Thread by @MartinBeraja

🤖🛑 Hitting the breaks on AI?

(1) Alignment worries: perhaps legit. But don't have much to say.
(2) Job automation concerns: LLMs ≠ robots. Economic arguments for slowing down automation are much weaker for LLMs.

Why? Let's dive in! #EconTwitter #FutureOfWork #AI #ChatGPT

Economists have put forth two main arguments for slowing down technologies that automate jobs and displace workers. The first is based on redistributive considerations. The second is based on efficiency considerations.

In a nutshell, the redistributive rationale is as follows. If automation displaces workers who are relatively poor, then this increases inequality. Absent other tools to directly redistribute (e.g., transfers to workers) a government should curb automation to reduce inequality.

This argument is well developed in work by @jptguerreiro-@SergioRebelo6-Teles; Costinot-@IvanWerning; @akorinek-@JosephEStiglitz.

The redistributive argument is much weaker for LLMs than it was for robots.

Robots automate jobs intensive in routine tasks. These tend to be low-to-middle-wage jobs (think workers in a car manufacturing plant). So automation increases inequality.

From what we know so far, LLMs are likely to automate cognitive-intensive jobs. These tend to be middle-to high-wage jobs (think lawyers or software devs). So LLM-led automation would likely decrease inequality. The redistributive rationale for slowing down automation is gone.

The efficiency rationale is in my paper with @Nathan_Zorzi. See this thread for the argument:

https://twitter.com/MartinBeraja/status/1545018823463473153?s=20

In a nutshell, firms automate excessively because
there is a conflict between how firms and displaced workers value the gains from automation.

The efficiency rationale for slowing down automation relies on two assumptions. First, automated workers are financially vulnerable. They are unable to borrow or don't have enough savings to finance consumption while relocating to a new job. Second, worker reallocation is slow.

Both assumptions seem much less relevant for workers displaced by LLMs than by robots.

Again, LLMs will displace relatively richer workers in cognitive-intensive jobs. These workers are likely to have sufficient savings or able to borrow while relocating to a new job.

That is, in practice, lawyers and software devs are not the first that come to mind when thinking about "financially vulnerable workers." Workers displaced by a robot in a car manufacturing plant in Detroit, on the other hand, much more so.

Moreover, and this is more speculative, relocation is likely to be faster/easier for high-cognitive-skill workers than for high-manual-skill workers.

So, while relevant for robots, neither of the two assumptions that the efficiency rationale relies on seem that relevant for LLMs. The case for slowing down LLM-led automation due to inefficiencies is weak.

In all, neither redistributive nor efficiency arguments seem to support slowing down LLMs due to its impact on workers in a way that do support slowing down robot adoption

AI misalignment remains a valid concern, in principle.

But I’d like to see more formal economic arguments rather than armchair theorizing; to spell out assumptions and evaluate them. At this point, I don’t feel we have such arguments to so confidently call for an AI moratorium.

Some who inspired this thread and may be interested: @avicgoldfarb @joshgans @erikbryn @ben_golub @Jabaluck @Prof_Yildirim @annastansbury @emollick @sama @ManlikeMishap @elonmusk @AndrewYang @GaryMarcus @geoffreyhinton @tegmark @ESYudkowsky @Grady_Booch @JeffLadish @Noahpinion