Context for posting link:
Sixteen months ago, I read a draft by a researcher whom few in AI Safety know about, Forrest Landry.
Forrest claimed something counter-intuitive and scary about AGI safety. He argued toward a stark conclusion, claiming he had nailed the coffin shut. I felt averse about the ambiguity of the prose and the (self-confirming?) confidence of the author.
There was no call to action – if the conclusion was right, were we not helpless to act?
Yet, profound points were made and stuck. I could not dismiss it.
But busy as I was, running research programs and all that, the matter kept dropping aside. It took a mutual contact – who had passed on the draft, and had their own doubts – to encourage me to start summarising the arguments for LessWrong.
Just before, I tried to list where our like-minded community fails to "map the territory". In at least six blindspots, we tended to overlook aspects relevant to whether work we scale up, including in AI safety, ends up having a massive negative impact. Yet if we could bridge the epistemic gap to different-minded outsiders, they could point out the aspects.
Forrest’s writings had a hippie holistic vibe that definitely marked him as a different-minded outsider. Drafting my first summary, I realised the arguments fell under all six blindspots.
Forrest wrote back feedback, which raised new questions for me. We set up a call.
Eleven months ago, Forrest called. It was late evening. I said I wanted to probe the arguments. Forrest said this would help me deal with common counter-arguments, so I knew how to convince others in the AI Safety community. I countered that my role was to find out whether his arguments made sense in the first place. We agreed that in practice, we were aligned.
Over three hours, Forrest answered my questions. Some answers made clear sense. Others slid past like a word salad of terms I could not grog (terms seemed to be defined with respect to each other). This raised new questions, many of which Forrest dismissed as side-tangents. It felt like being forced blindly down a narrow valley of argumentation – by some unknown outsider.
That was my perspective as the listener. If you click the link, you will find Forrest’s perspective as the explainer. Text is laid out in his precise research note-taking format.
I have probed at, nuanced, and cross-checked the arguments to understand them deeply. Forrest’s methods of defining concepts and their argumentative relations turned out sensible – they felt weird at first because of my unfamiliarity with them.
Now I can relate from the side of the explainer. I call with technical researchers who are busy, impatient, disoriented, counter-argumentative, and straight-up averse to get into this shit – just like I was!
The situation would be amusing, if it was not so grave.
If you want to probe at the arguments yourself, please be patient – perhaps start here.
If you want to cut to the chase instead – say obtain a short, precisely formalised, and intuitively followable summary of the arguments – this is not going to work.
Trust me, I tried to write seven summaries.
Each needed much one-on-one clarification of the premises, term definitions and reasoning steps to become more comprehensible to a few persons who were patient enough to ask clarifying questions, paraphrase back the arguments, and listen curiously.
Better to take months to dig further, whenever you have the time, like I did.
If you want to inquire further, there will be a project just for that at AI Safety Camp.
I did try to read some of Forrest Landry's writing just now. I understand that this was a light reading, not the kind of deep engagement you are suggesting. I'm just saying that I'm not going to engage deeply (and don't feel like I should) and will try to briefly explain why.
I ended up convinced that this isn't about EA community blindspots, the entire scientific community would probably consider this writing to be crankery. That may reflect a blindspot of the scientific community, but I think it's relevant for establishing that the next step probably involves clearer explanations.
Following links for the claim that AI safety is impossible, I got to the article Galois theory as applied to the hypothesis of AGI terminal extinction risk mitigation, which claims:
It's worth noting up front that this sounds pretty crazy. There are very few examples of anyone saying "It is 100% possible to know that X is 100% impossible" without having a very clear argument (this would be an exaggerated claim even if X was "perpetual motion machines," which is based on a really strong argument built by a normal scientific process over many years). So this is looking pretty cranky right from the top, and hopefully you can sympathize with someone who has that reaction.
The articles goes on to say "The technique of our proof is akin to Galois Theory... There are some aspects of the Godel Theorem in this also." I understand those pieces of machinery, I'm very much in the market for an argument of similar type showing that AGI alignment is impossible. But it's also worth being aware that the world is full of cranks writing sentences that sound just like this. Moreover, this is an incredible project, it would probably be one of the most impressive projects of formalization ever. So the odds are against.
(Also note that it's a red flag to call this kind of informal argument a "proof," not for fundamental reasons but because that's the kind of thing cranks always do.)
The article says "Our interest is:"
I was hoping for a summary but this did not illuminate, it continues to allude to limits without helping articulate what those limits are or how you become confident in them. It goes on:
I admit that I don't know what an IM triple and I'm not going to go looking (having failed to find it by Google or in the article itself) because I don't see how this could possibly help build a sense of how the purported impossibility result is going to go no matter what the definition is. This sounds really crazy.
This doesn't sound true? (And at any rate this is not the kind of argument that makes one confident in things.)
It's clear that AI systems can change their environment in complicated ways and so analyzing the long-term outcome of any real-world decision is hard. But that applies just as well to having a kid as to building an AI, and yet I think there are ways to have a kid that are socially acceptable. I don't think this article is laying out the kind of steps that would distinguish building an AI from having a kid.
These technical results are at best an allegory for the proposed argument, there's no mathematical meat here at all (and note that the problems we care about obviously aren't undecidable, again in order to make sense of this I have to read it as an allegory rather than words I can take literally, but at some point this article probably needs to say something I can take literally).
Though it doesn't affect the correctness of the arguments, I think this shows a lack of self awareness. Right now the state of play is more like the author of this document arguing that "everyone else is wrong," not someone who is working on AI safety.
And that's the end. This was linked to justify the claim that safety is impossible. There are no pointers to somewhere else where the argument is explained in more detail.
I somewhat recently updated away from this stance. The rate is not anomalous so it can't work as evidence. It is not a thing that cranks do, its a thing that people talking about a subject do.
I agree that people often use "proof" to mean "an argument which I expect could be turned into a proof." There is a spectrum from "I'm quite confident" to "I think there's a reasonable chance it will work" (where the latter would usually be called a "proof sketch.")
But this argument is not on that spectrum, it's not even the same kind of object. If you talk to a mathematician or computer scientist you shouldn't call something like this a proof.
(I have much more sympathy for someone saying "Yes this isn't what a mathematician or computer scientist would call a proof, I'm just using language differently from them" than someone saying "Actually this is the same kind of thing that people usually call a proof." Though you lose a lot of credibility if you do that while peppering your writing with references to other theorems and implying a similarity.)
I does feel like isolated demand of rigour. Mathematicians writing to other mathematicians about new results seems like a fair comparison of speech activity and this expresses a similar level of confidence (carefully combed analysis willing to defend but open to being wrong and open to details on questioning).
I don't understand what the two types that would make a type error would be. Both are the one shared by "It can be shown that an angle can not be trisected with compass and ruler". People that are far in inferential distance have some license to remain a bit clouded and not reach full clarity in short sentences. And I think it is perfectly fair to classify someone that you can't make sense of to be a nutjob while that distance remains.