Reflections on Asterisk's AI Fellows Applications

And why we still think you should start a blog

We received over 900 applications for the Asterisk AI Blogging Fellowship. We don’t mean this modestly: We really are flattered. As a magazine that regularly publishes forecasts, we try and practice what we preach, so we’re a little embarrassed to admit that when we each made predictions about how many applications we’d get, only one of us (Avital) even had the correct number in her confidence interval. (Luckily Avital is leading the program.)

We’ve extended invitations to 18 of those 900 (we initially planned for 15, but this was way too difficult), and we’ll be announcing our Fellows next week.

Until then: Sheer numbers prevent us from responding to each person individually, but we wanted to share some of the themes we saw, and explain how and why we made our selections.

We spent a lot of time — some 70-80 hours collectively — reviewing submissions. At least two of us read each application, and all three of us scored a long list of just under 200. There were many more people we would have gladly admitted into the program than we had room to accept.

Above all, we hope you use this as words of encouragement. We received some fantastic pitches (and we’ll be reaching out to some of you who weren’t accepted to see if you’re interested in writing a piece anyway). The entire purpose of this program is to get more people writing in public, and we still want to nudge more of you to do so, fellowship or not.

So here’s who applied:

In total, we received 913 applications. 120 (13%) had PhDs or are in PhD programs. There were a couple dozen tenured professors, many more folks with masters degrees, and at least one person still in high school! Applicants were heavily concentrated in the US (48%) and Europe (36%), but we were encouraged by the 8% who applied from Asia and 5% in Africa. (Note to selves: Do more outreach in Australia?)

Nearly a third of applicants work directly in ML or AI, either on the technical side (we had representation from every major lab, and many small ones) or in research. Our second largest pool worked in policy and governance. None of this really does justice to the breadth of backgrounds, though. We received applications from chemists, sexologists, hedge fund managers, neuroscientists, librarians, returned peace corps volunteers, journalists who cover cults, and even a few folks with — how do we say — sui generis theories of LLMs.

What we selected for

First, we emphasize again: We received many more qualified, suitable, and wonderful applicants than we had the space to admit. If you applied, thank you for taking the time. We’re really honored.

For several dozen of you, non-selection came down to either 1) our desire to have a cohort that varied in terms of expertise and experience, or 2) the difference of half a point on our grading criteria.¹ We’ll be sharing the backgrounds of admitted fellows next week, but here are a few of the common attributes that really excited us.

A unique perspective on some aspect of AI, and an eye to write about it. With 900 applicants, some pitches got repetitive, variations on a What is a transformer? Is AI conscious? How can we regulate AI? theme.

We were most excited by applicants who pitched specific stories on areas of AI that they know well, and that we considered under-covered in the blogging sphere: what AI investment funds are doing in the Middle East, how LLMs are being used in African judicial systems, or why monitoring via chain of thought thinking is likely to fail.

Many of you we considered to have a unique vantage on AI. You described in your backgrounds how you led AI integration in your workplace, or how you were in charge of AI policy for a company or government, and we grew excited as we read those descriptions! But then your pitches were not connected to your actual work. We valued confluence between your story ideas and experience very highly.

Excitement and curiosity. An admittedly fuzzy distinction between the pitches that popped and those that were simply competently good we would best describe as “a story that you want to write” vs. “a story that you think should exist.” These obviously aren’t mutually incompatible, but in our experience, the former produces more engaged writing. It’s something we could sense in the curiosity that a writer expressed. Think: “There’s this thing I’ve been trying to work out about how LLMs solve math problems, and here’s what I think it is,” vs. “No one to my mind has written the perfect 2000 word piece on what a transformer does.” A related failure mode was when we read a pitch that seemed like it was written with the idea of “something Asterisk’s editors would like.” We like to be surprised.

A voice. We think that many aspects of writing are skill issues — hence this fellowship. However, it is an uphill battle to develop a distinctive voice, and a common feature between everyone on our shortlist is that they were able to make themselves stand out despite the brevity of the application. Some people did this via style, others through very crisp articulations of their perspectives, others through having an angle we’ve never heard before, and others through some riskier answers.

