3か月でAI安全分野に転身する方法（キャリアアドバイザーMatt Beardによる記事朗読）

Hi listeners, my name is Matt. I'm a former advisor and writer here at 80,000 Hours. This is a narration of my article, How to Get into AI Safety in Three Months. You can find the full piece, including the many links that I mention, plus a few memes, on the 80,000 Hours Hubstack. As an advisor, I've spoken with hundreds of people who want to use their careers to mitigate AI risks, but don't know where to start. While the most efficient route to enter depends on your background, there's some foundational advice for breaking into AI safety that I find myself repeating on almost every advising call. These steps are hard, but they're not particularly complicated. Almost everyone I've advised who found work in AI safety followed them, and most experts or hiring managers that I've spoken with have endorsed similar ideas. If you're efficient and driven, I think it's possible to pivot within three months. So consider this a crash course for breaking into AI safety. It covers many steps, like how to build foundational knowledge, identify your comparative advantages, test your fit, develop side projects, tighten your feedback loops, network in AI safety, start applying. 1. Build your foundational knowledge and AI safety context. There are a

couple dozen hours of foundational reading that I think everyone trying to enter the field should do, but it might be more approachable than you think. If you take 30 minutes each day to read while drinking your morning coffee, listen to podcasts and article narrations on your commute, or make a reading group with your friends, you can build up your knowledge quickly. When my advisees have enthusiastically jumped into the AI safety rabbit hole, I've seen them get a better sense of where to focus and how to improve their fellowship or job applications. What should you read or listen to? Our AI Risk Essential Reading List and Problem Profiles are great places to start. Then, you might take a Blue Dot course or read foundational papers like the Alignment Problem from a Deep Learning Perspective. Once you have a sense of what you want to specialize in, which the next section will cover, you can also review more specific resources like our collections for technical work, policy work, etc. Your goal should be to reach the point where you can have a 30-minute conversation at an AI safety conference, and the other person would come away thinking that you understand the core problems and a few potential solutions. Research-heavy roles will require more specialized expertise, but for example, most jobs won't require you to implement

the transformer architecture. It's often enough to know what the transformer is and why it's significant. Specifically, you should have an answer for how fast you think AI progress might be over the next few years, which problems you think are most pressing based on this, and a few ideas for which organizations' theories of change you think make the most sense. Your views might evolve over time, but to ensure you're efficient with your job search, you should know some organizations might expect to hear different kinds of answers from competitive applicants. You might think, who am I to form opinions about this? I'm not an expert. And while I understand that concern, I'd say, you don't have to be certain, but you should have something to say. It's fine to keep your views loosely held and to defer to experts. Deferl is a skill, after all. But you should take ownership over forming your own views. Hiring managers will want to hear them, and they'll be interested in your reasoning for how you got there. Some people refer to all of this as developing your AI safety context. The EA forum posts why experienced professionals fail to land high-impact roles explains why some otherwise promising candidates can get rejected too early.

Some don't communicate that they understand AI safety issues and an organization's approach to solving them. In other words, they haven't developed sufficient context about the field. In other industries, this isn't as necessary. But in most jobs tackling catastrophic risk from AI, organizations want to find candidates who share their mission and are willing to think carefully about problems in a scope-sensitive way. This can also include non-research jobs. These organizations aim to solve big problems that many other people haven't quite grasped as possibilities yet, so it's important that you're eager to think about them directly. Of course, you don't have to agree with everything that a hiring manager or an organization thinks. Thoughtful pushback can show that you've developed your own views and aren't just parroting the right phrases. It's helpful to show your reasoning. 2. Identify the in-demand skills you have or could build. And in the article here, there's a picture of three ladders. They're quite precariously perched, and it says, with a caption, It's fine if your career looks like this for a while. You should build a ladder of cheap tests to discover what you're good at. To solve problems, you should identify and build skills.

