How safe “safe” AI development?

There are now at least two AI de­vel­op­ment com­pa­nies pur­port­ing to re­search safe AI and both have ex­isted for a cou­ple years now, so I think it’s worth tak­ing an­other look at how safe “safe” AI de­vel­op­ment is.

One of those com­pa­nies is GoodAI. I don’t know a lot about them be­yond what’s on their web­site. They seem clearly to be aware of AI safety con­cerns and the need for al­ign­ment but are also pur­su­ing ca­pa­bil­ities re­search. OpenAI is the other and is similarly pur­su­ing ca­pa­bil­ities re­search but, at least based on what I know about them pub­li­cly, only go so far as to say they want safe AI, al­though OpenAI does em­ploy at least one per­son known to be ac­tively work­ing on AI safety. There may be other com­pa­nies in the safe AI de­vel­op­ment space but to the best of my knowl­edge other efforts do not make ex­plicit state­ments about safety and are only fo­cused on ca­pa­bil­ities (al­though AI for self-driv­ing cars, for ex­am­ple, has cer­tain mun­dane safety con­cerns differ­ent from those of AI safety re­search).

When OpenAI was started there was some dis­cus­sion about whether it was a good idea. Ben Hoff­man said no, Nate Soares had a pos­i­tive re­ac­tion, and oth­ers had mixed re­sponses. Some quick search­ing hasn’t turned up any ex­plicit opinions on GoodAI, so pre­sum­ably peo­ple would feel similarly about them as they do about OpenAI. A quick re­hash on some of the ar­gu­ments:

  • No AI is safe AI un­til we are sure we can al­ign it.

  • Ca­pa­bil­ities re­search is in con­flict with safety re­search so no ca­pa­bil­ities effort can be trusted to be fully com­mit­ted to safety.

  • If you are go­ing to re­search AI any­way, bet­ter to try to be safe than not at all.

  • More peo­ple work­ing on safety is always bet­ter, and the as­so­ci­ated ca­pa­bil­ities re­search these com­pa­nies are do­ing would have hap­pened any­way.

We’re a cou­ple years on now, though, and given that AI seems to be on a strong up­ward swing, does it make sense to en­courage more com­pa­nies to be like OpenAI and GoodAI and tar­get safe AI, per­haps via a self-reg­u­la­tory or­ga­ni­za­tion, or is en­courag­ing safety as an ex­plicit goal in ca­pa­bil­ities re­search un­likely to have much effect?

I’m in­clined to sus­pect that some at­ten­tion to safety is bet­ter than none be­cause it gives a wedge with which to push for more safety later, so I’m es­pe­cially cu­ri­ous about ar­gu­ments that it wouldn’t help.