Stefano Romano

Tech startups guy, parent, trying to figure out the AI situation. Previously math, physics, poker

[Dwarkesh Patel @dwarkesh_sp]

I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower (a Mafia or Werewolf type social deception game).

The results were fascinating. Models have gotten much better over the course of this year at creative on-the-spot scheming and bluffing, and also at reasoning as a team to break apart these underground conspiracies.

I go through some of the most standout games these AIs played in this video. The new models are getting pretty good. If my friends pulled off some of these moves while we were playing, I’d give them a fist bump.

Fwiw, OpenAI’s newer models are much better than Anthropic’s at this game.

You can see the full benchmark at the website, and I’ve also made it so that you and your friends can play again the AIs: https://t.co/IF4d0EMSD8

[Stefano Romano @SSte_romano]

@GaryMarcus What does "new generation" mean in #5? Do you mean specifically this will come from companies other than the current frontier labs?

[Stefano Romano @SSte_romano]

@erminalt19216 @suddenlyFleck @imobulus @robbensinger Sure. I mean I personally think this is a very poor rendition of the x-risk from AI argument, and would encourage you to read more about it, but at least you're engaging with the argument. My point here was just in relation to your original "falsifiability is a blocker" claim

[Stefano Romano @SSte_romano]

@erminalt19216 @suddenlyFleck @imobulus @robbensinger In other words, you can disagree on how the AI safety community has argued that the threat model from AI is more probable than background noise, but not dismiss it on the grounds of it being unfalsifiable

[Stefano Romano @SSte_romano]

@erminalt19216 @suddenlyFleck @imobulus @robbensinger Yes, exactly! And by definition, for new emerging threats, that involves some amount of logical deduction, reasoning by analogy, and forward prediction. The threat model is by default not falsifiable, but that doesn't mean much

[Stefano Romano @SSte_romano]

@erminalt19216 @suddenlyFleck @imobulus @robbensinger Yes, exactly! And by definition, for new emerging threats, that involves some amount of logical deduction, reasoning by analogy, and forward prediction. The threat model is by default not falsifiable, but that doesn't mean much

[Stefano Romano @SSte_romano]

@erminalt19216 @suddenlyFleck @imobulus @robbensinger Let me ask you this way: I agree that AI posing existential risk is an unfalsifiable claim short of the thing itself happening. Does that mean we are safe?

[Stefano Romano @SSte_romano]

@erminalt19216 @suddenlyFleck @imobulus @robbensinger What's wrong with idea that a smarter-than-you entity might do something you hadn't thought of, or you might not be able to understand even in retrospect? Yes, by definition you can't falsify this before it happens. So what?

[Stefano Romano @SSte_romano]

@patrickc @humphreybc interesting idea

[Stefano Romano @SSte_romano]

@littmath @WKCosmo I appreciate you put in more effort than most others who stop at "math is about understanding", which is not really an argument for why math research matters. But it still warrants more words imo, e.g. why is math a good way to build human capital, relative to alternatives?

[Stefano Romano @SSte_romano]

@ylecun @francoisfleuret What's an example or a "regular LLM" that anyone has used in the last year?

[Stefano Romano @SSte_romano]

@clairlemon Yes. The principal axis of the AI debate is not safe vs risky, but big deal vs no big deal. Risk deniers tend to think of AI as normal technology, and conversely if it's not normal tech is natural to believe it carries abnormal risks

[Stefano Romano @SSte_romano]

@Ognifedefingo To an extent you can test this, eg they appear to materially change behaviour in response to (what they describe as) pain, in a way that seems consistent with the concept of pain but not of pleasure

[Stefano Romano @SSte_romano]

@labububibibobo @nic_carter @timricchuiti it's not clear how the existence of math that no one bothers to read or understand hurts human math progress in any way

[Stefano Romano @SSte_romano]

@perrymetzger Could you point at the logical fallacies you noticed? No need to explain why they are fallacies if you have done so in the past (though a link would be appreciated!)

[Stefano Romano @SSte_romano]

@clairlemon I don't know anything about your past work, but I "low quality voice on AI" seems straightforwardly true. I don't remember you attempting a single substantive argument, your entire engagement afaict has been at the level of simply stating your views and mocking opponents

[Stefano Romano @SSte_romano]

@AndrewCurran_ @aapakari It'd be kind of insane if they announced a solution to Hodge in this fashion

[Stefano Romano @SSte_romano]

@binarybits @mentalgeorge This was very good, though I think it too easily dismisses the idea that broad generalisation might just happen at some point from sufficient optimization in narrow domains. But overall, solid points and well articulated, thanks for the link!

[Stefano Romano @SSte_romano]

@nathanhamiel I don't think the difference in perception is really about what strong control looks like and can achieve, but more about the extent that control is a sufficient approach to safely developing a smarter-than-human entity. No one has work experience dealing with the latter

[Stefano Romano @SSte_romano]

@clairlemon What are examples of AI safety people rejecting security practices that would be effective today? (rejecting as in arguing these practices are net negative, as opposed to just insufficient)