pretraining LMs on historical text paper
3 experts across 2 network communities independently surfaced this.
3 experts
2 communities
1 sources clustered
“For fans of Talkie-1930 and all the methodological questions raised by historical models, here's a new entrant into the field, TypewriterLM, trained up to 1913. Corpus, instruction-tuning datasets, and event dataset are released. arxiv.org/abs/2606.02991”
“A study presents "TYPEWRITERLM," a 7.24B language model pre-trained on texts before 1913. It achieves strong language understanding without biases. The "HISTORY-EVENT" benchmark tests historical accuracy and model performance, aiding future research. https:…”
7 experts discussed this · 38 posts
Ted Underwood: For fans of Talkie-1930 and all the methodological questions raised by historical models, here's a new entrant into the field, TypewriterLM, trained up to 1913. Corpus, instruction-tuning datasets,…
Ted Underwood: Also the model itself. Inter alia, I am glad that I decided to work on a benchmark rather than a model.
Ted Underwood: Not as far as I know, and I haven't set it up locally yet, though that would be easy at this size.
Open the full discussion →
OpenAI angers mathematicians again
11 experts across 5 network communities independently surfaced this.
11 experts
5 communities
1 sources clustered
Research & technical analysis
1 expert
Evidence, methods and technical implications.
“Link: github.com/openai/math P.S.1: This reminds me of the Chinese room argument, though only indirectly. P.S.2: I imagine a few people at OpenAI have looked at the results. Also maybe some other people have looked at some papers since ~1h ago. These are o(…”
Questions & unknowns
1 expert
What remains unresolved or contested.
“OpenAI is working on at least 722 (so far) math manuscripts, all on GitHub in various stages of verificantion github.com/openai/math Why? I guess because they can?”
Context & explanation
1 expert
Background, chronology and why it matters.
“Oh my god, OpenAI just made the single largest advancement in mathematics and number theory in modern history. They released a quasi reimann hypothesis. This is just one of 722 proofs and advancements they released all at once. github.com/openai/math”
3 experts discussed this · 27 posts
Dashiell: I don't think there's a world where mathematicians aren't pissed about this. 1. Dumping everything is what the advisory group said 2. Just yesterday people were mad that Ant paid two experts to dig…
Naomi Saphra: I think (2) is actually good. It is weird, and makes mathematicians into some kind of benchmark evaluation annotator which they don’t like, but it’s just a hard time to be a mathematician. Destroyi…
Naomi Saphra: One big consequence here is you can’t use the duration of an open problem as a sign of difficulty/importance. But at least right now, there’s no way to develop interesting open problems through RLV…
Open the full discussion →
AGMAI September 29 general content
5 experts across 4 network communities independently surfaced this.
5 experts
4 communities
1 sources clustered
“It's possible I'm wrong, and they didn't actually closely follow the advice given by that board (AGMAI). Depends on how well we think they did on part 2B here.”
“I think the recommendations are the only thing that has been published agmai.org/general-sep29/”
4 experts discussed this · 47 posts
SE Gyges: i would like everyone complaining to state what they would have preferred to have happened here.
SE Gyges: why? and, like, i guess they could but i assume they tried basically all of them
SE Gyges: i'mma keep it real with you i don't remember who i was quote posting and did not intend it as a dunk on them, i am simply displeased with the discourse overall
Open the full discussion →
neural networks already clean essay
11 experts across 6 network communities independently surfaced this.
11 experts
6 communities
1 sources clustered
Research & technical analysis
2 experts
Evidence, methods and technical implications.
“Quite a claim in this new work from Goodfire. Among many points, they argue that some of the main examples of superposition and polysemanticity in large models were misunderstood. This definitely deserves some deep, rigorous peer review. already-clean.goodf…”
4 experts discussed this · 14 posts
Ryan Moulton: Haven't read this yet, but I'm posting it because it seems important and I want to. already-clean.goodfire.com
Ryan Moulton: Claim, as they say.
Ryan Moulton: @norvid-studies.bsky.social some sparkling fresh juicy claims for you.
Open the full discussion →
New
AI field signal
Signal
20h ago
⚡ 3269 h early
10 experts across 5 network communities independently surfaced this.
10 experts
5 communities
1 sources clustered
Policy & governance
4 experts
Rules, institutions and accountability.
