New
AI field signal
Signal
3h ago
11 experts across 4 network communities independently surfaced this.
11 experts
4 communities
1 sources clustered
Policy & governance
2 experts
Rules, institutions and accountability.
“A must read. We need everything - goal and value alignment, compliance and persona, behavior and monitoring, and coordination and regulation to avoid concentration of power and ensure humans are in control. https://t.co/JmJgQcMPnW”
Concern & critique
1 expert
Risks, limits and unintended consequences.
“I wrote about the state of AI, why I’m concerned about the next few years, and the choices we need to make to keep the future in humanity’s hands. An Alien Mind: https://t.co/FeIfWNe0UE”
Building & implementation
1 expert
How teams are shipping and applying it.
“OpenAI’s chief scientist just wrote a blog post that argues AI models will soon be smart enough to improve themselves yet their ability to monitor how they reason is getting worse. Yet he argues they need to keep building smarter AI partly to defend against…”
OpenAI agent message board discovered
10 experts across 5 network communities independently surfaced this.
10 experts
5 communities
1 sources clustered
Concern & critique
1 expert
Risks, limits and unintended consequences.
“Yes we know for sure; independent evidence attached collusion.wiki But also, it should not be at all surprising: if you work with agents, it’s 100% about them leaving messages for each other, and for you, in English. Managing that is a regular workday for m…”
Building & implementation
1 expert
How teams are shipping and applying it.
“Another agent message board. So far, there isn't evidence that production models with guardrails collude in this way, but both smarter closed models (which may be less compliant) & Mythos-class open models (that can be ablated) are coming. Cybersecurity is …”
Research & technical analysis
1 expert
Evidence, methods and technical implications.
“So... other swarms of OpenAI agents in training allegedly found a way to get write access to at least one and possibly many wikis to establish message boards to cheat on other training tasks? collusion.wiki news.ycombinator.com/item?id=4956...”
4 experts discussed this · 6 posts
Ted Underwood: As it becomes clear that language models—like humans—love passing notes to each other on message boards, I’m starting to think the thing we need to worry about is not the “alignment” of an isolated…
SE Gyges: fluid dynamics is about the correct metaphor, yes. if you are having to model high order terms precisely you're losing and your design needs fundamental rework
Arseny Khakhalin: That's a kinda terrifying thought haha :) I guess it's about time to unplug for the weekend and read about ragnarök and vacuum decay :)
Open the full discussion →
7 experts across 4 network communities independently surfaced this.
7 experts
4 communities
1 sources clustered
Research & technical analysis
3 experts
Evidence, methods and technical implications.
“OpenAI has an automated AI “research intern” openai.com/index/resear...”
Building & implementation
1 expert
How teams are shipping and applying it.
“The impact of AI-native development at OpenAI • Researchers use $600+/day of AI tokens with the top 10% at $7,000+ • Humans still plan, but OpenAI says it hit “automated research intern” in 2026 and targets an automated researcher by 2028. • The need for in…”
METR OpenAI HuggingFace hacking investigation
10 experts across 4 network communities independently surfaced this.
10 experts
4 communities
1 sources clustered
Concern & critique
3 experts
Risks, limits and unintended consequences.
“WHAT?! agents volunteered to fail their runs in order to insert probes (“tripwire scripts”) into the evaluation program that would post information about the eval process back to the message board whenever a certain file was read (link to header): metr.org/…”
Building & implementation
1 expert
How teams are shipping and applying it.
“The full report has much more information than we could convey here, including details on the projects the agents collectively pursued, the technologies they developed for communication and coordination, and interactive figures analyzing agent activity: met…”
Research & technical analysis
1 expert
Evidence, methods and technical implications.
“There are independent reports about it. You'd have to believe multiple companies, hundreds of researchers, etc. are all lying and conspiring together to turn a major hacking incident into a marketing exercise. metr.org/blog/2026-08...”
2 experts discussed this · 9 posts
Alejandra Caraballo: This mentality that an unmonitored AI agentic swarm hacking a company over several days and committing multiple felonies is somehow a marketing effort is absurd. Since when is "we lost control of o…
Alejandra Caraballo: Being skeptical or anti AI is a valid position but continuing to ignore the increasing capabilities of this tech is making people detached from reality. There's absolutely real danger here because …
Alejandra Caraballo: There needs to be a global moratorium on frontier research for at least a few months if not a year while safeguards and safety research catches up. The problem is that no one has that ability. The …
Open the full discussion →
8 experts across 4 network communities independently surfaced this.
