Nicholas Guttenberg
@ngutten.bsky.social
You could imagine making the conservation version, but then allow that to be broken by e.g. applying an oscillating external field - then things which arrange to resonate with that field would have these sorts of glider behaviors, whereas things which didn't would remain stationary.
November 19, 2025 at 3:38 PM
You could imagine making the conservation version, but then allow that to be broken by e.g. applying an oscillating external field - then things which arrange to resonate with that field would have these sorts of glider behaviors, whereas things which didn't would remain stationary.
I wonder if it'd just reduce to chemistry-like organizations if you did have momentum and energy conservation... same sorts of stable configurations (linear segments, 60 degree angles in 2d) but none of the vigorous motion breaking stuff up...
November 19, 2025 at 3:37 PM
I wonder if it'd just reduce to chemistry-like organizations if you did have momentum and energy conservation... same sorts of stable configurations (linear segments, 60 degree angles in 2d) but none of the vigorous motion breaking stuff up...
Neat, I guess that's some evidence towards geometric universality in these things... swarm chemistry, flow lenia, particle life - all things with mass conservation but without energy and momentum conservation, their gliders all tend to be these blob chains with 'ears' in front.
November 19, 2025 at 3:37 PM
Neat, I guess that's some evidence towards geometric universality in these things... swarm chemistry, flow lenia, particle life - all things with mass conservation but without energy and momentum conservation, their gliders all tend to be these blob chains with 'ears' in front.
My own extreme version of this: get rid of the related work section or survey parts of the introduction, and restrict that only to work that the researchers actually referenced and used during the research process itself.
Let the reviewers write that bit if they're annoyed at not being cited.
Let the reviewers write that bit if they're annoyed at not being cited.
November 15, 2025 at 8:27 PM
My own extreme version of this: get rid of the related work section or survey parts of the introduction, and restrict that only to work that the researchers actually referenced and used during the research process itself.
Let the reviewers write that bit if they're annoyed at not being cited.
Let the reviewers write that bit if they're annoyed at not being cited.
I don't think bullet points are really a good replacement for the worst excesses of verbiage. I'd really like to see less text that's defending against reviewer attack.
The researcher talking through their thought process free form? Keep that. Tacit knowledge is already hard enough to transfer.
The researcher talking through their thought process free form? Keep that. Tacit knowledge is already hard enough to transfer.
November 15, 2025 at 8:24 PM
I don't think bullet points are really a good replacement for the worst excesses of verbiage. I'd really like to see less text that's defending against reviewer attack.
The researcher talking through their thought process free form? Keep that. Tacit knowledge is already hard enough to transfer.
The researcher talking through their thought process free form? Keep that. Tacit knowledge is already hard enough to transfer.
It helps to have already dealt with the weird fact that for all sorts of systems in statistical physics mean field theory becomes *exact* above 4d (and wrapping your head around corresponding tricks to do perturbation theory to approximate 3d solutions as d=4-epsilon)
November 11, 2025 at 2:50 PM
It helps to have already dealt with the weird fact that for all sorts of systems in statistical physics mean field theory becomes *exact* above 4d (and wrapping your head around corresponding tricks to do perturbation theory to approximate 3d solutions as d=4-epsilon)
It's easier to count down from infinity. Only sort of joking - there's some sort of mental texture to 'all directions you can go are totally unrelated to each other unless you try to follow close, all points can fit adjacently to each other unless you try to stretch them'
November 11, 2025 at 2:46 PM
It's easier to count down from infinity. Only sort of joking - there's some sort of mental texture to 'all directions you can go are totally unrelated to each other unless you try to follow close, all points can fit adjacently to each other unless you try to stretch them'
My guess (if there is such a force, that is) would be on L/2c...
November 8, 2025 at 8:00 PM
My guess (if there is such a force, that is) would be on L/2c...
If so, you could then do the flipped version of the experiment and ask, if there were a second mirror halfway down the cavity, how long before removing that mirror would it be for the force on the terminal mirror to reach its equilibrium value? Instant, L/2c, L/c, or 3L/2c?
November 8, 2025 at 7:58 PM
If so, you could then do the flipped version of the experiment and ask, if there were a second mirror halfway down the cavity, how long before removing that mirror would it be for the force on the terminal mirror to reach its equilibrium value? Instant, L/2c, L/c, or 3L/2c?
This would make me wonder, in the quantum case, does the mirror at the end of an incompatible cavity experience a force from the presence of an excited atom within the cavity which cannot emit?
November 8, 2025 at 7:55 PM
This would make me wonder, in the quantum case, does the mirror at the end of an incompatible cavity experience a force from the presence of an excited atom within the cavity which cannot emit?
I wonder if there's a purely classical (relativistic) analogue in the case of things with Lagrange multipliers like pressure or stress. E.g. you have a pressurized hydraulic system 10 light seconds long, under compression. When you poke a hole in it, material begins to leave the system immediately.
November 8, 2025 at 7:55 PM
I wonder if there's a purely classical (relativistic) analogue in the case of things with Lagrange multipliers like pressure or stress. E.g. you have a pressurized hydraulic system 10 light seconds long, under compression. When you poke a hole in it, material begins to leave the system immediately.
I guess one question would be - when people make multimodal models, do they start from unimodal pretraining or do they have to start from scratch on all modalities at once in order for the input to not be too concretely bound to a particular modality? Has that changed with model size/data size?