A mixture of perspectives. About half of the cohort we selected are aligned with safety research in some form (this is the magazine shaped by the philosophy of Effective Altruism), but we mean it when we say are open to all perspectives. We selected a mixture of writers because we believe that to succeed in blogging for a wide audience, it helps to understand a multitude of perspectives. AI discourse seems to slough its skin every 6 months or so, acquiring deeper nuance and creating new coalitions each time it does. We selected for writers who we think have an eye towards that process, either because they’re privy to the research driving that process, or paying attention to those who are.

What we selected against

We rejected some applicants because they were over-qualified. Writing is never easy, and so we fully appreciate the value of a writing community and accountability system for anyone.² But if you have, for example, published a book, already have a well-subscribed blog, maintain a large Twitter following, or were mentioned as one of the intellectual inspirations that other applicants listed in their ‘Why’ sections (this happened at least three times), we thought we’d probably be less counterfactually helpful to you.

This was a pretty unique applicant pool. At an acceptance rate of less than 2% it took a lot to stand out. The most common reason we gave someone with a highly qualified background — and we are turning down academics with high h-indexes and research leads at large labs here — a low score is that we just weren’t excited about their pitches. Admittedly we can afford to be selective here, but we think this is useful information for those of you with substantial credentials who want to write more in the future: editors care as much if not more about the story than your background.

We received a number of submissions from researchers or established journalists interested in learning and writing more about AI. We chose not to accept these applicants because experience, not writing, is your bottleneck. Our best advice here is simply to start: AI is ubiquitous in the news, but there are still less than a couple dozen journalists on this beat full-time. There are many stories to be told, low-hanging fruit everywhere for those with eyes to see! You, journalist or researcher, do not need Asterisk’s imprimatur to start!

Last, while we do appreciate the irony, we unfortunately decided to turn down all obviously LLM written applications. We don’t mind if you use Claude for the copyediting & clarity assist, if the resulting application comes out really good. But there were some applications that pitched a boilerplate story that seemed eerily close to the pitches we got from GPT when we asked it to give us ideas at the intersection of AI and [your field of choice.]

Some reflections for aspiring writers

One of the most common reasons people cited for applying is that they feel held back by perfectionism, or imposter syndrome. As editors who tend to belabor a sentence, we empathize. Our best advice here is to ignore all your instincts. The internet is big, and in our experience, earnest efforts attract earnest feedback and like-minded collaborators. A blog post is never finished, only published. So this is us giving you permission. Hit submit. Just get it out there! Seriously, do it — many of you told us you have multiple pieces sitting in your drafts folder. What value do they have there? At least let Claude see them. And send them to us too — we’ll read it, even if we don’t have time to respond.

We think more people should write to your strengths. This is really a requirement for starting. The more junior you are, the more you need to narrow your scope: What do you know particularly well that no one else does? Can a CS major just out of undergrad write a transformative piece on how LLMs work, or the definitive blog post on the AI race between the US and China? We’re not saying no, but we think you’ll write a better blog post explaining the niche research project you’re working on.

The impetus of this fellowship was that there are lots of interesting conversations about AI happening offline — at lunch, at parties, on your anon discord server — that deserve to move into public view. The depth of applications we received made this clear. We’re extremely excited to start bringing some of this discussion to you on this page, Asterisk’s new AI vertical. Over the next few months, we’ll be featuring some of the best posts from our fellows here. To read them, subscribe below.

But we mean it when we say you should write the piece without us.

Thanks for reading. Thank you so much for applying. Check back here soon for more.

– Jake, Avital, and Clara

1. The three of us independently assigned everyone on our long list a score of 1 through 4. 24 received a 12, and 51 received an 11 or higher. We discussed all 51 of those in detail, and our final cohort is not comprised only of 12s.

2. In the future, we may explore setting up a Discord for this purpose.