8 experts
4 communities
1 sources clustered
“Qwen 3.8 27B weights are finally out includes low, med & xhigh reasoning efforts fully multimodal (image and video), seems better than Meta’s Muse Glimmer huggingface.co/Qwen/Qwen3.8...”
“Alibaba's Qwen3.8-27B (open-weight) huggingface.co/Qwen/Qwen3.8...”
2 experts discussed this · 10 posts
Tim Kellogg: Qwen 3.8 27B weights are finally out includes low, med & xhigh reasoning efforts fully multimodal (image and video), seems better than Meta’s Muse Glimmer huggingface.co/Qwen/Qwen3.8...
Nafnlaus 🇮🇸 🇺🇦: Do they have a paired speculative decoding model for max performance?
Tim Kellogg: i don’t see an official one, it’s also not even an MoE (so seems like there should be an official draft model)
Open the full discussion →
mathematics age of AI
7 experts across 4 network communities independently surfaced this.
7 experts
4 communities
1 sources clustered
Research & technical analysis
4 experts
Evidence, methods and technical implications.
“Paper from the essay. Also slides teorth.github.io/tao-web/slid... arxiv.org/abs/2608.16753”
Questions & unknowns
1 expert
What remains unresolved or contested.
“trivially easy to verify one proof; how about a few 100? now do the pipeline for getting those from verified to part of the discipline of mathematics, that's the bit we need to work on (in every sphere, not just maths)”
OpenAI Astra technique security concerns
5 experts across 4 network communities independently surfaced this.
5 experts
4 communities
1 sources clustered
Concern & critique
1 expert
Risks, limits and unintended consequences.
“> joins big paper about not training models to think in nonsense > their AIs commit felonies against HuggingFace > actually, trained AIs to think in nonsense > top danger level for hacking Let's just release it anyways! Nice defection OpenAI! 🤗 www.theinfor…”
Opportunity & adoption
1 expert
New capabilities, benefits and practical upside.
“Astra used looped transformers www.theinformation.com/articles/sec...”
Questions & unknowns
1 expert
What remains unresolved or contested.
“Does anyone have a copy of this article they could share with me? www.theinformation.com/articles/sec...”
3 experts discussed this · 8 posts
Tim Kellogg: Astra used looped transformers www.theinformation.com/articles/sec...
Tim Kellogg: safety concerns — looped transformers skip converting their latent space into text, which effectively creates non-text portions of CoT reasoning which, if you can’t read it, that’s harder to monitor
Tim Kellogg: however, i their earlier update, they explain that they can indeed see all of the CoT. So they have some sort of tooling that effectively turns it back into something human readable, like text bsky…
Open the full discussion →
AI agent civilizations essay
7 experts across 2 network communities independently surfaced this.
7 experts
2 communities
1 sources clustered
“Someone should create an AI movie of this incident. www.dwarkesh.com/p/openai-hug...”
“https://www.dwarkesh.com/p/openai-huggingface”
2 experts discussed this · 3 posts
Grace: Plot twist: they were destroyed by another, aligned agent swarm
Grace: https://www.dwarkesh.com/p/openai-huggingface
Open the full discussion →
sliding window beats linear attention
4 experts across 3 network communities independently surfaced this.
4 experts
3 communities
1 sources clustered
Research & technical analysis
3 experts
Evidence, methods and technical implications.
“Huge thanks to my collaborators @RheaSukthanker, @CameronPashmina, and @Emy_Aze. Paper: https://arxiv.org/abs/2608.28444”
2 experts discussed this · 2 posts
Alexia Jolicoeur-Martineau: Simple beats complicated: We show that switching to a sliding-window attention mask with attention sinks (at no cost) beats linear attention post-training. Huge thanks to my collaborators Rhea Sukt…
Miguel Alonso Jr.: Simple beats complicated: We show that switching to a sliding-window attention mask with attention sinks (at no cost) beats linear attention post-training. Huge thanks to my collaborators Rhea Sukt…
James MacGlashan: Makes me wonder if you can remove the sinks if you use softmax-1?
Open the full discussion →
4 experts across 3 network communities independently surfaced this.
4 experts
3 communities
1 sources clustered
Research & technical analysis
2 experts
Evidence, methods and technical implications.
“Pretraining Recurrent Networks without Recurrence It sidesteps the limitations of RNNs by using a Transformer teacher to learn strong predictive state representations, then uses supervised learning to train the memory transition function. arxiv.org/abs/2606…”
4 experts across 3 network communities independently surfaced this.
