I'm at School for Advanced International Studies here in Washington DC. We're primarily policy people, but I'm much more than my peers focused on AI, and strictly AI now.
For policy folks to know and have a clue about what AI is, I strongly insist that they need to start with linear algebra, should work through some perceptron math, and read the deepmind 'attention is all you need'.
This is linked to ai safety I would argue because obviously I am aligned with not doing crazy things with clankers. I'm wondering if i can get some club advice.
-------
I want to bring frontier AI comprehension, frontier AI knowledge to my policy classmates.
I'm primarily a researcher however and I just don't know and can't do clubs. I don't know why. I can't say or do something that makes sense. I just don't get clubs, or something.
I've asked people who have been interested by my initial pitch for thoughts or advice but I get this response that isn't really specific or exact? And I think it reveals either i'm doing something a bit off or I don't really know how to handle this situation too well.
- I've tried running an 'all-hands on jev' day. This was badly timed and only one person showed up. Arguably it was mostly a timing issue.
- I've sent this message out (paraphrased)
"[hope to] least make it possible for students to learn and obtain hands-on comprehension of ai for no reason other than that if you can't do linear algebra, you can't understand how clankers do their clanking things or essentially what's real and what isnt. This is quite critical to anyone who wants to do any type of AI or tech policy.
Other than that, is to drive serious knowledge and research and to fixate on the mysteries currently in the field. The technical mysteries are things like, will clankers scale to artif. superintelligence? Those are out of reach. But the mysteries that are a fit for our field are ones like is Jacob Coxton telling the truth or not, or at least, what is motivating him? What motivates Sam Altman and who is he (similarly, intelligence analysis also analyzes leaders). Given that Sam Altman now has permission to commit felonies without risk against the United States - he could be argued to have more power than your average nation-state. His motivations and background is within or adjacent to our field in the social sciences.
There are plenty of people around who do econ and business analysis (also critical to understanding the AI world), and international relations is a good way to assess national-level competition; this is the ideal place to bring all that together."
Not much of a reaction to #2
I've also made the argument that policy students are fundamentally disadvantaged when engaging in tech policy because of their lack of technical comprehension and we need to learn the hard stuff in the technology, even if it's survival-level linear algebra to understand it. For some reason this pitch worked when I said it to people.
Anyway I'm thinking of putting up a poster soon that's more about
"XYZ events happened.
Do these mysteries interest you?
Are you into frontier AI research? Does mechanistic interpretability fascinate you?
(Maybe see if people have interest in Jane Street's AI-related challenges"
It would have a bit of a very fixated and tailored audience, sort of like how those jane street questions are catered towards a narrow group of people.
I've had people say it sounds like I don't know what I want to research in response to #2. That's just wrong . There's too much to research alone. I don't see why policy folks wouldn't be interested in what I might term 'frontier AI research' here.
It doesn't matter what question it is.
Let me illustrate with an example. If we're talking about 'will models scale', this very simple question starts with how you know where their existing performance is at. This goes to the benchmarks. Well, the benchmarks weren't written on a stone tablet by god's hand. Not only that, they aren't trustworthy. Therefore the inquisitive researcher has no choice but to understand how the benchmarks were built and what they actually represent. You can make conclusions about scaling after that.
Or the recent OAI stuff where they move on to stealing math. On the same page of scientific american you can get almost entirely contradicting stories, right now, about the capabilities of the models and what happened.
This is precisely where pristine epistemics and primary source investigation benefits us and gives me justification for my research.
I'm sad I can't get more work going here with this 'club'. Maybe this isn't what I would need but what I would tell you is what I need is just someone who knows how to do the club part of things. I can just do the non-managerial parts of things.
I could 'teach' people things I know about AI at this school in its tech club, but my time's worth is not defined by me spending time rambling on about things. It's about collaborative research and cross-field knowledge.
I would consider me spending my time in the rest of the school year teaching linear algebra by math once a week a severe downgrade even if it guaranteed me work to do in the official club.
-- I'm interested in people who say AI is a 'footnote' in the economy right now. Okay, how is that measured. Will it remain a footnote? It's a footnote even though many people are using it? How long do you think that holds up? Etc. I would like to know about these perspectives but even if it's true that AI is a footnote what makes it more important is thinking about the basis of that statement, thinking about the velocity and acceleration of it.
Anyway, not even sure why i'm in this position and I don't know why it's so hard to just pitch straight research. I can't imagine any tech-interested student here not believing that this is core to their comprehension of the world right now.
Anyway thanks much to the commentators if any can offer any help. This is really about me preparing my AI policy colleagues with the skills they (partially) don't know how to obtain
