[This rant is directed mainly at people who probably won’t read it.]
Proposals to slow AI development have been criticized as special-interest lobbying. E.g. Marc Andreessen:
One-hundred percent, there’s a regulatory capture element to it. And they’re driving hard for that.
I’ll concede for the sake of argument that AI companies would like regulatory capture.
Dario Amodei, Sam Altman, and Elon Musk are all ten times too ambitious to settle for mainly achieving regulatory capture. Regulatory capture is effective at preserving existing corporate revenue and jobs. It’s a terrible way to colonize Mars or cure all disease. Those CEOs have made it clear that their ambitions are at least as big as colonizing Mars.
If regulatory capture were an important goal, I’d expect AI companies to focus on scaring voters about the risks of open weight AIs. Instead, they seem to treat open weight AIs as a low priority issue.
A stereotypical company seeking regulatory capture would be saying something like “our products are safe, please protect them from unproven alternatives”. Instead, they’re saying things that amount to “we’re scared of our products, they’re almost as dangerous as those of our leading competitors”. With the possible exception of Musk – I can’t tell whether he’s scared of what he describes as a 10 or 20% chance of doom.
See Alex Tabarrok for an expert opinion about what regulatory capture look like.
Science Fiction
Andreessen has been consuming too much science fiction, and not consuming enough expert analysis on the nature of intelligence. He imagines that our main risk is poverty, rather than being conquered.
Andreessen seems focused on the kind of future depicted by Star Trek, where the privileged are mostly too poor to afford their own starship, and some planets are so poor that a small outbreak of tribbles might cause them to starve.
The more I study AI-related topics, the more implausible that kind of poverty looks. It would take some really weird forces to go 240 years from now with approximately no improvement in AI or robotics. That’s roughly what it would take for the Star Trek future to make sense.
Maybe a Star Trek level of poverty could be achieved by destroying all semiconductor fabs. I promise you that AI CEOs will oppose that plan.
Longer Term Concerns
A steelman of the regulatory capture claim would be that regardless of what they want now, AI CEOs will pivot to goals such as banning open weight AI once they’ve accomplished their current goals.
I see some fairly strong limits to how much they could accomplish. There are powerful companies whose business is helped by the availability of open weight AIs: Nvidia, Intel, Micron, etc. They will lobby against many of the harmful regulations that might help AI companies. They have a bit more experience at lobbying than do the younger AI companies.
There’s also a fair amount that can be done to reduce regulatory capture without reducing regulation. E.g. regulators can be prohibited from going to work for AI companies within a few years after leaving the regulatory agency.
How bad is regulatory capture?
Typical concerns about regulatory capture are that it increases prices and reduces innovation.
One large set of effects would be on AI’s ability to replace human labor. Regulatory capture might well mean that the AI which outcompetes me at investing will cost $100k per year rather than the $50k it would cost if open weight AI were unrestricted. And my investing skills might not become obsolete until 2035 under regulatory capture, compared to 2030 without regulation. I have mixed feelings about these effects.
I am pretty concerned about the risks that regulation will delay cures for aging and infectious diseases. But that risk seems to mainly come from regulators who have either been captured by big pharma, or who are uncaptured by industry and averse to any changes which could cause negative publicity. I’m unclear as to why regulators who are captured by AI companies would cause anywhere near the kind of delays that we’re currently seeing in medicine.
Regulatory capture seems likely to have an important effect on military drone swarms. I’d really like to see delays in innovation in this area, provided that the regulatory capture spreads internationally. This seems like it ought to be the biggest influence on whether I consider regulatory capture to be good or bad. But I’m confused as to how to evaluate it.