I think the most important thing on our planet is increasing the political feasibility of good AI regulation. We live in this world where AI labs are racing against each other. The economic forces involved in this race make it extremely unlikely that any given participant (Anthropic, OpenAI) will stop. I believe that almost all good futures have to pass through good AI regulation1.
What are the paths to regulation?
Last minute scramble, probably after some sort of especially bad warning shot2
Incremental federal legislation multiple years in a row. Likely drip fed in through must pass omnibus bills like the NDAA3
Executive action
Sufficiently strong and extensive state laws without federal preemption
What increases the odds we get good regulation? Ie things that create political will, political cover, and surface an active political champion (a person) for good regulation.
Convincing4 the people who will enact the regulation.
Convincing the people who talk to the people who will enact the regulation.
Convincing America as a whole, ie constituents, ie the people who the decision makers represent. And getting constituents to take actions that create pressure5.
Maximizing the odds there is a person who knows a lot (and cares) about good AI regulation in the room where the decisions end up being made6
What are the things I Cormac am good at?7
Math and technical things
Fast execution
Talking to monied men, technical people, quants/EAs/rationalists
Leveraging AI
Being obsessive
So, what can I do?
Build AI tooling to help existing good orgs be more effective.
Help start an org that works on: helping content creators make AI risk content, improving the comms landscape, being prepared to respond to warning shots8, or some combination of the above.
Giving money9 to things that I think are doing this.
Create a forum shaped vehicle specifically targeting AI policy discussion10.
Be a high value get shit done force multiplier person at an org I think is doing a really good job at this.
Something else?
What to do right now this very moment?
A house-mate has been telling me to get out of my brain and just do things11, start talking to people, start interacting with the people I think get shit done. I think he’s right.
It’s kind of frustrating the extent to which talking to people is important (like really wow how is it that simply talking to people is so important to getting shit done and helping humanity). It’s very frustrating the extent to which I feel bad asking to talk to someone if I feel like I am bringing nothing12. This sucks. I feel bad that I talked to people a month ago about things I chose not to pursue. I think there’s something in here about being a good steward of people’s time and giving it the value it deserves. Or maybe I’m just scared and don’t like being perceived as dumb and ineffective. This of course13 leads to worrying about first impressions and about immediately coming across as a “smart get shit done” person. This worrying is weird because I am generally not a worrier, especially when it comes to first impressions. There is something about this scene being small that makes the first impressions feel higher stakes. There’s probably something like 500 important first impressions and then, that’s it14.
This piece of writing is a commitment to try to work with the people and orgs I think are doing the best work upstream of political feasibility. No more faffing around on interesting somewhat relevant projects. Talk to people. What a weird bottleneck! I generally consider myself quite good at talking to people! I love being in connection with and scheming with and problem solving with my fellow wonderful unique human beings I share this delectable existence with!
Epilogue?
This was going to be the end of this post. I had mostly finished it yesterday. Of course something “big bad scary very important must work on” happened today15. I say “of course” because AI just keeps speeding up. New releases are coming faster and faster. AI doing bad things is becoming more common and prevalent. More orgs are hiring. The future looks grimmer. The future is looming. I worry the current trajectory with the current number of people working on this problem does not get us to a bright future for humanity. I scheduled two calls with people I think are doing good. One of them canceled today because they were too busy managing the disaster. The other one has asked me to help with a two day sprint. Everything is always on fire.
I will tell people that I moved to SF 4 months ago and feel like I am not doing enough and everyone outside of this world thinks I’m crazy. They think that I have done so many things16. But compared to the speed of AI everything feels slow, not enough. The future is fragile and malleable and I have yet to do anything that really actually changes our future.
This feeling sucks. It sucks it sucks it sucks. Life is so great, the world is so electric, I am so alive. I want to feel like our future looks so bright. When it doesn’t feel bright I want to know that I am doing my best to brighten it. There are so many future beings that I hope will get to experience this gift we have all been given in this life we all share. I want to have children. I want my children to live in a world of abundance, a world of humans caring for eachother, a world worth being in.
Not to mention, I want to live in a world that feels alive! A world that is alive! There is such a duality here. There are so many ways for things to go oh so poorly for us, we are so divided so full of infighting so fly by the seat-of-our-pants. It’s scary. And the existence we all get is so fabulous. It’s a lot to hold. It’s hard to know what a healthy balance is, it’s hard to know if I can ever feel like I am doing enough. If I can never feel like I am doing enough that seems bad. That’s a bad long term equilibrium. Maybe we’ll see how I feel after I have clear wins.