But which ones? It's not always obvious from the outside which skills are most important in AI safety, especially because there are so many options for how to contribute. Some of the broad categories include and these are all links to career reviews on the 80 Hours website technical AI safety AI policy strategic operations scaling field building communicating ideas and various others like AI bio macro politics etc You should read a few of these career reviews and reflect on how your own skills and experiences might fit one of these paths. You can also chat about this with our free AI Career Chatbot online. You can also look at job postings for roles where, even if you aren't ready to apply yet, you can get a sense of which skills and experience these organizations are looking for. If there are experts you admire, you could read about their own career paths or listen to interviews that they've given. Similarly, one of the fastest ways to learn about a field is to speak directly with the people working in it. A later section of this article covers strategies for how to do this, or you can see the 80,000 hours article about networking.

But ultimately, what you'd be good at and could quickly learn is an empirical question. Self-reflection and reading are a good start, but soon you'll want to run some experiments. At first, these could be short, time-valid experiments that you could do over a weekend or a couple evenings. You can test your fit with small projects to see what you enjoy and could get better at. This might involve taking a free technical AI safety course, writing a blog post about a think tank report, building a data dashboard that someone else might find useful, or hosting your own event part-time in your city. If you enjoy your experiment, that's useful data. But if it puts you to sleep or makes you feel miserable, that's also important to know. you're seeking to efficiently find out which direction you should be pursuing. Most people should start by expanding the number of options they're considering, but then narrow them down through these kinds of experiments and outside feedback. I'd also encourage you to apply for advising with 80,000 Hours to get more tailored advice. Three, apply for fellowships and jobs. Compared to other industries, AI safety has an unusually high number of incubators,

fellowship programs, and resources to help you find jobs, including our own job board. The AI safety field has grown by about 3,000% in the last eight years, driven by the urgency of these problems and a large talent development ecosystem. 80,000 Hours has a list of technical AI safety fellowships linked in this article, and AISafety.com lists many resources as well. In AI Policy, we have a list of fellowships here, and EmergingTechPolicy.org has a broader list on their guide and database. Fellowships can be a quick way to get feedback, work on concrete projects, and get referrals for new potential opportunities. Outside of research, there are fewer fellowships for generalists who can get things done, manage teams or projects, and unblock bottlenecks for organizations. If that skill set describes you, you should try to learn from experience as soon as possible by applying for full-time operations jobs and, again, developing side projects. Even if you're unsure how competitive you are, it can often be worth your time and expectation to apply for a few jobs directly. In my time as an advisor, I've seen many people underwrite themselves.

Applications can also provide useful information about your skills and teach you new things through interview prep and the work tests that you do along the way. You can set up alerts on the 80,000 Hours Job Board for particular kinds of roles to be proactively emailed to you. Some positions are only open for a couple weeks before they close, and I've seen competitive candidates accidentally miss a deadline. Also, if you do have a particularly relevant experience, you could consider submitting an expression of interest to one of our top organizations. I know this can be a lot of work and you only have so many hours of free time in a week. Job searching is really hard. So you might divide your applications into a few categories. Maybe some percentage for your dream job applications where you really want to learn from the process and aim high, but then some other percentage for your core focus area. And depending on your situation, as time progresses, you could even allocate some percentage of your applications to safer opportunities for career capital building jobs that allow you to develop specific skills for a while outside of AI safety. and not just rolls on our job board. Four, networking is cliche, but seriously,

you should go talk to people. And there's an image here of a neural network with the caption, networking can seem messy and inefficient, but it produces surprisingly good results. People often hate networking. It can feel transactional, inauthentic, or like the only way to get a job is to already know somebody. I really empathize with this, but I think a better way of framing it is to think you should treat your networking like a research project and you should try to remember that talent markets are shockingly inefficient. It's just extremely time consuming for an organization to assess thousands of resumes and even after an interview or work test it's not always clear whether a candidate is a good fit. So a referral from a trusted source can save an organization a lot of time at their first application review stage Even something like oh we had a good chat at an event can provide some useful signal because it can show that you thoughtful and you taking the time to engage with the community However this should not be seen as necessary for finding impactful work Many people in AI safety are hired from cold applications, and almost all organizations have application processes that include multiple stages,