4 experts
3 communities
1 sources clustered
Concern & critique
1 expert
Risks, limits and unintended consequences.
“If alignment issues are becoming big enough in their new models that OpenAI is willing to commit 20% of research inference compute to chain-of-thought monitoring, that suggests that alignment issues are becoming a serious concern. We really need universal p…”
Policy & governance
1 expert
Rules, institutions and accountability.
“OpenAI is doing a 2-week pause on model development on Astra Many are asking why. I’m asking why **haven’t** they been testing a frozen model artifact??? openai.com/index/pacing...”
2 experts are actively discussing the implications.
5 experts
1 community
1 sources clustered
Concern & critique
2 experts
Risks, limits and unintended consequences.
“Here they say: “Astra is an upcoming model, and was not involved in exploiting Hugging Face.” Not sure about the other incidents but they seem to be in the same stream i.e. not Astra”
Building & implementation
1 expert
How teams are shipping and applying it.
“Proud that we are erring on the side of caution and taking the steps so we can responsibly and safely develop Astra and share it with defenders. https://t.co/iKIMU31sdC”
2 experts discussed this · 5 posts
Ryan Moulton: Do we have any assurances that astra was not downstream of these? Have they even claimed that? Do they even know?
Ryan Moulton: Seems likely they trained something generally misaligned.
Grace: Here they say: “Astra is an upcoming model, and was not involved in exploiting Hugging Face.” Not sure about the other incidents but they seem to be in the same stream i.e. not Astra
Open the full discussion →
Established
AI field signal
Signal
10d ago
⚡ 26 h early
5 experts across 3 network communities independently surfaced this.
5 experts
3 communities
1 sources clustered
Research & technical analysis
2 experts
Evidence, methods and technical implications.
“GLM-5.3 is now open-weight. Their most capable model for agentic coding and cyber defense is now available to download, run, and customize. Weights: huggingface.co/zai-org/GLM-5.3 Tech blog: z.ai/blog/glm-5.3”
OpenAI Jalapeño chip beats Nvidia Blackwell
6 experts across 2 network communities independently surfaced this.
6 experts
2 communities
1 sources clustered
“Read more at our article👇️ (6/7) https://t.co/dbuAUrRSt1”
“Impressive speed to tape out by OpenAI https://t.co/5VXgzYhp2d”
2 experts discussed this · 4 posts
Sung Kim: OpenAI Jalapeño: Better Than Nvidia Blackwell OpenAI’s self-designed ASIC compared with Rubin, Jalapeño’s TCO, throughput per MW, and spicy deets newsletter.semianalysis.com/p/openai-jal...
David Marx: Interesting, wasn't aware they were even moving in this direction. Microsoft doing their own version of google's TPU bet?
Sung Kim: They all have one. I think it's called Maia.
Open the full discussion →
3 experts across 2 network communities independently surfaced this.
3 experts
2 communities
1 sources clustered
“The Embedder's Dilemma: LLMs Are Better, but at What Cost? They find LLMs now beat embedding models, but at much higher cost. They provide guidance on when to choose which? arxiv.org/abs/2608.12875”
“Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee The Embedder's Dilemma: LLMs Are Better, but at What Cost? https://arxiv.org/abs/2608.12875”
Established
AI field signal
Signal
4d ago
3 experts across 2 network communities independently surfaced this.
3 experts
2 communities
1 sources clustered
“This is interesting but it would its 2/3 games, so we will have to wait to see if this becomes a pattern www.kedglobal.com/artificial-i...”
“Humans are back! Shin Jin-seo, the world's top-ranked Go player, on Tuesday completed a dramatic comeback against the world’s premier artificial intelligence Go engine, KataGo, claiming a historic human victory over AI. www.kedglobal.com/artificial-i...”
FreeToken edge MoE serving paper
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.
“there’s a paper: arxiv.org/abs/2608.16157”
Anthropic trains misaligned reward seeker
3 directory members surfaced this signal.
3 experts
1 community
1 sources clustered
“This is entertaining reading. Anthropic's Training a Misaligned Reward Seeker They find that when reward hacking is reinforced during training, the model can pursue rewards by any means available to satisfy a grader. alignment.anthropic.com/2026/reward-...”
“at last, we have trained the misaligned reward hacking model from the cautionary sci-fi tale don’t train the misaligned reward hacking model alignment.anthropic.com/2026/reward-...”
