I'm interested in advice on retirement savings - mine are far smaller than Jeff's and reading this gives me slight anxiety haha. It would be action-guiding for me as maybe I should just not push myself to donate more and instead have a more solid retirement plan. Kudos to you Jeff on being transparent and generous!

Thanks so much Mo! I am tempted to make the following updates already - does this seem roughly right? Or is this still too high?

1. Token usage at 8 hrs centered on 5M tokens, with an upper limit closer to 100M. The reasoning for the
   1. Upper range of 100M being that more complex tasks (assuming those from the study you quoted were low hanging fruits) might push this higher (as indicated by the compiler example), while
   2. efficiency gains might push lower, it already seems that from METR's GPT-5.1-Codex-Max work <6 months ago it might, and this is very, very crude, be going lower.
2. Token price centered at $1 per million tokens, instead of $5. I could make this even lower as $1 might show a downward trend, but at the same time this low price seems more to be due to cache tokens which I had ignored in my analysis - the input and output tokens still seem priced at roughly the price I found

At the same time, I also feel like these numbers might still be too high - especially token price. The reason is that the super helpful links you sent point at pretty steep downward trends on token cost and point well taken on cache tokens being much cheaper.

Sent draft on DM - I prefer to keep this non-public both because my employer needs to review this before it goes out and also because I would like a few people to find my biggest errors before I put it out there. But I aim to get this posted publicly in the next couple of weeks. Thanks!

I mean if you can move DCs and production/mining/etc. to space that's a solid win for AI safety? We then basically delineate AI and the only human-inhabitable planet we know about. That's a heavy lift though, but maybe water on that one moon of Uranus (?) and other parts could be assembled off-Earth. This is literally galaxy brain thinking but something I feel might be worth pursuing, or at least showing AIs that it might be cheaper for them to pursue off-Earth living than engage in uncertain and high-cost conflicts on Earth. Maybe they don't care about time, and would be happy to build towards this over the next 1000 years instead of spending valuable compute and industrial capacity to prepare for AI-human conflict.

I have looked into AI and energy (happy to share my drafts with anyone interested). My impression is that it is not the cost of energy that drives orbital DCs, but instead the availability. It is not only orbital DCs that are being considered, the portfolio includes hopelessly naive stuff like floating DCs powered by ocean waves, restarting Three Mile Island, SMRs and much more. If energy consumed by human-equivalent AI task performed does not drastically reduce, the inference energy demands will far outstrip even the most electrical generation the world as a whole has ever added in a single year, even at low labor replacement rates. If anyone is working or thinking about this I am super interested in talking. I am hoping to publish my initial thoughts soon. The upshot: As with compute, if energy becomes a limiting factor, it might be a good point for interventions. For example, electricity regulating authorities (there are many and strong ones!) can incentivize disclosure of model capabilities in e.g. cyber, which is extremely relevant to cyber attacks on the electric grid and thus plausibly lands under their jurisdiction.

I now think principles-first EA is more important than I previously thought because it helps prevent effectiveness drift. My anecdotal evidence from AI Safety and especially biosecurity gives me the impression that without constant anchoring to EA and especially comparisons to the clear ToCs and tractability for e.g. AMF, it is easy to lose focus on the high demands of choosing x-risk as a cause area over others. I previously placed less value on having strong links between EA and the various cause areas but now think I should update to thinking strong, continuous engagement with EA is important to keep one's focus on each intervention's cause prioritization assumptions that make it comparable to e.g. AMF or cage free chicken campaigns. This is not to say that causes such as AI Safety and biosec are not important, but that unless constantly tied back to EA cause prioritization, there is risk of drifting away from what made the cause look good in the first place. An example from biosecurity is the very easy slippage away from human extinction scenarios to ones where nearly everone dies (the difference being a crux as it is the potentially enormous future one is saving, not the people living at the time of the catastrophe). That said, it is completely fine and commendable that there is AI Safety and biosecurity work that does not target existential threats, but for EAs such changes in the nature of the work means they should consider changing their career. I think this is also important for newcomers to EA: For those of us who were around when we discussed whether x-risks demanded attention we might take concepts such as Pascal's Mugging as obvious, but for newcomers it is important to engage with such concepts. Another observation I have made is that in animal welfare and global health one is constantly reminded by metrics of suffering alleviated per dollar, but such recurring reminders are lacking in x-risk focused cause areas.

I am really happy you are focusing squarely on existential risk from bio. I think there is a tendency in EA-adjacent biosec work to lose a bit of focus on how extremely bad such scenarios are. I also think it is great you raised this Michelle - I also feel like not enough EAs have contemplated the importance of 2 further assumptions needed to work on longtermism:

1 - Massive increase in value in the future (re: Arepo's billions of star civilization), and
2 - Very few or not other periods of existential risk for the rest of the infinite future

Sure — I may be mixing abstractions here, so let me spell out what I had in mind.

A 1000× reduction means that, in expectation, the protected environment has 1/1000 of the relevant airborne particle concentration compared with the outside environment. So if an unprotected person would inhale 1000 relevant particles over some period, a protected person would inhale about 1 over the same period, ignoring spatial variation, time dynamics, leakage events, deposition, behavior, etc.

My intuition was then: if the minimum infectious/lethal dose were effectively 1 particle, and if we wanted lifetime infection/death risk inside the protected space to be below ~1%, then the protected person’s expected inhaled dose would need to be on the order of <0.01 infectious particles over the relevant period. With only a 1000× reduction, that corresponds to an outside unprotected expected inhaled dose of only ~10 infectious particles over that same period.

That seems surprisingly low compared to my own work on mirror bacteria. So I was wondering whether the 1000× target assumes one or more of the following:

1. the relevant environmental concentrations are expected to be quite low;
2. the true infectious/lethal dose is meaningfully above 1 particle;
3. the 1000× PM10 reduction is only one layer, with additional reductions from UV, glycol vapor, surface controls, masks, behavior, etc.;
4. the target is meant as a practical near-term benchmark rather than a complete risk-reduction target.

So my question is basically: what outside concentration / dose / acceptable-risk model makes 1000× the right threshold?

Perhaps this is not all good news - we want AI to embody all of humanity's values and desires and possibility for flourishing. While it might be true that EAs have found some path and framework close to optimal for human flourishing, I think we should also be skeptical that both AI and EA has emerged from a very narrow set of humanity, tech-centered, etc. Seems like there is a lot to unpack at more meta levels.

Perhaps I am missing something but on the >=1000x criteria, if we target e.g. <1% of people succumbing to the disease over their lifetime (might want to set even lower, in order to make people comply with suggestions - precedent of similar risk reductions and uptake might be worth looking into if not done already), this means we expect people that are not protected to inhale only 10 particles over their lifetime, in expectation (assuming minimum lethal dose of 1 particle)? Asking as that seems like a small degree of environmental spread. I realize that perhaps the reasoning here might be infohazardous, but if not I would be very interested to know more. Or perhaps additional reduction comes from one or more additional measures, such as far UVC, glycol, etc.