LLMs read write neural network weights
16 experts across 6 network communities independently surfaced this.
16 experts
6 communities
1 sources clustered
Research & technical analysis
7 experts
Evidence, methods and technical implications.
“This is fascinating. Very large models can read the weights of small models and intuit what those models were trained to do, and make direct weight changes to improve them. rikiyat.github.io/blog/weight-...”
4 experts discussed this · 5 posts
Ethan Mollick: If this holds up, it is big: "We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x …
Antonino Greco: If this holds up, it is big: "We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x …
Thomas Dietterich: Can someone provide a plausible mechanism for this? Has GPT-6 Astra been trained on neural network weight (or activation) interpretation?
Open the full discussion →
AI-generated language comprehension difficulty paper
7 experts across 4 network communities independently surfaced this.
7 experts
4 communities
1 sources clustered
Research & technical analysis
2 experts
Evidence, methods and technical implications.
“Curious to dig into the analysis here arxiv.org/abs/2610.29162”
4 experts discussed this · 4 posts
Shubhendu Trivedi: Curious to dig into the analysis here arxiv.org/abs/2610.29162
Ted Underwood: Curious to dig into the analysis here arxiv.org/abs/2610.29162
Ted Underwood: Intuitively the qualities they list match my assessment, and the rationale passes a sniff test
Open the full discussion →
6 experts across 4 network communities independently surfaced this.
6 experts
4 communities
1 sources clustered
Building & implementation
2 experts
How teams are shipping and applying it.
“The exponential time hypothesis is apparently false: openai.com/index/buildi...”
Research & technical analysis
1 expert
Evidence, methods and technical implications.
“If this result survives independent scrutiny, AI will have done something more consequential than outperforming humans on a benchmark. It will have overturned a foundational scientific assumption held by experts for decades.”
Opportunity & adoption
1 expert
New capabilities, benefits and practical upside.
“Oh this is neat. ETH is assumed to be true in a bunch of theoretical CS work, so there's quite a lot of opportunities to revisit priors and maybe have a few faster foundational algorithms down the road openai.com/index/buildi...”
4 experts discussed this · 11 posts
Colin: It is interesting watching them try out all kinds of different stuff to try to stop everyone from getting mad. I swear, everyone getting really mad at a tech company really works. It’s one of the f…
tweety fish: It is interesting watching them try out all kinds of different stuff to try to stop everyone from getting mad. I swear, everyone getting really mad at a tech company really works. It’s one of the f…
Colin: Deep down the people who work there just want to be held in high esteem. They want everyone to say “thank you for changing the world with your wonderful products” and they don’t want to be yelled at.
Open the full discussion →
emergent symbolic structure neural networks paper
3 experts across 3 network communities independently surfaced this.
3 experts
3 communities
1 sources clustered
“Work done with @paulsoulos.bsky.social, @tallinzen.bsky.social, and Paul Smolensky For much more, see our paper! arxiv.org/abs/2608.29530 12/12”
4 experts discussed this · 30 posts
Tom McCoy: 🤖🧠NEW PAPER🧠🤖 (The result of an 8-year project!) LLMs seem very different from symbolic systems. Yet LLMs excel in symbolic domains (e.g., language/code/math). How do they do it? Our finding: LLM r…
Tom McCoy: Paper: arxiv.org/abs/2608.29530 We consider a variety of neural networks that perform seemingly-symbolic tasks, from small-scale models trained on simple list-manipulation tasks to LLMs performing …
Tom McCoy: Our hypothesis: The networks succeed on these tasks by implicitly building symbolic structure in vector space. What makes this hypothesis testable is the Tensor Product Representation (TPR) formali…
Open the full discussion →
ODE sampler discrete diffusion paper
2 experts are actively discussing the implications.
2 experts
1 community
1 sources clustered
“Looks cool arxiv.org/abs/2610.40618”
2 experts discussed this · 2 posts
Shubhendu Trivedi: Looks cool arxiv.org/abs/2610.40618
Ramon Astudillo: Looks like continuous-time text diffusion, but better
Ramon Astudillo: Looks cool arxiv.org/abs/2610.40618
Open the full discussion →
collective motion moshers metal concerts paper
4 experts are actively discussing the implications.
2 experts
2 communities
1 sources clustered
“Actually can't find it. It was from 10-13 years ago. Using LMs gave this, which I also remember, but it's not this one. This is too simple and whimsical arxiv.org/abs/1302.1886”
4 experts discussed this · 25 posts
Shubhendu Trivedi: One reason I have basically stopped scrolling social media (down to perhaps 15-20 minutes at max) is because I can't stand the sight of the endless procession of anons.