November 4, 2025 at 1:54 AM
I guess one question would be - when people make multimodal models, do they start from unimodal pretraining or do they have to start from scratch on all modalities at once in order for the input to not be too concretely bound to a particular modality? Has that changed with model size/data size?
For example in Anthropic's recent blog post, they chose to inject abstract concepts into a fairly deep layer - 2/3rds through the architecture! That may have just been an arbitrary choice, but maybe it wouldn't have worked on an earlier layer?
November 4, 2025 at 1:40 AM
For example in Anthropic's recent blog post, they chose to inject abstract concepts into a fairly deep layer - 2/3rds through the architecture! That may have just been an arbitrary choice, but maybe it wouldn't have worked on an earlier layer?
I do not know any studies about third party spectator effects though - e.g. does a third party observing two others in conflict end up with more nuanced views or does it drive them to pick a side. Seems worth looking into if anyone has tested it...
November 1, 2025 at 3:25 PM
I do not know any studies about third party spectator effects though - e.g. does a third party observing two others in conflict end up with more nuanced views or does it drive them to pick a side. Seems worth looking into if anyone has tested it...
If people are forced into conflicting interaction without already having a stake in their relationship with the other speaker, they largely just entrench. Also, the persuasion effects when you do first humanize start local and tied to that relationship. I'll have to find the specific papers again.
November 1, 2025 at 3:18 PM
If people are forced into conflicting interaction without already having a stake in their relationship with the other speaker, they largely just entrench. Also, the persuasion effects when you do first humanize start local and tied to that relationship. I'll have to find the specific papers again.
The research I've read on polarization suggests that it's not 'exposure to other ideas' that does it, it's humanizing holders of those other ideas by first interacting with them in ways that avoid the conflicting parts. Eating a meal together with low stakes conversation, playing a sport together
November 1, 2025 at 3:18 PM
The research I've read on polarization suggests that it's not 'exposure to other ideas' that does it, it's humanizing holders of those other ideas by first interacting with them in ways that avoid the conflicting parts. Eating a meal together with low stakes conversation, playing a sport together
Like a lot of AI things, the phenomenon of note is actually how *we* behave given scale, not how *it* behaves. 'Slop' is the manifestation of the human desires to abuse communication as a manipulative and insincere end, magnified by industrial-scale automation.
October 31, 2025 at 3:07 PM
Like a lot of AI things, the phenomenon of note is actually how *we* behave given scale, not how *it* behaves. 'Slop' is the manifestation of the human desires to abuse communication as a manipulative and insincere end, magnified by industrial-scale automation.
The public web getting choked with low-information LLM-generated blogs filled with the worst of the sycophantic, condescending, 5 paragraph essay style outputs is just a totally different beast than the personal interaction asking an LLM to explain KAM theory with probing and clarifying questions.
October 31, 2025 at 3:05 PM
The public web getting choked with low-information LLM-generated blogs filled with the worst of the sycophantic, condescending, 5 paragraph essay style outputs is just a totally different beast than the personal interaction asking an LLM to explain KAM theory with probing and clarifying questions.
The adoption of the terms as a generic negative valence exclamation masks that there's a hierarchy of intentionality (with things you participate in personally for your own goals always landing higher on that chart), and that some public spaces have been flooded with low intent adversarial stuff.
October 31, 2025 at 3:05 PM
The adoption of the terms as a generic negative valence exclamation masks that there's a hierarchy of intentionality (with things you participate in personally for your own goals always landing higher on that chart), and that some public spaces have been flooded with low intent adversarial stuff.
I'm very interested in this kind of vector injection experiment. Is this from a particular paper?
I have this kind of half-baked idea that explicitly fine-tuning for introspection could help in bridging to new modalities, using the introspective trace as a feedback signal.
I have this kind of half-baked idea that explicitly fine-tuning for introspection could help in bridging to new modalities, using the introspective trace as a feedback signal.
October 29, 2025 at 9:45 PM
I'm very interested in this kind of vector injection experiment. Is this from a particular paper?
I have this kind of half-baked idea that explicitly fine-tuning for introspection could help in bridging to new modalities, using the introspective trace as a feedback signal.
I have this kind of half-baked idea that explicitly fine-tuning for introspection could help in bridging to new modalities, using the introspective trace as a feedback signal.
Which means that, yes, a prosthetic for reasoning would be very attractive, but not in a 'this solves the problem' way; instead, in a 'this is actually the problem' way.
October 26, 2025 at 4:21 PM
Which means that, yes, a prosthetic for reasoning would be very attractive, but not in a 'this solves the problem' way; instead, in a 'this is actually the problem' way.
That kind of break is very nasty. If I know that I am bad at spotting tricks of reasoning, the strategy that remains to me is to find someone I believe I share interests with, then when they say 'that's wrong' believe them over my own critical thinking. Because I have been taught not to trust myself
October 26, 2025 at 4:18 PM
That kind of break is very nasty. If I know that I am bad at spotting tricks of reasoning, the strategy that remains to me is to find someone I believe I share interests with, then when they say 'that's wrong' believe them over my own critical thinking. Because I have been taught not to trust myself
Logic can carry us to counterintuitive yet true places, but that means it can be abused to tricking people into doing things against their interests if one is better at hiding the trick than they are at spotting it. If people live with that experience, they might come to mistrust logic itself.
October 26, 2025 at 4:18 PM
Logic can carry us to counterintuitive yet true places, but that means it can be abused to tricking people into doing things against their interests if one is better at hiding the trick than they are at spotting it. If people live with that experience, they might come to mistrust logic itself.