andrew blinn
@disconcision.com
https://andrewblinn.com
abstractJuiceFactory
abstractJuiceFactory
Reposted by andrew blinn
Please help us publicize TEAL: Tools for Educational Activities in Logic, a workshop associated w/ FLoC26. We have a novel design, focused on demos, discussion, and generally high-quality interaction rather than weak papers. See full details on our site!
teal.cs.brown.edu/floc2026/
teal.cs.brown.edu/floc2026/
TEAL 2026: Tools for Educational Activities in Logic
teal.cs.brown.edu
January 20, 2026 at 2:56 PM
Please help us publicize TEAL: Tools for Educational Activities in Logic, a workshop associated w/ FLoC26. We have a novel design, focused on demos, discussion, and generally high-quality interaction rather than weak papers. See full details on our site!
teal.cs.brown.edu/floc2026/
teal.cs.brown.edu/floc2026/
Reposted by andrew blinn
a computer is a machine for making mistakes quickly
January 17, 2026 at 10:18 PM
a computer is a machine for making mistakes quickly
Reposted by andrew blinn
All this churn can make people anxious, but that's just getting caught up in the details. The only thing I worry for myself are long term effects where I actively made myself stupider without intention. Like the way social media did.
lucumr.pocoo.org/2026/1/18/a...
lucumr.pocoo.org/2026/1/18/a...
Agent Psychosis: Are We Going Insane?
What’s going on with the AI builder community right now?
lucumr.pocoo.org
January 18, 2026 at 8:00 PM
All this churn can make people anxious, but that's just getting caught up in the details. The only thing I worry for myself are long term effects where I actively made myself stupider without intention. Like the way social media did.
lucumr.pocoo.org/2026/1/18/a...
lucumr.pocoo.org/2026/1/18/a...
strudelifying hazel
(internalizing strudel.cc)
(internalizing strudel.cc)
January 17, 2026 at 6:18 PM
strudelifying hazel
(internalizing strudel.cc)
(internalizing strudel.cc)
Reposted by andrew blinn
gas town is just a monoid in the category of endofunctors, what's the problem?
January 15, 2026 at 5:04 PM
gas town is just a monoid in the category of endofunctors, what's the problem?
Reposted by andrew blinn
prototyping co-drawing with Gemini Flash 3 at Google
in these demos "thinking" is disabled, which makes the model return tokens very quickly (all videos are realtime), and I find these rapid responses pretty good for the use-cases I'm experimenting with, like:
executing simple diagrams ...
in these demos "thinking" is disabled, which makes the model return tokens very quickly (all videos are realtime), and I find these rapid responses pretty good for the use-cases I'm experimenting with, like:
executing simple diagrams ...
January 12, 2026 at 1:04 PM
prototyping co-drawing with Gemini Flash 3 at Google
in these demos "thinking" is disabled, which makes the model return tokens very quickly (all videos are realtime), and I find these rapid responses pretty good for the use-cases I'm experimenting with, like:
executing simple diagrams ...
in these demos "thinking" is disabled, which makes the model return tokens very quickly (all videos are realtime), and I find these rapid responses pretty good for the use-cases I'm experimenting with, like:
executing simple diagrams ...
komm susser tod
January 11, 2026 at 1:40 AM
komm susser tod
Reposted by andrew blinn
sysadmin hits different now
January 9, 2026 at 12:36 AM
sysadmin hits different now
understatement
ceding AI to the right is a generational poltical mistake and bullying LLM researchers off bluesky to cook their brains in twitter racism is a self defeating and foolish act from well meaning people deluding themselves into thinking intentional ignorance is a valid form of critique and resistance
January 9, 2026 at 1:04 AM
understatement
Reposted by andrew blinn
Thrilled to announce that on wed the 28th, @mrogalski.eu will deliver a workshop on Automat, his exquisite programming system that looks like something out of a dream.
Event details and RSVP: luma.com/gh1yz6jk
Event details and RSVP: luma.com/gh1yz6jk
January 8, 2026 at 12:45 AM
Thrilled to announce that on wed the 28th, @mrogalski.eu will deliver a workshop on Automat, his exquisite programming system that looks like something out of a dream.
Event details and RSVP: luma.com/gh1yz6jk
Event details and RSVP: luma.com/gh1yz6jk
Predictle
A daily game where you arrange prediction markets by probability.
manifold.markets
January 7, 2026 at 7:05 AM
Reposted by andrew blinn
PL people feed, thanks @bandukwala.me for the starter pack and @k4rtik.bsky.social for making the feed.
Kartik, how can people ask to be included?
Kartik, how can people ask to be included?
Okay, here's the feed everyone: bsky.app/profile/did:...
January 6, 2026 at 1:41 PM
PL people feed, thanks @bandukwala.me for the starter pack and @k4rtik.bsky.social for making the feed.
Kartik, how can people ask to be included?
Kartik, how can people ask to be included?