“Maria hails back to the last #attention2026 conference in which Megan Garcia spoke about her lawsuit against Character.ai, which led her son to suicide. Ressa notes that Character.ai just went bankrupt. There are real ways to hold these systems accountable …”
Building & implementation
1 expert
How teams are shipping and applying it.
“character.ai did build their own LLM”
Research & technical analysis
1 expert
Evidence, methods and technical implications.
“Character.ai was gutted in 2024 for $2.5 billion so Google could rehire one of the original authors of the Attention Is All You Need paper, but it hung in there for two years under Google’s “licensing” deal. This is the first of many I’m afraid. This is all…”
2 experts discussed this · 12 posts
thebes: am sick so spent the night with opus 5.5 working on a chart (inspired by the one from lem's golem xiv) how did we got here / model genealogy / the heroic age of language modeling
Grace: am sick so spent the night with opus 5.5 working on a chart (inspired by the one from lem's golem xiv) how did we got here / model genealogy / the heroic age of language modeling
thebes: i think it's too late for the period i was focused on (i carried the branches to the modern day but only if they started in the period in question)
Open the full discussion →
Established
AI field signal
Resource
12d ago
⚡ 87 h early
arxiv preprint archive resource
8 experts across 3 network communities independently surfaced this.
8 experts
3 communities
1 sources clustered
Research & technical analysis
5 experts
Evidence, methods and technical implications.
“arxiv.org, doi.org, biorxiv.org, medrxiv.org, pubmed.ncbi.nlm.nih.gov, nature.com, science.org, pnas.org, semanticscholar.org, openreview.net”
Questions & unknowns
1 expert
What remains unresolved or contested.
“Just me or does @ChatGPT search the most random websites no matter the query? If I ask for the best bagel, it will “Searching ” + {https://t.co/vEQ6I1WmCF, https://t.co/5JR1Db9NqB, …}”
4 experts discussed this · 6 posts
Ted Underwood: I just eliminated 22 tabs by doing this: osascript -e 'tell application "Google Chrome" to get URL of every tab of front window' | tr ',' '\n' | grep arxiv.org list of urls -> LLM -> annotated read…
Ted Underwood: did I come up with that idea by myself? no I did not
Karandeep Singh: Any suggestions for the other 478 tabs?
Open the full discussion →
AI-generated language comprehension difficulty paper
4 experts are actively discussing the implications.
2 experts
2 communities
1 sources clustered
“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 →
AI impact on creative class essay
4 experts across 2 network communities independently surfaced this.
4 experts
2 communities
1 sources clustered
“Have seen a bunch of people marvel at How Much Concert Tickets Cost These Days but per @JosephPolitano it seems to be largely the mirror image of how much cheaper it’s become to listen to recorded music. https://t.co/rtZwZO5alY https://t.co/6EdBB0PwS2”
“It's just a parade of bad news for jobs in the media and entertainment industries nowadays, and I don't know if anything is going to stem the bleeding anytime soon www.apricitas.io/p/ai-and-the...”
2 experts discussed this · 11 posts
Joey Politano🏳️🌈: Thanks for making it to the end of the thread! If you want to read more, here’s the full article where I give more detailed data on AI & arts job loss While you’re there, consider subscribing to my…
Joey Politano🏳️🌈: Overall, media has always been a small share of the economy—the effect of AI art will be primarily cultural & social rather than economic Still, current arts displacement is a microcosm of disrupti…
Joey Politano🏳️🌈: Still, it's extremely difficult to predict how tech will interact with arts businesses Take US novel sales, which had been declining for decades up to 2020. Would anyone have anticipated that short…
Open the full discussion →
Last Translation Benchmark tool
3 experts across 2 network communities independently surfaced this.
3 experts
2 communities
1 sources clustered
Research & technical analysis
2 experts
Evidence, methods and technical implications.
“Sad to see no African representation in the last translation benchmark, despite LLMs and MT being so bad If you know who would be interested or are interested yourself in contributing(coauthoring), please do and share last-translation-benchmark.vilda.net Vi…”
3 experts discussed this · 9 posts
Vilém Zouhar: There are many things machine translation still can't do. Help us steer the next direction by contributing hard-to-translate inputs (and be on a cool paper).
Vilém Zouhar: Multiple things made us start this effort. Typical translation benchmarks are.. ...oftentimes trivial or saturated (so they can't be used for guiding the next steps in the field) ...not evaluatable…
Vilém Zouhar: In the Last Translation Benchmark we solve both by: - collecting hard-to-translate inputs (texts, images, audios) - requiring human-readable "verification rules", which enable provable evaluation o…
Open the full discussion →
cyber risk AI discourse critique essay
4 experts across 3 network communities independently surfaced this.
