Don’t sell yourself short
And other advice for the newly AI-concerned.
Back in 2023, I spent several months as an AI safety lurker. In my quest to ascertain if and when AI might destroy the world, I was spending several hours a day quietly observing a discussion that I was both terrified and confused by. I became familiar with the cast of characters that populated this online world, an ensemble of profile pictures and usernames espousing confident but conflicting opinions, whom I considered the People Who Know Things.
Determining precisely who ought to be counted among the People Who Know Things was its own challenge. Should I restrict membership to academics whose Twitter bios were decorated with qualifications from prestigious universities and whose work had been cited thousands of times? Or insiders at frontier labs? What of those whose online presence didn’t appear to signal any relevant formal qualifications or experience, but whose long and complicated-sounding LessWrong posts were being seriously engaged with by people in categories 1 and 2? The one thing I did know is that I was not and never would be a member. The People Who Know Things occupied a world on the other side of an unbridgeable divide.
One day, the barrage of conflicting apocalypse forecasts I had been subjecting myself to for nearly half a year was taking its emotional toll. After a few glasses of wine, I took to Twitter to complain about it. I wrote a long, earnest thread whining about how I was just so confused, because Expert A said we were all going to die in 2027 and Expert B said 2035 and Expert C said never, and all their reasoning was totally opaque to me because I was Just A Girl and didn't know what a FLOP was or how scaling laws worked and I was simply not smart enough to figure any of it out.
When I woke up the next day, I unlocked my phone to find that my thread was gaining traction. The People Who Know Things were starting to talk back to me. Responses were varied, from people who related to my plight to white knights who wanted to swoop in and save me from the AI-doom cult. But one in particular stuck out:
That evening, my housemate found me frowning at my phone while polishing off the remainder of the wine.
“Are you ok?”, she asked.
“I called myself stupid on the internet and people believed me”, I muttered.
“You should spend less time on Twitter”.
Can’t and won’t understand the arguments, I tutted indignantly to myself. Can’t master the object level. Of course, my annoyance was totally unfounded, since I had said exactly that! This interaction, and several others that I’d come to have over the subsequent months, taught me a valuable life lesson. The culture I’d grown up in, where every conversation took place against a tacit backdrop of false humility, is not representative of the world at large. At my North London girls’ school, self-belittlement was invariably met with a chorus of “omg you’re so smart and pretty and totally not fat at all!!”. But in the AI safety sphere, the self-belittling are assumed to be doing a well-calibrated assessment of their own attributes1.
Was I actually being falsely humble? I think the answer to this question is somewhat complicated. The emotions I expressed in the thread – of confusion, overwhelm and uncertainty about who I should be deferring to – were genuine. But did I consider myself incapable of understanding object-level arguments for and against AI risk? Not really. I believed then and believe even more strongly now that the high-level reasons for being worried about powerful AI are extremely simple. It’s not hard to understand why racing to build increasingly capable and goal-directed AIs – without any scientific solution to the problem of controlling them once they supersede us – may lead to bad things. It’s also not difficult to observe that the unworried are few and far between, and that they lack anything like a satisfying answer to this basic concern.
Of course, there are many more in-the-weeds questions that one could ask in order to form a well-rounded AI worldview (will takeoff be fast or slow, what is the likelihood that X or Y alignment technique will work, what will the political response to advancing capabilities be), and it is true that in 2023 I’d have had very little to say on any of these. But it is also true that I hadn’t really tried. My “research” into AI safety up until that point could be more accurately described as an obsessive quest for reassurance, which ended up culminating in an extensive collection of p(doom) estimates and timeline predictions. These had been cherry-picked from a much richer discourse about why people had reached those conclusions, which I had only superficially engaged with. I just wanted to know if the world was going to end. I was more guilty of being lazy than stupid. Over a year later, I’m still far from confident in any of my AI takes, but I have at least tried to do the intellectual legwork of refining them. I now consider myself an Aspiring Person Who Knows Things.
I don’t know whether there are many other AI safety lurkers out there who don’t feel qualified to join the conversation, let alone whether any of them are reading this. But on the off chance that they are (especially if any are female, since being one of the few women in AI safety can create the perfect storm of self-doubt), I wanted to offer some unsolicited advice.
#1 Don’t sell yourself short
This is of course generic advice that applies to the world at large – if you repeatedly downplay your own abilities, people will not always give you the benefit of assuming you’re just insecure or self-deprecating. They might take you at your word!
