers. Hence the supply for labour increases, and the wages decrease, and the demand for such labour, at the new price, increases. Because the labour costs have dropped, competing companies will lower prices to gain more market share. And new industries, which would've been unprofitable in the previous economy, are now profitable and viable. This leads to a real-terms wage increase.

After a few years, the economy reaches (or at least approaches) a new general equilibrium in which...

1. People have more money to spend.
2. The products and services are more abundant, cheaper, and of a higher quality.
3. People have more leisure to enjoy themselves.

This is what we call an "AI Summer Harvest".

Notice that the benefit of ChatGPT-3.5 dissipates slowly throughout the economy. Eventually, the economy approaches a new general equilibrium, but this is a non-instantaneous process.

(2) Exceeding this limit yields no socio-economic benefit.

Our story starts with ChatGPT-3.5 and ends with higher social welfare.

This is a common feature of technological progress — a new general-purpose technology will normally yield these three benefits:

1. People have more money to spend.
2. The products and services are more abundant, cheaper, and of a higher quality.
3. And people have more leisure to enjoy themselves.

This follows from standard neoclassical economics.

However, the neoclassical analysis only applies when the economy is near-equilibrium. When the economy is far-from-equilibrium the neoclassical analysis is agnostic, and we must turn to other macroeconomic methods.

According to mainstream macroeconomic theory, until the sticky prices are adjusted, fired employees are rehired, and investments are reallocated, the economy will not be in a better socio-economic situation.

(3) Exceeding this limit imposes significant risks and costs.

This has been covered extensively elsewhere. I'm alluding to —

• Economic shocks (unemployment, inflation, inequality)
• Misinformation, fraud, blackmail, other crimes.
• Extinction-level risks.

(4) AI growth-rate currently exceeds the adaptive limit.

ChatGPT-3.5 was released circa December 2022, and GPT-4 (in the form of Bing Sydney) was released circa February 2023. That's a three-month gap! This definitely exceeded the adaptive limit. The economic impact of ChatGPT-3.5 would probably have dissipated through the economy over 5–10 years. 

That's my rough intuition —

• Students study in university for 3–6 years.
• Employees change their job once every 5–10 years.
• Politicians are re-elected on a timescale of 5–10 years.
• Companies can persits for 5–10 years in an unfit economy.
• Big government projects take about 2–10 years.

So 5–10 years is the timescale over which the economy (and society as large) can adapt to socio-economic shocks on the scale of ChatGPT-3.5.

This implies that ChatGPT-4 could've been released in 2028–2033 and we would've yielded the same socio-economic benefits. Why? Because the bottleneck on labour automation isn't that companies need a GPT-4-based chatbot rather than a GPT-3-based chatbot, but rather that firing employees happens very slowly, over years, not weeks.

(5) Therefore, we should slow rate of AI growth-rate.

I think we should slow the growth-rate of AI to below 0.2 OOMs/year, equivalently a 58% year-on-year growth, equivalently a doubling-time of 18 months.^[1]

What does this mean in practice?

On March 15 2023, OpenAI released GPT-4 which was trained with an estimated 2.8e+25 FLOPs. If OpenAI had followed the 0.2 OOMs/year target, then GPT-4 would've been released on March 29 2029. We can see this by examining the proposed training-run compute limits, shown below. See the original post for details.

Implications

• Because economic equilibriation is non-instantaneous, there are smaller near-term financial incentives to train GPT-$(n>4)$. This makes coordination easier, and also lengthens timelines conditional on no coordination.
• Timelines are more aggressive in worlds with neoclassical economics — many competitors, instantaneous transactions, no transaction costs, non-sticky prices, etc. These things increase the timescale of economic equilibriation.
• Governments (which are mostly neo-Keynesian) avoid economic shocks with significant short-term downsides (e.g. mass unemployment), even if the long-run equilibrium resulting from the shock would be socio-economically productive. Government prefer to slow the timescale of exogenous shock to the adaptive limit. This increases the likelihood that policy-makers will coordinate on slowing AI.
• There is a trending metaphor of an AI Summer Harvest which compares slowing AI to a summer harvest. This article presents an argument underlying the comparision — there's no benefit in a farmers sowing seeds at a greater rate than they harvest crops. The benefits of an AI Summer Harvest should be communicated clearly and repeatedly to policy-makers and the general public.
• A thermodynamic system undergoes a quasi-static process if it remains close to equilibrium throughout the process. This happens when the exogenous change occurs slower than the timescale of the system's internal equilibration. For example, a normal atmospheric heat wave is quasi-static whereas a pressure wave is not. Hence, an equivalent way to say "slow AI growth-rate to the adaptive limit" is to say "AI development should be a quasi-static process".

1. ^{^}

   The "0.2 OOMs/year" figure was first proposed by Jaime Sevilla, Director of EpochAI (personal correspondence).