Shubhendu Trivedi: fast-moving collective dynamics, which it does, even if the participants didn't intend anything beyond the local venting (and don't even care about loosely coordinated adjacent communities). Take a…
Shubhendu Trivedi: On some days I might indeed doom scroll. But it's just the same thing. Big tech anon, frontier anon, reactionary anon, prof anon, grad anon, eigen anons, ai anons, anti-ai anons, this anon, that an…
Open the full discussion →
anytime-valid envelopes iterative algorithms paper
2 directory members surfaced this signal.
2 experts
2 communities
1 sources clustered
“Kind of interesting: a general framework for constructing data-adaptive, anytime-valid error bounds for stochastic iterative algorithms (reducing the analysis to time-uniform concentration of predictably weighted noise sums) arxiv.org/abs/2610.23570”
“A study presents a method for creating anytime-valid envelopes in stochastic algorithms, enhancing data-adaptive stopping strategies. This boosts convergence guarantees and broadens applications in optimization, aiding real-time decision-making and machine …”
softmax cross-entropy fix paper
2 directory members surfaced this signal.
2 experts
1 community
1 sources clustered
Research & technical analysis
2 experts
Evidence, methods and technical implications.
“This definitely couldn't have been anything but an "agent result." arxiv.org/abs/2610.31642”
Established
AI field signal
Signal
14d ago
2 directory members surfaced this signal.
2 experts
2 communities
1 sources clustered
“How do max and softmax layers approximate convex functions? Supporting planes give a constructive proof, error bounds, and a tradeoff between neuron count and temperature. https://www.gpeyre.com/blog/2026/10/24/convex-max-softmax-approximation/”
Established
AI field signal
Signal
5d ago
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
“It remains to be verified. The people looking into includes someone who stated the hypothesis originally. I have been very critical of their math dumps, tbc. But given that another group also gave an algorithm github.com/anthropics/f... it should be easy to…”
photoacoustic brain imaging skulls paper
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
equivariance convolution compact groups generalization paper
2 experts are actively discussing the implications.
1 expert
1 community
1 sources clustered
“My favourite paper is still this arxiv.org/abs/1802.03690 it took around 2.5 years to conceive and write..”
2 experts discussed this · 9 posts
Shubhendu Trivedi: We weren't sure what to do with it. The more ambitious problem we wanted to address (for which the exact problem I have stated in a thread here some months ago), had eluded us. But one of the autho…
Shubhendu Trivedi: Visited my own scholar profile after a few months to link a paper to someone. Embarrassed to find that a paper that we put online (and then submitted to TMLR where it was accepted) but were really …
Shubhendu Trivedi: My favourite paper is still this arxiv.org/abs/1802.03690 it took around 2.5 years to conceive and write..
Open the full discussion →
Established
AI field signal
Signal
16d ago
2 experts are actively discussing the implications.
1 expert
1 community
1 sources clustered
“Otherwise, you will have more of this huggingface.co/datasets/ula... there person on twitter who for the purpose of "advancing math" has even scraped questions people discussed in Oberwolfach sort of meetings. Imagine a descent on this. Basically would end …”
2 experts discussed this · 16 posts
Shubhendu Trivedi: I used to say the same thing to my advisor, and faculty collaborators (I graduated in 2018). Even used to say the same for poorly written papers and books that nonetheless had something unique abou…
Shubhendu Trivedi: The analogy I used to draw was akin to a denoising autoencoder, with the input a signal from a mind that had a reasonably-well-formed and unified conception, but presented it idiosyncratically, or …
Shubhendu Trivedi: With humongous AI proofs at least at the moment, I don't feel that. Consider the latest dump, only a very small fraction (in additive combinatorics, TCS, and Number Theory) had some expert whetting…
Open the full discussion →
martingale posterior Cox processes paper
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
“Martingale posterior Cox processes arxiv.org/abs/2610.24787”
Blackwell Bellman optimality paper
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
“Also interesting arxiv.org/abs/2510.13476”
LLM reliable inference cost learning
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
“A paper with an interesting combination of mechanism design and online learning. You have a reverse 2nd price auction to incentivize truthful cost reporting, and a learning component which is meant to discover which providers can reliably satisfy the qualit…”
1 directory member surfaced this signal.
1 expert
1 community
1 sources clustered
