Stanford Prof. @chrmanning says universities are losing AI professors to frontier labs because they still fund computer science like it’s 2005, when a laptop was enough: "Universities have just been too flat-footed in dealing with the way the computing and AI world has changed. There are really two problems." "Over the last couple of decades, universities have become enormously more bureaucratic and administrative, and that makes them a less good research environment than they used to be in the 20th century." "Universities got used to this nice period around 1990 to 2010, when computer scientists needed barely any money for their research. You bought a laptop or a PC, and it was fine. The reality now is that computer scientists, particularly in AI, need a lot of money." "It's established that if you're hiring a new faculty in physics or chemistry, you need some millions of dollars, but it hasn't trickled down that people in AI need that too. There are too many professors leaving academia." "Stanford made a start with the Marlowe cluster of 256 H100s. That's an okay baby step, but there probably needs to be about 10 times that much." @stanfordnlp @Stanford

This is exactly what I hear again and again when talking to experts. Biorisk is fake. The problem is not designing a dangerous pathogen -- AI is good at that -- it is creating it physically to make it dangerous. And that is what AI cannot do and is easy to safeguard against

RL is expensive, so every step should count. ~1 yr ago, we showed xent SFT isn’t the best way to prepare for RL (arxiv.org/abs/2510.15020). Now, we propose TailSFT (arxiv.org/abs/2608.25756), a lightweight + principled way to directly improve coverage and get better post-RL perf.

Prior to 2012 just about nobody in the core software industry (MSR doesn’t count) paid much attention to academic CS publishing. Also, prior to 2012, academics actually read papers and judged you more on the quality than the quantity of those papers. Despite the low cost of producing shit papers, even then, the system kinda worked mainly because there was little incentive to spray slop. The groundswell of interest in the outputs of CS academia broke the system. The reward for “has a NeurIPS”, divorced from the “I read this paper and want to interact with this mind” broke the system. Perhaps all that it takes for the system to rebuild is for it to burn to the ground. Once everyone with money and power thoroughly decides to ignore papers, the incentives to spam will evaporate. From the ashes of the scorched landscape of academic publishing a few fresh shoots will sprout. These will be weirdos with ideas that care about their ideas and don’t care if there’s a pot of gold at the end of the review cycle. Open science and academic inquiry will bloom again, until overgrowth and an untimely drought ignite the next wildfire.

this claim, that AI has a real risk of killing all of humanity in the near* future, has been stated regularly without evidence for over a decade. it hasn't even come close in that time. is it reasonable? two sanity check comparisons: climate change, and life w/o internet. 1/n

🤔4B params are not enough for a good coding agent... Think again! We show how to get a 4B model to be a good agentic coder without distillation! Introducing FrogNano 🐸 Check the thread for more information 👇

This is cool, but a bit confused how Table 1 can claim that our work, Parcae, doesn’t compare isoFLOPs results between looped and fixed depth? A major contribution (section 5.2) of our paper is that looping is compute optimal under isoFLOP and isoParams. arxiv.org/abs/2604.12946

It's exciting to see the rumors that Astra is a looped model. For those interested in the math on how to stably train them, I would check out our work Parcae. We also show that looping follows predictable scaling laws! arxiv.org/pdf/2604.12946

Rumor has it OpenAI's Astra is a looped model. If you want the math on how to actually train one of these without residual explosion and loss spikes, we (@hayden_prairie @zacknovack @realDanFu) worked it out in Parcae (spectral norm constraints on the injection params, per-sequence loop sampling, scaling laws). Curious whether OpenAI hit the same instabilities and fixed them the same way. arxiv.org/abs/2604.12946

In a new episode of stealing parts of production LLMs, we found a way to steal hidden model architecture and expose undocumented inference optimizations through ordinary streaming APIs. On Gemini Flash 2.5, latency jumped 3.2× at 130k tokens, revealing speculative decoding and a hidden 128K draft-model context window. sadegh-majidi.github.io/leak… 🧵

New blog post: continuous diffusion for language is back! This research direction receded into the background for a while, but as of this year, it is once again a hot topic. I wrote down a historical perspective and some thoughts on the recent revival. sander.ai/2026/08/24/continu…

Rigid, hand-designed sampling schedules become inaccurate during autoregressive video generation as errors accumulate. We introduce Equilibrium Forcing (EqF), a modular framework for video generative modeling that removes noise level conditioning to enable adaptive inference. 🌀

I'm so happy to be able to announce the first general purpose sign-language-to-text translation (SL2T) model from my team that's out today and powering a new ASL input feature on @Android phones (and don't worry, we're not stopping at ASL).

@huskydogewoof is truly one of the looped model GOATs, awesome work here! AND they used Parcae-style stable input injection too🤘

why can the china labs build glm-5.2, kimi k3, and many more to come? it is because of the openness. not just the open weights but the whole ecosystem. most of the work done in the china labs is carried by interns. i met brilliant undergrad and graduate interns who deeply understand the model training details, and they are 100x more open to share. that means the talent that knows how to train llms in china is 100x greater in number than the talent in the us, and it is growing in contrast, the us ai ecosystem is too closed. frontier labs do not hire interns. i know brilliant phd students at stanford, berkeley, and so on. they struggle to get an internship and the compute to train a properly sized model. most of the secret recipes are locked away by a very small group of privileged researchers it is not about china or the us. it is about open and closed science. the fact is that every average cs student can learn how to train an llm. they just need the opportunity. labs should be more open and hire more interns, like how deepmind and fair did in the pre-llm era

🎉 Thrilled to share we had 4 papers accepted to COLM 2026! Huge congrats to Hayden Prairie @hayden_prairie, Ivan Lee @ivn1e (lead on two!), Yasaman Jafari @yasjafarii , Zixian Wang, Cheng Yang @ChengYANG_yc , and Zachary Novack @zacknovack — 🧵 below with all the papers: 1/ "Parcae: Scaling Laws For Stable Looped Language Models" Hayden Prairie, w/ Zachary Novack, in collab with Dan Fu @realDanFu Viewing layer looping as a dynamical system yields stable looped LMs, and a new predictable scaling axis at constant memory. 📄 arxiv.org/abs/2604.12946 2/ "Studying the Soupability of Documents in State Space Models" Yasaman Jafari, w/ Zixian Wang & Leon Bergen Encode docs separately with Mamba2, then average the SSM states into one "soup": better multi-doc QA at a fraction of the inference cost, scaling to 256 docs. 📄 arxiv.org/abs/2505.24033 3/ "The Format Tax: Measuring and Mitigating the Cost of Structured Output" Ivan Lee, w/ Loris D'Antoni Forcing JSON/XML output hurts reasoning. The culprit is prompt-side distribution shift, not constrained decoding. Reason freely, then reformat: most of the accuracy comes back. 📄 arxiv.org/abs/2604.03616 4/ "Optical Context Compression Is Just (Bad) Autoencoding" Ivan Lee, w/ Cheng Yang Vision tokens aren't magic: simple baselines like mean pooling match DeepSeek-OCR-style compression on reconstruction and beat it for language modeling. 📄 arxiv.org/abs/2512.03643 #COLM2026

Lol ok this auditorium is huge haha No pressure, go humans!! 💪

#3105!! Join us!

If you're interested in unified multimodal models, come say hi and check out our poster: 📅 Thu, Jul 9, 2026 ⏰ 2:30 PM – 4:15 PM KST 📍 HALL A #4411 My collaborators @Sakiazusaaa and @Chufan_Shi will be presenting our work! 🚀

Parcae has been accepted to COLM! Feel free to reach out if you want to chat in SF!