You can just
do things.

Survivorship bias.

• Roughly 50% of startups fail.
• Some of my attempts failed too.
• Had they all failed, you wouldn’t be hearing me talking about them!
• AI is crazy, and the world is changing quickly.

I did a lot of programming.

Then I left Berkeley.

• I had a place at Bristol University to study Physics.
• I delayed it by a year for medical reasons — I needed to rest.
• I got bored very quickly.

“That’d be fun.”

• I was interested in languages.
• I decided to make a tool that used fancy probabilistic data structures to estimate the difficulty of sentences on webpages, then translate ones at your level to another language, so you learn by immersion.
• For the first time, I would actually sell it.
• Near-zero strategic thinking.

Metaculus AI Forecasting Tournament

• Metaculus is a prediction platform.
• They were running a tournament to build AI bots to predict the future.
• I didn’t do it: “Markets are efficient; I have no edge; it’s not worth my time.”
• This was very stupid, because markets are not efficient.

I built the language tool.

• It was technically good.
• It made some money, but not loads: 20 paying customers.
• My personality is very, very bad at selling to normal people.
• “Don’t sell to consumers” was common wisdom. I knew that and ignored it, and the common wisdom was correct!

But I did some research.

• I needed research, couldn’t find it, and did it myself.
• I wrote a blog about what I found.
• It was ridiculously low-hanging fruit.

The world is smaller than you think.

• Nobody else was doing AI translation research.
• I was invited to do a paid talk at a company, and random Japanese professors emailed me to chat about it.

Let’s try Metaculus.

• I built something super simple.
• It turns out that nobody else was trying.
• I came second place. Good to know! Let’s keep at this.

I tried building something to trade Polymarket.

• That failed.
• Sometimes things fail!

A fancy translation system.

• I got a few business customers.
• Then I realised that building a translation company when I don’t know any other languages is patently stupid.
• I focused on forecasting instead.

It does translate poetry very well, though. Still a good party trick when I meet someone who actually speaks the language!

University was Basically Fine

• I emailed Exa for higher rate limits for my bot.
• The website said higher limits were for business users only. They raised the limits anyway. Cold emailing works!
• They offered me an internship, pending legal being OK with a British citizen.
• Then they never got back to me.

Again: failures are very normal.

Jasmine and Jeremy

Jasmine · 20

• An incredibly accomplished, intelligent AI safety researcher.
• She encouraged me to get into AI safety, explained the routes, and gave me a very useful reference.
• She knows everyone in AI safety.

Jeremy · 20

• An economics student in America.
• He knows everything about AI economics.
• He seems to know everyone in the San Francisco Bay Area.

Lots of other people:

A week later: San Francisco.

• Jasmine told me about BlueDot Context Week. I applied and went.
• I met lots of really cool people!
• Cultural difference: networking is not a dirty word, and everyone wants to help you succeed.
• I started the extensive application process for MATS, an AI safety research fellowship.

Access + eloquence

• Jasmine invited me to an exclusive event in San Francisco.
• I chatted with senior AI researchers.

“You’ve clearly thought this through a lot.”

“Once we start doing internships, you’re first on my list.”

You Can Just Talk to Americans

• Everyone wants to help.
• You can make friends with AI researchers while waiting in line at the airport. I did.
• It’s a very different culture. A week in the Bay Area will get you very far.

I said hi to London’s AI safety researchers.

• I asked Jasmine for contact details.
• I emailed or texted and asked nicely.
• I went around the MATS and LISA offices.
• It was totally informal. You can just do things.
• This probably would’ve worked without Jasmine too; she just made things smoother.

An opportunity via WhatsApp.

• Jeremy went to a rationalist summer camp in 2025.
• Someone posted in the group chat that they were looking for forecasters for a hedge fund.

I’m not a forecaster.

• I DMed him anyway.
• I pitched that AI forecasting was going to replace human forecasting.
• They should buy my software.

That somehow worked.

• A month of negotiations and presentations.
• We agreed on an exceptionally well-paid internship.
• I’m there until January, then either heading off for MATS or staying in a full-time role.
• I also went to an EA conference and met friends from Cambridge, plus lots of other cool people!

I’m smart, but I’m not a genius.

Compete as little as possible.

• Everyone is fighting over the same tiny sliver of possibilities.
• The normal track means thousands of applications and stupid psychological tests.
• Basically everything else has far less competition.

Give it a go.

Cold Emails are Unreasonably Effective

• People will usually reply if you do them well, and they don’t take long!
• Never use AI.
• Reduce fluff and increase signal, then KEEP DOING THAT until it’s comically compact.

Legibility

• A tournament position.
• A blog.
• A project people can actually look at.
• Internships, fellowships, and prior work.
• Even a Twitter account where you’re smart.

Put yourself out there in a way that fits in an application.

Eloquence is important. This is gameable.

• Nobody is asking you to do an equation at a party.
• Read a lot. Write a lot. Think about things.
• This isn’t about purple prose or English essays.
• Read Paul Graham’s essays, Dwarkesh podcasts, and LessWrong.
• Actively increase your active vocabulary: the words you draw upon in conversation.

Practice speaking to people.

• Presenting in a boardroom.
• Talking to a powerful person at a party.
• Discussing things in a small group.
• These come up more than talking to a big crowd.

Depending on how I’m doing right now, this may sound hypocritical.

The audience matters.

The success of the “weirdo 20yo” depends heavily on who’s listening.

Academia is a particularly hideous status game. Avoid it.

Twitter democratises access.

• In certain niches, it’s incredibly useful.
• If you have smart things to say, you will quickly find yourself randomly chatting with very powerful people.
• You find interesting people, research, and opportunities.
• It gives you lots of material and context to discuss.
• The format might not fit you. I’m terrible at posting!

Surround yourself with good people.

• People who are smart, driven, and encouraging.
• People who actually do things and help each other.
• I mostly found them through Twitter, Invisible College, and the Bay Area.

Effective Altruism/Rationalism

• These are big subcultures in the Bay Area, startups, and AI.
• They’re very friendly to people inside them.
• Disagreeing with half of EA/rationalism is pretty much what defines you as an EA/rationalist. You don’t need to believe it all to benefit from ingroup bias.
• The way you write reveals whether you’re an EA/rationalist, so you can silently express it in an email or job application.

What about AI?

• Everything is changing.
• Be willing to change what you’re doing. Pivot fast.
• Use AI to learn and build. Keep your own judgement.

Watch this lecture.Miles Cranmer / Cambridge
How to Avoid AI Brain Rot

There are loads of things to apply to.

• AI safety has a whole pipeline: courses, research programmes, and fellowships.
• Ask your AI to find things that fit your age, interests, and experience.
• Apply to Invisible College too. Your AI might miss it.

BlueDot · SPAR · MATS · ERA · LASR · ARENA · Invisible College

Todo lists and notes you actually USE.

• Keep track of applications, deadlines, and people to reply to.
• Write down ideas you’ll otherwise forget.
• Have something you actually use to keep on top of things.

80/20 your entire life

• You can generally get 80% of the way with 20% of the effort.
• Do many things, slightly worse.

But rest when you can

• Preserve the capacity to go 100% for weeks or months when you need to.