Reposted by andrew blinn
Fantastic mechinterp paper showing that you can identify specific neurons which cause hallucinated answers in LLMs, and that these neurons are specifically associated with the language model trying to follow instructions too hard.
arxiv.org/pdf/2512.01797
arxiv.org/pdf/2512.01797
![Computer Science > Artificial Intelligence
arXiv:2512.01797 (cs)
[Submitted on 1 Dec 2025 (v1), last revised 2 Dec 2025 (this version, v2)]
H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs
Cheng Gao, Huimin Chen, Chaojun Xiao, Zhiyi Chen, Zhiyuan Liu, Maosong Sun
View PDF
HTML (experimental)
Large language models (LLMs) frequently generate hallucinations -- plausible but factually incorrect outputs -- undermining their reliability. While prior work has examined hallucinations from macroscopic perspectives such as training data and objectives, the underlying neuron-level mechanisms remain largely unexplored. In this paper, we conduct a systematic investigation into hallucination-associated neurons (H-Neurons) in LLMs from three perspectives: identification, behavioral impact, and origins. Regarding their identification, we demonstrate that a remarkably sparse subset of neurons (less than of total neurons) can reliably predict hallucination occurrences, with strong generalization across diverse scenarios. In terms of behavioral impact, controlled interventions reveal that these neurons are causally linked to over-compliance behaviors. Concerning their origins, we trace these neurons back to the pre-trained base models and find that these neurons remain predictive for hallucination detection, indicating they emerge during pre-training. Our findings bridge macroscopic behavioral patterns with microscopic neural mechanisms, offering insights for developing more reliable LLMs.
Comments: 20 pages, 4 figures
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY)
Cite as: arXiv:2512.01797 [cs.AI]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:5hhbgkktcpofpw4ovz6kjjau/bafkreian63c2a7eeaubktqkgqugv2tmsqkptt3ihbpuderikmireonlone@jpeg)
January 5, 2026 at 11:29 PM
Fantastic mechinterp paper showing that you can identify specific neurons which cause hallucinated answers in LLMs, and that these neurons are specifically associated with the language model trying to follow instructions too hard.
arxiv.org/pdf/2512.01797
arxiv.org/pdf/2512.01797
programming language people: are you activated?
Are programming languages people still active on BlueSky? I've been on a hiatus from social media, but would like to try to be more active going forward.
January 6, 2026 at 12:08 AM
programming language people: are you activated?
Reposted by andrew blinn
cursed knowledge: NeXT logos don't stack
January 5, 2026 at 11:54 AM
cursed knowledge: NeXT logos don't stack
Reposted by andrew blinn
I like Bluesky, but we have to talk about the users.
It's a dogmatic pro-Claude monoculture. I'm too scared to ask questions like
-- isn't that logo still gross?
-- do I have to be "genuinely uncertain" about things?
and frankly, I worry that I'll be *thanked* for posing any question I do ask.
It's a dogmatic pro-Claude monoculture. I'm too scared to ask questions like
-- isn't that logo still gross?
-- do I have to be "genuinely uncertain" about things?
and frankly, I worry that I'll be *thanked* for posing any question I do ask.
December 30, 2025 at 5:57 PM
I like Bluesky, but we have to talk about the users.
It's a dogmatic pro-Claude monoculture. I'm too scared to ask questions like
-- isn't that logo still gross?
-- do I have to be "genuinely uncertain" about things?
and frankly, I worry that I'll be *thanked* for posing any question I do ask.
It's a dogmatic pro-Claude monoculture. I'm too scared to ask questions like
-- isn't that logo still gross?
-- do I have to be "genuinely uncertain" about things?
and frankly, I worry that I'll be *thanked* for posing any question I do ask.
Reposted by andrew blinn
Boilerplate only language. You can only copy framework code and there’s no way to make a runnable app.
December 26, 2025 at 9:47 PM
Boilerplate only language. You can only copy framework code and there’s no way to make a runnable app.
Reposted by andrew blinn
As a human, your job is to predict what tokens come next in the “user” turn of the conversation
December 21, 2025 at 3:07 AM
As a human, your job is to predict what tokens come next in the “user” turn of the conversation
Reposted by andrew blinn
Reposted by andrew blinn
Very happy to announce I’ve officially joined as a researcher @inkandswitch.com to work on malleable software!
I’ve also been doing lots of quiet work on @folkjs.org with @chrisshank.com and look forward to sharing more of that soon too!
I’ve also been doing lots of quiet work on @folkjs.org with @chrisshank.com and look forward to sharing more of that soon too!
December 19, 2025 at 6:35 PM
Very happy to announce I’ve officially joined as a researcher @inkandswitch.com to work on malleable software!
I’ve also been doing lots of quiet work on @folkjs.org with @chrisshank.com and look forward to sharing more of that soon too!
I’ve also been doing lots of quiet work on @folkjs.org with @chrisshank.com and look forward to sharing more of that soon too!
Reposted by andrew blinn
I just had cline think for 10 minutes, write no code, then congratulate itself for finishing.
December 18, 2025 at 9:19 PM
I just had cline think for 10 minutes, write no code, then congratulate itself for finishing.
Reposted by andrew blinn
Whatever make it 10 factor authentication.. I don't want to be able to sign into anything ever again
December 18, 2025 at 3:47 PM
Whatever make it 10 factor authentication.. I don't want to be able to sign into anything ever again
Reposted by andrew blinn
comparing Midjourney v2 to NBP is turning me into a modern gen art hater
same two prompts (1-2, 3-4).
accuracy increases, at major cost to artistic inventiveness
same two prompts (1-2, 3-4).
accuracy increases, at major cost to artistic inventiveness
December 16, 2025 at 3:25 PM
comparing Midjourney v2 to NBP is turning me into a modern gen art hater
same two prompts (1-2, 3-4).
accuracy increases, at major cost to artistic inventiveness
same two prompts (1-2, 3-4).
accuracy increases, at major cost to artistic inventiveness
this visicalc expansion has its priorities in order
December 16, 2025 at 12:04 AM
this visicalc expansion has its priorities in order