Lateral Workshop Applications: A Case Study

LessWrong，LessWrong 是一个专注于人类理性思考的社区博客，讨论认知偏差、决策理论、AI 安全与有效利他主义。关注

We’re running a three-day program for experienced professionals moving into AI safety. See our announcement post for more details.

Lateral Workshop received over 700 applications for its first cohort. We were happily surprised by the range and quality of experiences, and decided to expand the workshop to accept more applicants. Inevitably, the majority of applications were rejected, and I (Jacob) wanted to put down some notes for people considering next steps.

First off, there were some top candidates who we wanted to accept but didn’t have capacity for. We agonized over a few that we saw promise in but had to prioritize participants who we thought that we could provide more value to. In addition, we fully admit that our review can never be perfect, despite involving domain experts throughout the process. There are surely a handful of people who were rejected not because of their background but because we misread their presentation of it throughout their application responses and resume.

I’ll focus on the most tractable fixes to common patterns we saw. Applicants who portrayed clear competency in their career stood out, but it’s possible to make up for a lack of experience with deep thoughtfulness and working on building your field-level understanding.

  • Understand the priorities of the AI safety field. We decided to interview nearly all of the applicants who articulated field priorities cleanly. There are certain concerns that fall within what the ecosystem is presently built to support; many valid risks are better suited to work in adjacent fields. Good places to start include LessWrong, the 80,000 Hours podcast, or AI Futures Project’s scenarios.
  • Look into neglected risks. Current risks typically have a greater mass of corporate and governmental energy dedicated to them, whereas future risks are often neglected even in cases where early solutions are tractable. Learn how to reason about risks in a scope-sensitive manner.
  • But be authentic. There were applicants who gave us answers that felt like they were trying to guess at the “correct” answers. Ultimately we care about your thoughtfulness, not that you memorized the correct language to use.
  • Gain knowledge of AI safety organizations and their research directions or policy proposals. Normally this type of information is hard to glean from websites; reading a few reports authored by researchers at the organization or their personal Substack or LessWrong posts is a step in the right direction.
  • Include any costly signals of commitment to a career transition and clear-eyed thinking around next steps. If applicants put together a Saturday reading group or took time off work for self study, we want to know! We also appreciate people who had thought through how Lateral in particular would help them reach their goals, as opposed to other existing fellowships.
  • If asked to discuss your strengths, don’t spend too long rehashing your resume. People sometimes don’t realize that the skills important to a small, mission-driven field are quite distinct from the skills that got them to their current role. Curious and open-minded responses here fare well.
  • Write in your own voice. LLM-produced applications consistently underperformed (LLM usage was discouraged but not banned). The issue with LLM-generated text is that it makes you sound the same as everyone else, not only in tone but also reasoning style. Low quality applications can be improved marginally, but they are still so far below the bar that this doesn’t matter. Higher quality applications that would have had a chance to pass on to an interview are pushed down towards the mean, and increase the risk that reviewers miss your insights.
  • Ensure your resume is well formatted. Despite my best efforts to look past the format, when readability suffers, it makes it harder to advance the applicant.
  • Stay away from fluffy language. Transparent reasoning was rated higher. Clear explanations of the causal chain for the risks applicants care about fared better than gesturing in a broad direction.

Out of these points, the lion’s share of the misses were from people who hadn’t internalized the priorities of the field. My guess for why this is the case is that one learns this through diffusion, by parsing a variety of sources, although I don’t think that it would take an inordinate amount of time. See Matt Beard’s post for a clear roadmap.

Here’s a list of resources to look at if you haven’t considered them:

Thank you to Dillon Nguyen, Spencer Kitts, James Lester, and Eliana Du for feedback on a draft of this post.



Discuss

添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论