This is somewhat complicated by the fact I think my online AI ramblings have filled an underrepresented niche – that is “normal person who doesn’t understand AI but is stressed about it”. My femaleness complements this quite nicely. It has been easy to lean into the Just A Girl aesthetic. There’s endless mileage you can get out of takes that boil down to “wow guys it sure looks like the experts are spooked by AI and think it might literally kill everyone, maybe governments should be taking that more seriously but hey what do I know haha”. There have been benefits to doing this, but I think that to a degree, it has resulted in me letting myself off the hook. I believe it has meaningfully delayed my transition to an Aspiring Person Who Knows Things. If I could go back in time, I’d probably skip my Just A Girl arc.
#2 You can just say things
If you’re an AI safety lurker and don’t want to be – just say things. It really is that simple. Despite interacting with 100s of people who are smarter than me on topics I would once have considered beyond my intellectual bandwidth, I have suffered very few embarrassments. Even if you say something wrong, you’ll likely be politely corrected, not socially punished. You’ll probably find that The People Who Know Things often agree with you.
This isn’t to say that everyone’s contributions to the conversation have equal weight, or that AI safety is some unique field where we shouldn’t privilege expertise. I don’t want to sound like one of those people who thinks a layperson with acess to PubMed is as well-qualified to perform medical diagnosis as a doctor. But AI safety is a domain in which one can have knowledge or expertise along many different axes. For example, technical experts may be best placed to forecast the speed at which AI capabilities will improve, but not how governments ought to respond to them. The field is also nascent enough that, for better or worse, it hasn’t had time to become subject to the institutional gatekeeping that befalls many others. There are conversations at the frontier of AI safety and policy literally playing out on Twitter and Substack. There’s nothing to stop you joining them.
As was the source of my frustration in that original tweet thread, there is no expert consensus on what a future with powerful AI will look like. It seems like a safe bet that the future will be very weird, but beyond that, I see little grounds for confidence. I think this is the perfect context in which to Just Say Things. I like the way Nathan Labenz put it in a recent episode of The Cognitive Revolution podcast:
AI gods might be an emerging trend over the second half of the decade. I have no idea how we're gonna relate to these things. If they are meaningfully superhuman, will we even try to keep them under control? Will we worship them? No matter how weird your alignment idea is, I think it is worth pursuing. No matter how weird your thoughts are about where the future might be going, I would say they're probably worth entertaining.
#3 Ask questions
In trying to follow developments in AI, I find myself frequently confused. Anyone who isn’t delusional will too. Luckily, there is a whole community of well-intentioned and knowledgeable people who are happy to weigh in! I have on innumerable occasions contacted people I know (or don’t know!) with technical or policy expertise to ask them clarifying questions. More often than not, they have responded. I also often pose my open questions to Twitter. What’s the deal with o1 sometimes doing better on PhD-level science questions than high school level ones? Why do some people believe that an ASI in the hands of governments would pose an unacceptable risk of power concentration, but one in the hands of a private company wouldn’t? I have received many helpful answers.
#4 Actually try
As I said above, 2023 Sarah, who pleaded ignorance of all things AI in that long rambly thread, hadn’t actually tried very hard. In 2025, I still haven’t read or understood nearly as much as I would like, but I am sure as hell trying. The downsides of ever-accelerating AI capabilities are many (plausible short-term human extinction, job loss, a looming crisis of meaning…) but at least one major upside is that it has never been easier to learn things! A hill I will die on is that using LLMs to translate smart-person speak into dumb-person speak is, in fact, a smart-person move. This is an excellent use case for making sense of jargony papers or impenetrable posts on LessWrong. You can then have as long a conversation as you want with a private, infinitely patient AI tutor about anything you still don’t understand. Still, there will be things that remain the exclusive purview of CS PhDs or seasoned policy wonks. There is a level of expertise that you or I likely don’t have time to build up pre-singularity. But that doesn’t mean we can’t make a start.
In the early days of my confused-girlie-come-AI-worrier journey, multiple people reached out to me to ask if I’d considered working in AI safety. I dismissed them out of hand. Of course I hadn’t! That was the territory of People Who Know Things, of which I obviously was not one. My plan was simply to post through it.
I now do, in fact, work full-time in AI safety. I still consider myself to be at the bottom of a very steep learning curve, but I’m not Just A Girl anymore.
This is obviously an over-generalisation! I think this is more true among AI safety people than in other communities / cultures I’ve encountered.
Hell yeah
Oh wow, you work in safety now? That's great! What are you up to?