Arseny Khakhalin
@khakhalin.bsky.social
Data Scientist in Berlin
Former Bard College prof
For my after-work alter-ego, see @elstersen.bsky.social
Support Ukraine! 🇺🇦
Former Bard College prof
For my after-work alter-ego, see @elstersen.bsky.social
Support Ukraine! 🇺🇦
Pinned
Arseny Khakhalin
@khakhalin.bsky.social
· Apr 14
Minimalistic bifurcation :)
import matplotlib.pyplot as plt
f = lambda x,k,a: (x if k==0 else f(a*x*(1-x), k-1, a))
y = [f(.21, 17 + i % 19, 2.5+1.5*i/10000) for i in range(10000)]
plt.plot(y,'.',markersize=1);
import matplotlib.pyplot as plt
f = lambda x,k,a: (x if k==0 else f(a*x*(1-x), k-1, a))
y = [f(.21, 17 + i % 19, 2.5+1.5*i/10000) for i in range(10000)]
plt.plot(y,'.',markersize=1);
Reposted by Arseny Khakhalin
Ilya!!!!
www.dwarkesh.com/p/ilya-sutsk...
www.dwarkesh.com/p/ilya-sutsk...
Ilya Sutskever – We're moving from the age of scaling to the age of research
“These models somehow just generalize dramatically worse than people. It's a very fundamental thing.”
www.dwarkesh.com
November 25, 2025 at 6:07 PM
Ilya!!!!
www.dwarkesh.com/p/ilya-sutsk...
www.dwarkesh.com/p/ilya-sutsk...
Ugh I thought you were just busy writing books and running labs :)
November 25, 2025 at 5:58 PM
Ugh I thought you were just busy writing books and running labs :)
Reposted by Arseny Khakhalin
Good thread about dimensionality here
Y’all are reading this paper in the wrong way.
We love to trash dominant hypothesis, but we need to look for evidence against the manifold hypothesis elsewhere:
This elegant work doesn't show neural dynamics are high D, nor that we should stop using PCA
It’s quite the opposite!
(thread)
We love to trash dominant hypothesis, but we need to look for evidence against the manifold hypothesis elsewhere:
This elegant work doesn't show neural dynamics are high D, nor that we should stop using PCA
It’s quite the opposite!
(thread)
“Our findings challenge the conventional focus on low-dimensional coding subspaces as a sufficient framework for understanding neural computations, demonstrating that dimensions previously considered task-irrelevant and accounting for little variance can have a critical role in driving behavior.”
November 25, 2025 at 5:20 PM
Good thread about dimensionality here
Are you less active than you used to be back before the Events?
And what can we do to make this place better?
And what can we do to make this place better?
Before the monstrous transformation of Twitter to X, it was considered normal to be active on academic Twitter. Now, even though we've migrated to Bluesky ("safe space"), I feel that many ppl (including myself) are less active in posting anything. Have you noticed this trend? Or am I imagining this?
November 25, 2025 at 5:53 PM
Are you less active than you used to be back before the Events?
And what can we do to make this place better?
And what can we do to make this place better?
we need to be the change we want to see in the world... There's no other way. Even this may not be enough, of course, but without forced optimisim, it is guaranteed to not be enough :)
Easily for me to say, I know, but I still think atrproto has a chance.
Easily for me to say, I know, but I still think atrproto has a chance.
November 25, 2025 at 5:51 PM
we need to be the change we want to see in the world... There's no other way. Even this may not be enough, of course, but without forced optimisim, it is guaranteed to not be enough :)
Easily for me to say, I know, but I still think atrproto has a chance.
Easily for me to say, I know, but I still think atrproto has a chance.
Can we have windows in Saturnian flying cities? Probably no, we can't? Or rather, only inwards-facing windows? Despite the magnetosphere and the atmosphere?
November 25, 2025 at 5:36 PM
Can we have windows in Saturnian flying cities? Probably no, we can't? Or rather, only inwards-facing windows? Despite the magnetosphere and the atmosphere?