4 experts
3 communities
1 sources clustered
“In case you missed my hot take yesterday. www.interconnects.ai/p/the-cyber-...”
“The Cyber Risk Discourse is Broken www.interconnects.ai/p/the-cyber-ri… #AI #cybersecurity”
3 experts discussed this · 3 posts
Nathan Lambert: In case you missed my hot take yesterday. www.interconnects.ai/p/the-cyber-...
Ted Underwood: In case you missed my hot take yesterday. www.interconnects.ai/p/the-cyber-...
Alex Hern: From stage to screen
Open the full discussion →
LLM story diversity diagnosis paper
5 experts across 3 network communities independently surfaced this.
5 experts
3 communities
1 sources clustered
Research & technical analysis
5 experts
Evidence, methods and technical implications.
“In the "Limitations" section they acknowledge that their study was monolingual, but say a multilingual generalization would be interesting. arxiv.org/abs/2605.26492”
2 experts discussed this · 2 posts
Ted Underwood: was reading this and enjoying it — and then enjoyed it more when it turned out to be a paper by Sil Hamilton and @dmimno.bsky.social
David Mimno: was reading this and enjoying it — and then enjoyed it more when it turned out to be a paper by Sil Hamilton and @dmimno.bsky.social
David Mimno: The preprint is arxiv.org/abs/2605.26492, and @srhm.ca is working on a post!
Open the full discussion →
Established
AI field signal
Signal
19d ago
3 experts are actively discussing the implications.
2 experts
2 communities
1 sources clustered
“It is wild that this framework: compression progress as the basis for beauty, curiosity, art, science, and even jokes, was laid out in full formal detail in a Japanese journal 15 years ago.”
3 experts discussed this · 5 posts
hardmaru: Recently learned that Schmidhuber’s Formal Theory of Fun and Creativity, covering compression progress and intrinsic motivation and their relationship to beauty, curiosity, art, science, music, and…
hardmaru: It is wild that this framework: compression progress as the basis for beauty, curiosity, art, science, and even jokes, was laid out in full formal detail in a Japanese journal 15 years ago.
Ted Underwood: Recently learned that Schmidhuber’s Formal Theory of Fun and Creativity, covering compression progress and intrinsic motivation and their relationship to beauty, curiosity, art, science, music, and…
Open the full discussion →
2 experts are actively discussing the implications.
2 experts
2 communities
1 sources clustered
“Check out this practical developer guide for EmbeddingGemma 2 covering multimodal retrieval across text, code, images, audio, and video, modular encoders (270M to 740M), and Matryoshka truncation (768d down to 128d): https://developers.googleblog.com/embedd…”
2 experts discussed this · 7 posts
Gus: Check out this practical developer guide for EmbeddingGemma 2 covering multimodal retrieval across text, code, images, audio, and video, modular encoders (270M to 740M), and Matryoshka truncation (…
Gus: I don't think there is a plan for a re-ranker models as of now. for tutorials, we are building some more since people enjoyed the model so much
Ted Underwood: Check out this practical developer guide for EmbeddingGemma 2 covering multimodal retrieval across text, code, images, audio, and video, modular encoders (270M to 740M), and Matryoshka truncation (…
Open the full discussion →
multi-perspective LLM annotation subjective tasks
2 experts are actively discussing the implications.
2 experts
1 community
1 sources clustered
“Check out our new paper! arxiv.org/abs/2603.21404”
2 experts discussed this · 2 posts
Kristina Gligoric: Check out our new paper! arxiv.org/abs/2603.21404
Ted Underwood: Check out our new paper! arxiv.org/abs/2603.21404
angela zhou: This looks great!
Open the full discussion →
Established
AI field signal
Analysis
18d ago
AI overconfidence risk essay
2 experts are actively discussing the implications.
1 expert
1 community
1 sources clustered
“this is the source btw tedunderwood.com/2024/08/31/w...”
2 experts discussed this · 4 posts
Ted Underwood: Lots of gratitude to @carlbergstrom.com for quoting probably my favorite of my own sentences about AI at @arl.org
Ted Underwood: this is the source btw tedunderwood.com/2024/08/31/w...
Ted Underwood: Needs more dancing Milla Jovovich
Open the full discussion →
