You believe that working on AI safety is one of the most important things you could do. So what do you do next?
You feel in your bones how urgent this is. AI is improving rapidly. The next OpenAI-Hugging Face incident could be close, and much worse. You hear people in the field talk about how short their “timelines” are. You’re not entirely sure what that means, but you have a looming sense that the time to act is now.
So you do a BlueDot course (perhaps all of them). You spend hours consuming LessWrong, blog posts, newsletters, podcasts, research papers. You apply to all the jobs and fellowships that seem relevant.
Yet you still feel lost.
You struggle to explain why you want to work on mech interp or compute verification or US-China coordination. These all seem important. But it feels like you’re working backwards to find the reason.
That’s because you’re optimising for the wrong thing.
Job ≠ impact
When your goal is landing a job, you start from who’s hiring and look for a reason to want it. When your goal is solving a problem, you have far more options than the jobs currently open.
Many of the jobs needed to solve these problems don’t exist yet.
The field is expecting tens of billions of dollars in philanthropic funding, and funders are looking for good ideas to back. That makes creating your own opportunities more realistic than it has ever been.
So focus less on getting hired and more on solving problems. That starts with…
Understand what’s going on
Spend at least 100 hours understanding what is going on in the field. If you’re open to a wide range of problems, it could be closer to 500.
This might sound like a lot when time feels short. But if you pick the wrong problem, going really hard at it achieves nothing.
AI safety is a wicked problem. People who have been working on this for years still disagree (often vehemently) about what a good future looks like, let alone how we get there.
Some believe we should prepare to hand our future over to a superintelligent AI. Others think that we should ban the technology from being built altogether. Others believe it would be immoral not to unlock the benefits this technology could bring us.
Every option has trade-offs.
For example, preventing people (and AI) from bioengineering pandemics pushes us to restrict who can access and build powerful models. But that means betting that the few who do have access will not abuse it to stay in power. It also means betting that they’d be able to keep these increasingly powerful AI systems under control.
If the problem were simpler, perhaps the few thousands of people working on it full-time might have solved it by now.
Resist the temptation to latch onto your first idea. Figure out your own answers to:
What is the goal of AI safety? What does it mean to “win”?
Why do people disagree so wildly about this?
How has AI progressed over the past decade, and how do you expect it to progress over the next 1-5 years?
Who is building it, and why do they continue to do so?
How exactly might AI cause catastrophic harm? What’s so hard about preventing this?
You’ll end these hours feeling more confused. That’s how you’re supposed to feel. Your confusion will be more specific, and that’s where you want to be.
This requires:
Time. Understanding the field is hard to do alongside a full-time job and personal responsibilities. If time is what’s blocking you, apply for a career transition grant.
Quality conversations. Spend a few weeks with people in the field at places like London Initiative for Safe AI (LISA), Constellation, FAR Labs or other communities.
Structure. Apply to in-person, multi-week experiences which help you build the context to pick your problem or join communities that run structured reading groups, so that you can get frequent useful feedback. Find them here.
Narrow down to 1-3 problems worth solving
Then, pick one and start. You’ll learn more about what the problem entails as you work on it. The field’s priorities change rapidly, and yours will too.
The problems you could work on are not as they seem
Don’t assume you can’t contribute before you’ve understood what the work is.
Maybe you think a bioengineered pandemic is coming and we’re unprepared. You assumed that without a bio background there was nothing you could do. Then, you learn that there is biosecurity work that doesn’t require it. For example, building a respirator stockpile for critical workers is a problem that requires supply chain expertise more than it does bio.
Your background doesn’t limit you either.
A computer science degree doesn’t mean you can only do technical work. A legal qualification isn’t all you have to offer. Often, a problem just needs someone willing to step up and figure out what to do next.
Be ambitious
You’ve landed in a young field.
Tens of thousands of people have been working on problems like climate change, nuclear disarmament and eradicating malaria since the 1950s. Whereas AI safety had a handful until the 2010s and has only a few thousand now.
You are still early. You can pick a topic and, within months, be one of the few people in the world who understands it. On many topics, you could speak to everyone working on it in a week.
So now is the time for moonshots. Think expansively about what’s important to you and what you could do:
What about the future with AI scares you most?
What part of it do you most want to work on?
What bets is the field not making that you think they should?
Learn from the people already working on it
They know things that aren’t written down anywhere: experiments that failed, results that aren’t published yet, the reasons behind their strategy. Speak to as many of them as you can. Ask what problems they see and what’s causing them.
Then plant your own flag on what you think needs to happen, and move it when you’re wrong.
You’ll know you understand a problem when you’re shocked by how simple it is to explain.
Start working on it
You don’t need a job to start working on a problem. Here are a few ways to get started:
Apply for a rapid grant to fund your project.
Offer help to someone already working on it. Figure out who else is working on your problem, why, and what they need. Then email them: “You might need X. I can do it for you in Y way.” The field is starving for talented people who want to help.
Pitch a role. Work on the problem independently for a while. If you realise that it’s a full-time job, pitch an org to hire you to do it. Orgs want to hire people who can hit the ground running.
Start a project or org of your own. If no one is working on it, consider stepping up to do it yourself. Apply for Incubator Week or a Seed Grant.
You may well end up in a job. But you’ll have chosen it because it’s the best way to work on your problem.