In my occasional advising calls with aspiring AI Safety folks, one of the most common questions I get is “What courses should I take next?” I often find myself replying: “None; go do stuff instead.”
Fabricando fit faber. By making, one becomes a maker.
There are a lot of courses in AI safety and governance. I’ve helped teach a few. Some are quite good! But after the tenth or twentieth person tells me “I’ve read If Anyone Builds It, Everyone Dies and taken BlueDot’s AGI Strategy and the CAIS AI Safety course, I’m not sure whether I should do ARENA or apply to MATS…” I start to notice a pattern. I want to take them by the shoulders and say “THOSE ARE NOT YOUR ONLY CHOICES.”
Look. MATS is pretty cool. Many of their graduates do good work. It’s also, AFAIK, swamped with promising applicants and eager young ML engineers.
If you think MATS sounds like a good fit for you, sure, you should probably apply. And then, while your keyboard is still cooling from the red-hot fire of your earnest expression of interest, you should turn that incandescent determination towards something you can do right now.
First of all, have you contacted your representative to tell them your thoughts on AI? No? Go do that!
Second, have you looked at other action pages for inspiration?
Third, have you considered Just Doing The Thing?
You know, the Thing. The Thing you’ve been thinking about doing, the Thing that leapt out of your mind as an important step while you were worrying about AI, the Thing that perhaps you’ve been putting off until you “have learned more” or “feel more ready”.
Verily I say unto thee: Thou’rt probably ready.
Here’s a good example of a Thing: How can OSINT be used for the enforcement of the EU AI Act?
This policy memo explores the potential of Open-Source Intelligence (OSINT) for the enforcement of the EU AI Act. Comparing the monitoring of AI development to the existing OSINT efforts in nuclear nonproliferation, this memo advocates for civil society organisations supporting compliance with the EU AI Act through OSINT. Despite its limitations, OSINT is a useful method to support the regulation of especially high-risk systems and it should become a part of the AI policy enforcement toolbox…
This short post was written as a final project for an AI safety course. It took a decent chunk of time—about a month of part-time work and some unknown-to-me number of hours on the part of the author. It makes a straightforward point about a narrow subject: Open-source intelligence can help enforce AI policy. It’s niche, it’s not perfect, but it exists. The author did some research, made a case, cited a few sources, and moved on.
If you’re reading this, you probably can too. There are plenty of open problems in policy or technical governance or, if you’re focused on machine learning, interpretability.
Pick a topic that seems you-shaped. Set a deadline measured in weeks, not months. Make something exist.
You do not need twenty weeks of online courses or a Ph.D. in machine learning to become an Officially Licensed Person Whose Opinions Matter. You can just make things.
They will not necessarily be good things, right away. But the act of making a Thing will make you more the sort of person who makes that sort of Thing. And you can point to the Thing later, when you are applying for MATS or CAIS or UKAISI or wherever, and say, “I did this Thing on my own initiative, because I care. I want to do more Things like this, but with you [for money].”
Go forth and do the Thing.
Although I very much agree with 'go build a thing' motto, I think you risk sounding a lot like a self-help author, if you don't actually address the multiple elephants that are in the room. I list a few of them for you, just as an example:
- Building a thing, whether it be a research paper, a piece of code or anything else, requires some resources. At minimum, the person who is going to build the thing needs to be able to dedicate time to it. In some cases other resources are also needed. Dedicating time usually means lack of time or resources in other areas. Your assumption here is that people are generally privileged.
- I have worked for 22 years in the Tech industry. I have worked for governments, NGOs, commercial SMBs, academic research centres and so on. The organisations that you mention -- plus a few others that you did not include here -- not only control the narrative, but also the funding and in general what counts as an opportunity and who gets to benefit from those opportunities. And group theory in psychology tell us that groups of people will tend to maintain status quo, if they are the beneficiary of it -- ask me why no one gets to do anything other than LLMs these days or why theoretical alignment is virtually dead. This translates to elitism, gate-keeping, or at best just purely biased and unfair behaviour.
- Since said organisations do essentially call the shots, any meaningful amount of work to materialise has to go through their filters, comply with their beliefs, and do not go against their narratives -- ideally, strengthen it, but in some cases they may let you get away with being a bit unorthodox. This means even if you manage to generate any meaningful amount of work on your own, it is going to be dismissed, at best. Why? Think about like this: what would happen if people who didn't have a PhD (as you mentioned in your post as an optional qualification), wrote papers, published them in prestigious journals, got air time in TV, advocated politicians and so on. The very direct side-effect of that is the institution that issues the said PhDs from free funds handed over to them via tax-payer money, now being less relevant, and, of course, that's not ideal because it threatens many people's position and income. This is just an example. I am sure you can use your imagination to see how this could apply to other organisations.
So, although I can agree with "You can just make things." to an extent -- because you are assuming the person who is committing to the work already has the resources needed for producing the work e.g. everyone must be at least as privileged as I am; I highly disagree with "You do not need twenty weeks of online courses or a Ph.D. in machine learning to become an Officially Licensed Person Whose Opinions Matter."
This claim is surprising to me because you have been close enough to this circle to help teaching some of the courses, but you didn't notice that unless you get a 'badge' from people who control the narrative, your opinion doesn't matter. This isn't anything new. It has been the case in many, many fields from academia to foreign policy, and for centuries. It's just your framing of it that makes me react to it, otherwise, it is not anything new or surprising, or frankly worth discussing in most situations.
Regardless, I hope I am wrong and you are right and people without certain privileges get to just do things and those things end up making a difference, even though they don't get the seal of approval from the people and organisations who control the narrative, funding, and opportunities of an entire field. Mind you, none of those people and/or organisation were elected democratically. It's just privilege leading to more privilege.
If that is not the case and somehow I have been misreading the whole society for two decades now, that would be actually a great news to me and I would personally celebrate it, but until then I will have to take your claims as self-help material, at best. I don't think your goal is to promote a certain agenda, even though I don't have anything to back it up.
If you think I am wrong, please do educate me by all means. I am here to learn.
[Edited for fixing some typos.]