including anonymous work tests, to assess candidates independently. But speaking with people in the field is also one of the fastest ways to learn. I really think you might get more tacit knowledge and feedback from a 30-minute coffee chat than you could get with a month of self-study alone. Networking can also help you quickly test your own fit. Could you see yourself doing the kind of work that this person you're speaking with does? What would it take to get there? 80,000 Hours tries to distill this expert advice into our career reviews and our podcasts, but when possible, I think you should go to the source. How can you start speaking with AI safety professionals? This is easier said than done, but the 80,000 Hours Networking Guide has some tips. You could attend a local AI safety, rationalist, or effective altruist meetup, go to talks or events in person or online, start participating in online forums or on Twitter, get coffee with friends of friends, or use well-designed cold emails. I think that succinct cold emails that engage with someone's work are often underrated. It can be really worth your time. I also recommend taking a trip to a hub city like San Francisco, London, or DC,

or even moving if you can. Multiple people who successfully found work in AI safety have told me that moving to where the action is was their most important decision. There's no substitute for being surrounded by the tacit knowledge and expertise that you can get in a specific community. But of course, you can also get this from visiting at conferences like Effective Altruism Global. Five, if you aren't breaking through to jobs or fellowships yet, do deliberate practice and develop a portfolio of side projects. Movies skip over training sequences with a montage because it's a boring grind to get good at anything. but in real life, it takes time. When working through AI safety courses or projects, you should do some deliberate practice, meaning that you set specific goals, work at the edge of your comfort zone, and build in concrete feedback loops. This is a habit that can take some time to cultivate, but it does pay off. It helps to build routines to do this every week if you can, even if you only have a couple hours. This might take the form of a structured course, where feedback loops from the curriculum and an instructor are already built in, but it also might mean developing your own.

AI safety is a relatively new and fast-moving field, so it's often better to learn by doing. AI models themselves are also increasingly good at helping you design your learning, find new ideas, and get outside feedback. Which side project you choose will depend a lot on your background and what topics interest you, but there's no shortage of options. Our technical and policy resource lists include projects, and this what-are-some-projects-I-can-try page, it's linked in the article, includes many others. You don't need to have a novel contribution early on. It's more important that you're working your way at something to just build up those skills. Of course, side projects are not the silver bullet. If you don't have much relevant direct career experience for the kinds of roles you're targeting, again, you might need to focus on finding some intermediate job for a while that helps you build skills for about a year first. For example, you could try to find jobs outside AI safety in software engineering, political staffing, communications, operations, advocacy, etc. Six, work in public and tighten your feedback loops.

There's a screenshot of a tweet here. The tweet reads, wait, sorry, did he say that seems reasonable or did he say that might be reasonable? What's the difference? Oh, God, just scrap the idea. Just scrap it. It's over. Just forget it. Just delete the doc. Just, just. And I've written the caption here. Getting feedback is scary, but essential. In our survey of almost 40 AI safety organizations, we often heard that applicants should publish their outputs. Quote, respondents strongly emphasize the importance of being public with your work and personal journey. Show projects, write blogs, posts and comments, make videos, build and share useful things, or just build something using AI to show off. Why? Employers need to know that you can do the work, and that you're proactive, energetic, and reasonable. Doing work in public is one way to give them that information. Hiring managers just can't see inside your head. They sometimes only have a few minutes to review your job application at the first stage. You can differentiate yourself with job experience, of course, but you can also showcase your thinking and personal agency by having these concrete projects. Oh, she wrote a response to one of our papers,