They are getting better, and they are increasingly refusing to answer. They also learned to mark facts (or rather, knowledge), inference, and wild guesses better. I can still lead them into rarified area where they are not helpful, but they are not failing in ways in which GPT used to fail...
November 25, 2025 at 5:13 PM
They are getting better, and they are increasingly refusing to answer. They also learned to mark facts (or rather, knowledge), inference, and wild guesses better. I can still lead them into rarified area where they are not helpful, but they are not failing in ways in which GPT used to fail...
Cooking is one of the early absolute win use cases, 100/100. I love doing this for example: describe it my cooking plans, and ask: if it were a cheap but good family restaurant, how would you have adjusted this recipe? What should I absolutely do, not skip? What is wrong? What doesn't matter?
November 25, 2025 at 5:07 PM
Cooking is one of the early absolute win use cases, 100/100. I love doing this for example: describe it my cooking plans, and ask: if it were a cheap but good family restaurant, how would you have adjusted this recipe? What should I absolutely do, not skip? What is wrong? What doesn't matter?
tragic, as llms were useful for recipes already a year ago, if not earlier! Primordial chatbots already wouldn't tell you about how the yard of their aunt smelled when she was about to start cooking your search term. Were able to adjust to your fridge content, and provide required/optional analysis!
November 25, 2025 at 5:07 PM
tragic, as llms were useful for recipes already a year ago, if not earlier! Primordial chatbots already wouldn't tell you about how the yard of their aunt smelled when she was about to start cooking your search term. Were able to adjust to your fridge content, and provide required/optional analysis!
If you want to work for a company that makes this world better, you have three choices:
- find one
- found one
- fund one
Unfortunately (3) is not an option for me, (2) is unlikely, so I'm kinda stuck with (1)
- find one
- found one
- fund one
Unfortunately (3) is not an option for me, (2) is unlikely, so I'm kinda stuck with (1)
November 25, 2025 at 4:59 PM
If you want to work for a company that makes this world better, you have three choices:
- find one
- found one
- fund one
Unfortunately (3) is not an option for me, (2) is unlikely, so I'm kinda stuck with (1)
- find one
- found one
- fund one
Unfortunately (3) is not an option for me, (2) is unlikely, so I'm kinda stuck with (1)
Ah, gotcha! Thank you! So it's kinda pre-filtered! Makes sense, sorry for not realizing it earlier! :)
November 25, 2025 at 4:52 PM
Ah, gotcha! Thank you! So it's kinda pre-filtered! Makes sense, sorry for not realizing it earlier! :)
How's the radiation situation? is it tolerable, or is it Mars/Jupiter-style?..
November 25, 2025 at 4:37 PM
How's the radiation situation? is it tolerable, or is it Mars/Jupiter-style?..
I'm still confused about the Caribbean: there are 16 countries there, but your map (beautiful map!) only shows 4. Is it because a threshold on population? Or am I missing something obvious?..
November 25, 2025 at 4:33 PM
I'm still confused about the Caribbean: there are 16 countries there, but your map (beautiful map!) only shows 4. Is it because a threshold on population? Or am I missing something obvious?..
"Letting humans die is moral, as there's no difference between humans and other animals" is somewhere in the same direction for me. I am aware of Jainism etc, it is a coherent worldview! But I don't have the stamina to argue with it. My goal was to object, then disengage out of self-preservation :)
November 25, 2025 at 4:11 PM
"Letting humans die is moral, as there's no difference between humans and other animals" is somewhere in the same direction for me. I am aware of Jainism etc, it is a coherent worldview! But I don't have the stamina to argue with it. My goal was to object, then disengage out of self-preservation :)
I am afraid I have to chicken out of this discussion, sorry, as I'm at risk of getting irrationally angry, and I don't want to. There's a whole bunch of thought systems that are formally rational, like "having kids is amoral, because environment", but I can't engage with them. I get too angry.
