Every Platform Will Be a Recruiting Platform
Featuring 7 generative AI roles
Welcome to the sixteenth edition of Black Box. Last time, I described a human-first approach to consumer investing. This time, I reflect on the new way startups are hiring.
I’ve been working with several generative AI startups over the past month and the one thing they’ve all asked me to help them with is hiring12. My job has always been some version of finding the “right” people — competitors as a consultant, users as an operator, founders as an investor — so I thought this would be familiar territory. I quickly realized I didn’t even know where to start. It turns out that most generative AI talent live in the corners of the internet.
Interestingly, the same is true for web3, the last big wave in tech. And niche communities have long been the standard-bearers of whole categories such as gaming, open-source software, and capital-light hardware (think drones and PC builds). Perhaps this is but a corollary to Chris Dixon’s observation that the next big thing always starts out looking like a toy. In any case, it is worth understanding the nature of these corners and how to source talent from them at scale.
Although I call them corners, the host platforms for these communities are generally well-known. The public ones are Twitter, Reddit, and 4chan; the private ones are Discord, WhatsApp, and Telegram; and the technical ones are GitHub, Hugging Face, and Hacker News. What makes them corners is the fact that they are decentralized, pseudonymous, and non-professional (i.e., not exclusively career-focused like LinkedIn). These characteristics make such communities inaccessible and difficult to navigate for recruiters.
The problem of decentralization is straightforward: Outsiders must either know where the community lives (e.g., specific subreddits) or explicitly be added to the group (e.g., Discord servers). This was the biggest problem for me, and I spent an exorbitant amount of time discovering where to look3. To make matters worse, the identities of users on many of these platforms are obscured by usernames. This means recruiters can evaluate them only on what they have written on-platform, which is limited by the amount of content available and the (unclear) extent of its originality.
And finding candidates is just the beginning. They may not be interested, respond to outreach, or continue the conversation. In fact, I found those in corner communities particularly resistant to job opportunities. For many, going deep into these technologies is a passion project — an escape from their work that would be ruined if it became their profession. Others do it for the status that comes from being a community leader or (better yet) a lone genius, and are thus unmotivated to work for anyone else. Others still are ideologically opposed to keeping their work proprietary. (Unfortunately for companies, this segment also tends to include technically sophisticated talent such as leakers.)
Despite these challenges, companies must meet talent where they are. As I explored the corners, I wondered if there was a way to source from them at scale. My first thought was using scouts who are part of these communities since they’re already part of the right groups and familiar with some of the members. They also have the subject matter expertise to evaluate them and help legitimize the role. But then the question becomes how does one find these scouts? The same issues apply as sourcing candidates directly except the company must also convince the scout to work with them. This is hard because the scout would risk their community status given the resistance to job opportunities described above. Yet the members who make the best scouts are the group founders, mods, and other community leaders. This is a wicked problem.
I suspect there is no general solution to corner recruiting. But I think there is a special case for hiring engineers. The technical platforms are not only public but are also relatively centralized; GitHub and Hugging Face are the repositories of choice for code and machine learning models. A secondary benefit is these repositories generally contain a relatively large amount of timestamped content. This makes it possible to algorithmically evaluate an engineer from on-platform content alone. It also makes it difficult to cheat (e.g., by injecting relevant samples) as the arc of historical code within and across projects acts as proof of development.
An engineer recruiting startup could exploit these qualities by building a masked contrastive model that compares its predicted code to the actual code written by a user. Using RLHF, it should be possible to fine-tune the model to understand whether differences are desirable (i.e., more skillful). (This effectively treats the prediction as “par” and asks how many “strokes” an engineer is over or under.) Then the model just needs to repeat over the corpus of that engineer’s code and apply the “proof of development” check. Thus it should be possible to programmatically source talent from at least the technical corners of the internet.
Actually, all social platforms should be building recruiting products. The next big thing is developing in some niche community right now, and the people who build it will come from those same corners. When everyone is always a potential candidate, then anything anywhere is a potential source of signal. In the end, every platform will be a recruiting platform. ∎
If you’re a founder, where are you finding talent these days? Let’s chat @jwang_18 or reach out on LinkedIn.
Lex is looking for a founding React engineer with experience in YSJ and ProseMirror and a founding backend engineer with experience in fine-tuning LLMs. Candidates should have 3–5 years of experience, ideally at a startup. Send me a note at jiarui[at]lex[dot]page if this is you!
Spellbrush (YC W18) is looking for an LLM hacker with experience fine-tuning chatbots and loves anime (seriously), a mobile game producer, a digital artist specializing in isometric pixel sprites, a technical recruiter specializing in AI/ML talent, and an EA with experience working with a startup founder or VC. Reach out to jobs[at]spellbrush[dot]com!
Some hacks I developed include finding thought leaders on Twitter and going through the accounts they follow, checking the organizations that a star Hugging Face user belongs to, and searching for subreddits with Discords (the invite is often in the description or a pinned post).