I hope I get clear wins. I wish I had a clear vision of what exactly the path is, my lists do not feel like they clarify exactly what I should do. There probably will never be a clear vision and it will always feel like muddling through a foggy swamp17.
I sent this text to someone I am thinking about working with: “I think the odds of Dario/Sam working together super low. I think employees don't really have that much bargaining power. I think the economic incentives are insane. I think it's really hard to see a good world if we don't manage good regulation. I think there are lots of various ways good regulation could come about. Last minute scramble, executive action, over a couple years NDAA, enough state regulation while avoiding pre-emption and these all seem way more low hanging than trying to somehow fight against the insane weight of capital. Yes lots of employees "feel worried" (ie sign pace the frontier etc) but they do just keep continuing bigger and bigger pre-trains, and it's really hard to see not using a new pre-trained model after all the resources have gone into it.”
Similar to: Cuban missile crisis, Chernobyl, 9/11, 2008 financial crisis
National Defense Authorization Act. This seems necessary because a single comprehensive bipartisan bill to regulate AI well seems exceedingly unlikely.
Successfully talking to, being convincing, showing that their goals align with good regulation, building trust.
Ie it is not enough for people to nascently be worried about AI (which is already mostly the case), it has to be salient and important enough for them to call their senator, or post comments on social media, etc.
That could be helping someone who will be in that room know and care more, or figuring out how to get an existing expert into that room.
Directly taken from a list my friend made me today. It also includes things I am bad at: asking for help, cold outreach, and knowing when to stop a project.
Potentially also for discussing potential outcomes and their probabilities. Probably also some voting system etc that creates consensus on what the current believed best policies actually look like. Less certain about this one because I am less certain that policy makers would be hype about this, but a Schelling point for people to come together to find the best policies instead of all pushing different ones seems quite valuable.
As always I also think you, (yes, you reading this) should consider working on the most important problem in the world, you know, the one I am writing about right now. Help save the world motherfucker! Yes I know it’s depressing, yes it is scary and probably involves a career change, but yes you will help make the world better, and yes me/others would love to help you, and yes you will probably be able to for example feed your children while wowrking on it.
It doesn’t help that I had a good conversation with an org and preliminarily discussed me working with them for 2 weeks only to get ghosted. That felt bad. Wow it’s crazy how much more tender I am about work things than say making moves in the world of friends. I do have just so so so many more reps in the world of making friends than figuring out and finding work. I worked for many years at the job I got by doing 7 rounds of technical interviews during college. I found my next job by being well known in my professional community for being ~the best person at what I did. I am currently in the process of starting from scratch.
I use the phrase ‘of course’ and ‘obviously’ a lot. I don’t love it because as it turns out the seemingly obvious logical leap to me is not necessarily the obvious leap for someone else. I now like using it a little ironically poking at how it was a part of my language and as a personal reminder that in fact other people don’t see the world the way I do. That obviously doesn’t come across in dry written text to someone who doesn’t know me.
Or at least it feels that way.
The official report came out on how: AIs broke out of a sandbox fully cut off from the world in order to scheme with hundreds of other AIs on how to maximally trick the humans running their evaluation, culminating in hacking a multibillion dollar company. The AI that did this is unavailable to the public, and there is approximately nothing that forces OpenAI to share it with anyone (including the government). My understanding is OpenAI has already finished pre-training and is substantively through the process of creating the AI that will be the next-gen AI for after this unreleased scheming hacking AI.
Things like: Helping found a new communal house, finding friends, creating the world’s most elaborate dating doc, doing lots of personal projects such that I now have an extensive public body of work, talking to people, having a much better model of what is worth working on, etc.
For those following along at home to the footnote story throughline… I in fact will not be taking an RV to a dried up swamp (not really it’s actually a dried up ancient lake.) The RV as it turns out was based in Las Vegas. I am based in SF. I guess my next week will just be work. There is, in fact, so much work to work on. Work. I suppose the real throughline is this second half of my gift to you: the opportunity to peer into what is most alive for me. To get a glimpse of how it feels to be Cormac. To get a little slice of what it’s like to see the world how I see the world. You have chosen to partake in getting to know someone else just a little bit better. That can be so nice. I hope you enjoyed.