Everybody is (was?) always talking how important distillation is. A field of the size of AI safety needs efficient onboarding, and people writing a textbook with exercises & (afaik?) flashcards is overdue.

@Adrian-Regenfuss I would be even more enthusiastic if there were plans to also train LLMs on human brain signals, since the activations look to me to be too inflexible to bridge potentially-extremely alien cognition to human cognition. But that's a much higher ask.

I've referenced this proposal a bunch of times in conversation, and find it pretty cool.

@Adrian-Regenfuss I've also accepted a job offer which is a better fit and lasts for at least ten months. Due to opportunity cost for charitable donations it seemed prudent to give the money back and not wait for a better opportunity to arise, instead of applying again with a clear idea when I am ready for opportunities again.

What progress have you made since your last update?

I've decided to not pursue this project, or any other project of the type startup, because I've learned of three competitors doing ~the same thing, and the two of those I talked with didn't want to team up and seemed more advanced in their startup.

What are your next steps?

I am going to transfer the funds back to Nuño Sempere.

@Isaac-King There are both an e/acc wiki¹ and an e/acc forum².

¹: https://www.eaccwiki.com/wiki/Main_Page

²: https://www.effectiveacceleration.org/, rather quiet for a while though

@Adrian-Regenfuss The other one being https://www.withrhythm.xyz/.

@IsaakFreeman I talked with Simon from https://presale.getbrighter.co/, and I don't think anymore that this is a great opportunity (there's apparently >3 different people trying to start projects in the space already).

In case anyone is wondering, I've seen this competitor. I don't have any strong opinions yet.

@Adrian Hi Adrian! (lol) I am quite interested in talking to you about this. Would you be up for a chat this week? My email address is cenabzrfgeb@tznvy.pbz (rot-13ed).

For 1 & 2, I have done similar things in the past and their impact looks very limited at best, and especially the AI alignment things I'd be doing would have a pretty big theory-practice gap that seems really hard to cross.

Having met a bunch of SERI MATS scholars and people who barely did/didn't get in, I think I don't quite get to their level of technical competence.

4 is more a statement of "get more information on what plan to pursue", which I can do in any scenario.

On 5 & 6 I believe that I'd be way more replaceable than if I started the lumenator company, my intuition is that people have a bias towards joining existing organisations rather than starting new ones.

7 onwards appear obviously inferior to me in terms of ambition/impact,

but I can elaborate if this is not clear.

In general, if I succeed in following through with this project, it looks to me like there are at least three classes of success:

1. Steady small profit: The startup succeeds but doesn't return enough profit to support me full-time, but enough to keep it running on the side. In this case I could go on to pursue the other options.

2. Steady medium profit: After some years, the company generates enough profit to support me (and some more beyond that), but having plateaued, and can be maintained with below-full-time effort. In this case I would be free to pursue projects that I evaluate as being valuable in the rest of the time.

3. Full-time with medium-large profit: This would be the standard entrepreneurship-to-give scenario, in which the startup generates enough profit and grows enough to justify me spending my full time on it, and donating the money to support good altruistic projects.

   I expect there to be some path-dependence here, that if I pursue any of the options in the counterfactual I'll very likely not try to start the company (out of complacency/comfort), but if I start the company I am okay with going back to the rest of the plan.

Oh, also it looks like I can not edit the time by which this project runs out. Is that intentional?

Thanks for the feedback, Austin! These are very good points and I feel embarassed that I didn't make them clearer (and in the case of 2. hadn't thought through on it).

Having done some low-stakes iteration before writing anything about it publicly would have been a wiser decision.

1. I have not, and I should've at the least made that clear in the proposal, or waited until I'd built one and then posted it. This was my biggest point of hesitance in posting this, and in retrospect I should've just gotten some hands-on experience first.

2. Good point. I hadn't considered the pre-selling with discount.

1. My model here is that a surprising number of people have built lumenators themselves, resulting in a sort-of-shitty product at some significant time investment (a friend of mine spent ~5 hours researching the right materials before buying), and extrapolating from how many people get filtered out by trivial inconveniences

2. The "Tesla Roadster" strategy is kind of hindered by the existence of Coelux and InnerScene, which sell expensive lumenator-like lamps. Their stuff is expensive enough that they even don't list it in their catalogue or on their website, but this website says it costs ~$40k and someone from InnerScene also states that their product is about as expensive as the Coelux product, and their lamps are a bit too dim (I've seen up to 6k lumen), so it could be possible to slot below them on the price scale.

3. Okay, good to know.

4. I guess I just don't feel motivated to do software, outside of a dayjob? And there's no thing I want to make that could also generate profit?

1. I haven't run any side projects that generate profit before. (I have run a number of side projects, some of which have involved making physical things).

5. I would really like to have a cofounder. I'm in contact with two people, but think both are unlikely (<25%) to join in. Given how much I'd like to have a cofounder I should invest much more work in finding someone.

Wow, that's a strong vote of confidence! Both 1. and 2. look like I can just do them in <2 weeks at low cost, so I'll do them and report back.