won't guarantee that you get a job, but it'll be noticeable that you're aware of an organization and are proactively engaging with their work Working in public is also important because it helps you tighten your feedback loops You might save months of effort by getting feedback on a project early and you might learn a lot faster that way People working in AI safety spend a lot of time online They might respond to your questions on LessWrong, the EA forum, Twitter, Substack, if you engage there in good faith and, briefly, post something of your own. It can really sting if you do get negative feedback on your public work. I know how scary this can be when I put up writing myself. But negative feedback can often still be really useful data. Either you haven't built the requisite skills yet, which constructive feedback can actually help you with, or you just don't know yet how to frame your outputs in a way that others will understand. Again, I know this is scary, but you'd rather know that information early so that you can learn more, change how you communicate, or pivot to some other subfield within AI safety that you might be better at. 7. Don't wait for permission to do the kinds of work that you want to be doing.

For years, AI safety has been building the plane as we fly it. While the field is more mature today, with established organizations and research agendas, it's still moving fast and has many urgent gaps that will need to be filled. If you can find a way to be helpful and solve some problem, you might get noticed. While it won't guarantee you a full-time job, I've seen relatively new researchers get noticed by publishing interesting writing online or by having a thoughtful Twitter presence. Even at 80,000 Hours, we recently hired someone who we discovered because of his AI safety talent network side project. Luca noticed a gap within AI safety, then built an online career platform that had a login, automated opportunity matching, support for hiring managers, etc. When we at 8,000 Hours started building similar tools to improve our job board, someone in our network recommended that we speak with Luca. We were excited by his experience and invited him to a multi-stage job assessment, then we made him an offer to join our team full-time after he did well in those. Not all hiring works this way, of course, but as a general principle, I think it's valuable to start doing the kinds of work you want to be doing,

even if it's part-time, before you get a job offer. You'll learn a lot, and it'll be easier for people to notice you. Your goal isn't just to get a job, of course. It's to build the skills to solve the problems that you care about solving, and this is one way to do it if you have extra time. Caveats. Imposter syndrome, burnout, and acknowledging how hard this all is. Many people considering an AI safety career are held back by imposter syndrome. I've had advisees tell me they think they have nothing to contribute to this field, only to land a job in AI safety a few months later. When you're comparing yourself to others, I think you should make sure that your comparisons are, one, time-sensitive. So, are you sad that you're not as successful as someone who's five years older than you, or who's been at this for three times as long as you? Why would that be the case? Two, in the right reference class, are you comparing yourself only to highly visible top performers? The right reference class for someone who could do great work at the current margin is much wider than that. And three, empirical. Have you actually tested if you could get good at this skill?

For how many hours? With how many different approaches and how much feedback have you gotten? Similarly, burnout is real and it's terrible. AI safety has many urgent problems that we need to solve, but I do want your work to be sustainable. 30 minutes a day sustained over a few months will take you much farther than a couple frenzied weekends where you work on something and then give up. Lastly, I want to acknowledge how stressful job searching can be, particularly in a new and quite competitive field like AI safety. When you deeply care about a problem like this, you put in hours learning and applying and then get rejected, it can really sting. I hope that you'll be patient and kind with yourself and acknowledge how hard you're working. But if you've listened this far into the article, kudos to you for wanting to keep people safe and make the future go well. If we're right about these risks and the possibility of the risks they might pose. This could be one of the most important problems in human history for thoughtful people to shape decisions about the technology. I think it's worth it for you to investigate seriously whether you could contribute. And you should feel proud that you're paying attention to these problems

and looking for ways to contribute now. From here, I think you should get good and be known. Great, so that's my article. I really hope you enjoyed. If you've listened this far, I just want to encourage you one more time to check out the 80,000 Hours Advising page. I mentioned a lot of links throughout this article, but I think that's a really great place to start. We have a free career advice chatbot you can use that draws on all of our content and that I know the team has worked really hard on to refine the advice that it gives. But you can also apply to speak with one of our advisors for personalized advice and connections into the AI safety field.