November 25, 2025 at 4:11 PM
I am afraid I have to chicken out of this discussion, sorry, as I'm at risk of getting irrationally angry, and I don't want to. There's a whole bunch of thought systems that are formally rational, like "having kids is amoral, because environment", but I can't engage with them. I get too angry.
find the probability of a gridlock p_{n, L} ( r ) as a function of rounabout radius r, where n is the number of buses, and L is the length of each of them
November 25, 2025 at 3:22 PM
find the probability of a gridlock p_{n, L} ( r ) as a function of rounabout radius r, where n is the number of buses, and L is the length of each of them
I guess you could also have a list of news people, repeatedly click "see less of this" on news in your ForYou, and have an Iron Will of Never Liking News :) You'd get the same end-effect. But iron will is hard!
November 25, 2025 at 3:15 PM
I guess you could also have a list of news people, repeatedly click "see less of this" on news in your ForYou, and have an Iron Will of Never Liking News :) You'd get the same end-effect. But iron will is hard!
This entire topic is both awesome, accessible (in formulation), and apparently full of unproven conjectures and best results!!
en.wikipedia.org/wiki/Opaque_...
en.wikipedia.org/wiki/Opaque_...
November 25, 2025 at 3:08 PM
This entire topic is both awesome, accessible (in formulation), and apparently full of unproven conjectures and best results!!
en.wikipedia.org/wiki/Opaque_...
en.wikipedia.org/wiki/Opaque_...
Reposted by Arseny Khakhalin
another thing I like is asking image models 'A cyclops wearing sunglasses'.
This is Nano Banana, the "ultra-realistic" image generation model, responding to that prompt
This is Nano Banana, the "ultra-realistic" image generation model, responding to that prompt
November 25, 2025 at 2:00 PM
another thing I like is asking image models 'A cyclops wearing sunglasses'.
This is Nano Banana, the "ultra-realistic" image generation model, responding to that prompt
This is Nano Banana, the "ultra-realistic" image generation model, responding to that prompt
Reposted by Arseny Khakhalin
Decades after it was raised as an issue in neurophysiology it is nice to see this topic confronted in the ML/neural net community. I just wish the intellectual connections were stronger and more explicit, because we can't sensibly separate ideas from neuro with those in AI.
📍Excited to share that our paper was selected as a Spotlight at #NeurIPS2025!
arxiv.org/pdf/2410.03972
It started from a question I kept running into:
When do RNNs trained on the same task converge/diverge in their solutions?
🧵⬇️
arxiv.org/pdf/2410.03972
It started from a question I kept running into:
When do RNNs trained on the same task converge/diverge in their solutions?
🧵⬇️
November 24, 2025 at 5:10 PM
Decades after it was raised as an issue in neurophysiology it is nice to see this topic confronted in the ML/neural net community. I just wish the intellectual connections were stronger and more explicit, because we can't sensibly separate ideas from neuro with those in AI.
Reposted by Arseny Khakhalin
November 22, 2025 at 3:12 PM
That Volt and the Greens overlap so much feels very true, but also such a waste. I wish they had a color to unite. Or maybe embrace the viridis colormap.
November 25, 2025 at 2:55 PM
That Volt and the Greens overlap so much feels very true, but also such a waste. I wish they had a color to unite. Or maybe embrace the viridis colormap.
Reposted by Arseny Khakhalin
S-Values are much more interpretable than P-values, yet adoption seems near impossible. I wonder what it would take to make the leap? #statssky #episky #rstats #statistics
October 10, 2025 at 3:50 PM
S-Values are much more interpretable than P-values, yet adoption seems near impossible. I wonder what it would take to make the leap? #statssky #episky #rstats #statistics
I got two accounts here: on one I reward news in "For you", in another one I discourage them. It works flawlessly! When in need of news, or when tired of them, just switch the profile :)
November 25, 2025 at 2:50 PM
I got two accounts here: on one I reward news in "For you", in another one I discourage them. It works flawlessly! When in need of news, or when tired of them, just switch the profile :)
But other than that, for other topics, it's the most fun I had in years. And it feels way less toxic than even the old Twitter, which is nice actually!
November 25, 2025 at 2:48 PM
But other than that, for other topics, it's the most fun I had in years. And it feels way less toxic than even the old Twitter, which is nice actually!