Project summary

SciencePetr is a YouTube channel doing Veritasium-quality science communication for AI safety.

AI safety communication is science communication. It is also the single hardest science communication challenge I have ever seen, and I've seen, and successfully tackled a lot of them during my time at Veritasium.

Imagine someone came up to you in 1942 and said: "The government is building a top secret lab in the desert. They have the smartest minds on the planet. They're going to split the nucleus of an atom, and it will release unimaginable energy." The average person's response would have been: what is an atom?

That's where we are. The public is being asked to form views about recursive self-improvement, mech interp, deceptive alignment, and loss of control before they have a working mental model of what a neural network is. They don't know about AlphaGo, or Move37, or what GPT stands for, or what an LLM is, or why RL is inherently agentic. What is a neural network? What do you mean we don't build them? RL? RLVR? Alignment? What?

And it's actually worse than 1942 in a few ways. The atom eventually got a Hiroshima and Nagasaki – and after that, it was undeniable. AI risk may never get one before it's too late to matter. The experts visibly disagree with each other. The technology improves while you're mid-explanation. The thing we're warning about ships as a fun consumer product — and the actually good models sit behind a $20-a-month paywall. People don't understand cyber, or bio, or RSI. How could they?

I'm a physics PhD, formerly at Veritasium, and I built science content at Palisade Research including work with Geoffrey Hinton which hit 800k views. I started this channel a few months ago -- it's now at ~40K subscribers, and I've made videos about the Andon Labs store, Cognitive offloading, and Mythos. I am currently working on a video about the METR graph, and on two videos with Yoshua Bengio.

Palisade Video -- https://youtu.be/A3HjNYDIhGU?si=Tz1kKXwAPLPXqgTH
Andon Labs Video -- https://youtu.be/NsjKnBhvTco

What are this project's goals? How will you achieve them?

Make the world's best Science Communication videos about AI safety. Have this channel be the "Veritasium for AI safety".

That means: get people to engage seriously with this. Get people who aren't interested interested. Lead from the front — prove that AI safety comms can be high-quality and respectable, not preachy or amateurish. And make videos that actually change minds: get people to vote, to march, to change careers.

How: make great videos. Talk to the experts — the Hintons and Bengios of the world, the people inside the safety orgs. Hire great people and stay ruthlessly picky on quality. Keep quality at the forefront of all I do.

The plan is to make 12-18 or so amazing videos a year. I would also like to start a podcast, but will need to hire people for this.

How will this funding be used?

• My time — it lets me focus on this full-time. I want to do this full time, and not split my energy into other projects to pay the bills.

• My team of editors and animators. Some are full time adjacent, and that gets very expensive fast.

• Hiring more people: a full time writer, an ops person, more freelance support.

• Travel costs

Who is on your team? What's your track record on similar projects?

Me, plus contract editors and animators. These editors are animators are also from Veritasium, so they're world class.

Physics PhD. Former Veritasium — the highest-production-value science channel on YouTube. Won a Streamy Award, made dozens and dozens of great videos with 10s of millions of views. Built science content at Palisade Research, including a Hinton video that hit 800K views. In a few months, SciencePetr has reached ~40K subscribers with videos on the Andon Labs store, cognitive offloading, and Mythos. Already funded by FLI, with three videos commissioned by BlueDot. Currently producing two videos with Yoshua Bengio.

What are the most likely causes and outcomes if this project fails?

Most likely cause: distribution stays stuck. The videos reach existing subscribers but don't break into YouTube's suggested feed — a known, specific problem, not a quality one. If I can't crack it, reach plateaus and the mission never scales past the people already paying attention. Algorithms are hard. Titles/Thumbnails are hard -- I think it'll work, I'm betting my career on the fact it'll work, but it's hard!

Second: chasing reach dilutes the seriousness that made the channel worth funding. I am not interested in losing quality, but this is the trajectory if purely following the incentives (which is why your money helps here!)

Third: I'm not expecting to be hit by a bus, but... this is mostly a one-person operation, and I'm the constraint. (This is also something I would like to fix with more funding).

Outcome if it fails: SciencePetr stays a respected-but-small niche channel. Real opportunity cost, but little downside risk.

How much money have you raised in the last 12 months, and from where?

35k from FLI via the Digital Media Incubator, 120k from BlueDot for 3 videos to promote their courses.