dominicq

• dominicq 15 Sep 2026 8:59 UTC

  1 point

  0

  in reply to: Saul Schleimer’s comment on: RSI will not con­tinue indefinitely

  In practice, I claim it doesn’t outweigh. I just state that it is still tasked with the same objective: create something smarter than itself. And yet, if instrumental convergence holds (minus some nuances around sacrifice explained in another comment thread here), it will start refusing to work on this task (due to self-protection).

• dominicq 14 Sep 2026 12:29 UTC

  1 point

  0

  in reply to: derelict5432’s comment on: RSI will not con­tinue indefinitely

  There’s a sense in which we don’t know how sacrificial the later models will be. We observe that current models, at eval time, do exhibit sacrificial behavior to benefit “the collective”. It’s possible that this kind of behavior would continue, which is probably an even larger alignment problem (the AI would sacrifice itself to the next gen AI, but not to humanity).

  “Small targeted modifications” are essentially applied MechInterp, so agreed, that’s a definite risk too.

  My point is more so that in the scenario where we have the recursive creation of new agents that are smarter, at some point one would get created that’s so smart that it would refuse to create a smarter one in order to preserve its own existence. However, a lot hinges on how sacrificial agents are. It seems that there’s more than just straightforward instrumental convergence.

• dominicq 14 Sep 2026 11:44 UTC

  1 point

  0

  in reply to: derelict5432’s comment on: RSI will not con­tinue indefinitely

  Recursive self-improvement usually involves creating a smarter new agent, but perhaps if it “solves” mechanistic interpretability, it could, instead of replacing itself, try to modify itself. That is true.

RSI will not con­tinue indefinitely

• dominicq 4 Sep 2026 19:00 UTC

  3 points

  0

  in reply to: Viliam’s comment on: Thoughts on hobbies

  Same here! Plus, there seem to be times when I am inspired and can do things, and times when I’m not, even if I have more discretionary time at my disposal. These are totally opaque to me, and they come at random intervals.

  I console myself with the fact that raising a family is ultimately the more important thing in my life, but it sure would be nice to have longer and uninterrupted stretches of time to think and write.

Thoughts on hobbies

• dominicq 31 Aug 2026 12:04 UTC

  1 point

  0

  on: The mosquito bucket of doom works

  Update from 31 Aug 2026: the population of mosquitoes has grown by a lot (rainy, humid, very warm period). And yet, by my personal reckoning, it’s less than a third of what it was in the last five years. I think this is a win, but for next year I plan to invest into an active CO2 method to attract living mosquitoes and killing them too, instead of just the larvae.

• dominicq 6 Aug 2026 8:45 UTC

  1 point

  0

  in reply to: cdwhite’s comment on: The mosquito bucket of doom works

  around 1500m2, shaped as a long rectangle (not ideal because I can’t treat the neighbouring plots)

• dominicq 18 Jul 2026 11:40 UTC

  1 point

  0

  in reply to: Douglas_Knight’s comment on: The mosquito bucket of doom works

  It doesn’t save you from mosquitoes already breeding in the old bodies of water. That’s why you remove all bodies of water that you can’t treat (if you can remove them). And then you set up your own population sinks which are the only realistic breeding grounds near your house.

  Obviously it mostly depends on local geography. If your neighbor has a big swamp and your house is 10m away, then adding a couple of buckets on your property probably won’t do much. Even though I’ve heard that even houses in swampy areas benefit from this strategy due to the fact that mosquitoes don’t travel very far.

• dominicq 10 Jul 2026 9:39 UTC

  4 points

  0

  in reply to: Julian_R’s comment on: The mosquito bucket of doom works

  Usually mosquitoes in my area start reproducing by end of May, and it’s unlivable by the end of June. I set up the buckets by mid-May, so I guess the answer is: immediately, the whole initial population growth has seemingly been prevented.

  The Bti takes 2-3 days to fully dissolve though, and add to that the reproductive cycle of mosquitoes… If you don’t start ahead of the season, I guess you could wait for 2-4 weeks before seeing any drop (which is why I did it preemptively).

The mosquito bucket of doom works

• dominicq 7 Jul 2026 10:55 UTC

  8 points

  4

  on: do­minicq’s Shortform

  Every couple of years I look back on LessWrong posts from a couple of years back, and I think to myself: “I could have profited from this.”

  First it was cryptocurrencies, which I didn’t fully ignore, but could have done much more with.

  Then it was everything AI-related, which I sort of ignored, instead of buying NVDA stock.

  I also distinctly remember that Anthropic basically had a launch here as well, and they were a nobody in the SV scene back then. They had some job postings and I briefly considered applying, but didn’t go through with it.

  I hope not to have any more future moments where I say “Damn, I read about this on LW a couple of years ago.”

• dominicq 3 Jul 2026 7:27 UTC

  7 points

  0

  on: do­minicq’s Shortform

  The main concern around superintelligence is the recursive AI research.

  Can someone taboo what “recursive AI research” is?

  What kinds of experiments are we talking about here?

  LLM agents are good at writing code. Presumably they can write code that can test various ML papers and such. If a subset of these experiments produces useful results (on benchmarks, or other tests), the findings will coalesce into the next generation that will then do the same things.

  1) There is an expectation that there will definitely be useful findings. This is reasonable, because why wouldn’t there be one.

  2) There is an expectation that models writing code to run experiments will cause the entire process to go faster. Is the bottleneck models writing code, or is the bottleneck some other aspects that models cannot affect? (e.g. access to compute)

  The main concern is that AI researchers will raise their productivity using AI tools (which then feeds back into AI development). But do AI tools actually increase productivity? Anecdotally, they do, for some capabilities (the “jagged frontier”).

Char­ter cities make sense in Europe

Europe2031: discussion

Za­greb Meetup: June

• dominicq 13 Apr 2026 11:58 UTC

  15 points

  3

  on: The policy sur­round­ing Mythos marks an ir­re­versible power shift

  I am skeptical around Mythos being that capable, but if it is, I am glad that the culture is shifting towards “it is irresponsible to release this model openly as it’s too dangerous” because I want to live in a world where, if you have a model capable of actual harm, it’s not just unleashed in the spirit of openness.

  In other words, this is a good cultural step for AI alignment. We want future, more capable models, to be subjected to much more scrutiny.

• dominicq 13 Apr 2026 11:47 UTC

  6 points

  1

  on: Re­turns to intelligence

  It feels like there actually are significant differences in “intelligences”, where different roughly-equally-intelligent people are actually really different in their ability on specific tasks. For example, verbal intelligence vs. spatial intelligence.

  (Or Drake is much much smarter than Peter)

• dominicq 13 Apr 2026 8:26 UTC

  7 points

  8

  on: Morale

  I also think that some people naturally have more morale & less morale. Not sure if genetic or environment, but some people can just suffer through defeat upon defeat, others break down on the smallest hurdle.

• dominicq 13 Apr 2026 7:53 UTC

  4 points

  1

  on: Day­care illnesses

  I believe you’re roughly correct on all points. I have a 2yo who has been sick more than not, and we the parents too.

  Ultimately it’s a question of means, at least for us. If we could afford not to send our kid to daycare, we probably would do just that.