R
Yao Fu a relayé @reflection_ai
Beam’s model technical report is now available. Read the full report to learn about the methodology our team used to train Beam end-to-end in-house, here:
AI High Signal
Ce qui est entré en base, sans jugement.
35226posts
528auteurs
3175liens
1452 hde recul
626membres
Recoupements sur la fenêtre : 2 liens
et 9 tweets relayés par plusieurs
membres · au mieux 6 personnes sur un même objet.
A
swyx 🔜 aie.codes a relayé @vtahowe
@swyx Let's gooo!!! Big congrats on 1 million! 🎉
My favorite is any debate with @ianlivingstone and @dexhorthy
Turing Post (Ksenia Se) a relayé @TheTuringPost
D
Stanford NLP Group a relayé @DSPyOSS
Nice section in this AWS post on ShipStation Global’s AI IT help desk.
A dspy.RLM module holds their entire Slack history as a variable, and mines it into a knowledge base of resolved issues.
T
Stanford NLP Group a relayé @thomaschen__
How can an LLM self-improve at theorem proving? Inspired by STP (Dong & Ma, 2025), we propose a theoretical framework for understanding prover-conjecturer self-play algorithms.. NeurIPS 2026, Sydney Poster Session 2, Tue 5-8pm @ Hall 1-4. 1/8
Ethan Perez a relayé @nytimes
Breaking News: Nvidia opened an investigation into Jensen Huang, its CEO, after he acknowledged to the company’s board that he had been intimate with a business executive who U.S. intelligence determined to be a Chinese spy.
Victoria X Lin a relayé @soumithchintala
Inkling 2 Small on OpenCode for free for a limited time, to help us improve our models further:
Leandro von Werra a relayé @ariG23498
This week we were blessed by:
1. @Nouamanetazi dropped MoE training handbook:
2. @yb2698 started training harness adapters:
3. @joelniklaus released the harness leaderboard:
Have a lovely weekend y'all. 🤗
A
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) a relayé @norpadon
I am a bit pissed that more and more papers are coming out which cite EVERYONE other than us, despite our paper coming out at the same time as DSpark and significantly earlier than DFlash2
Idk, maybe the phone-friendly format makes it look less “serious”
R
Eric Hartford a relayé @imryven
This is f*cking insane, the CEO of Y Combinator open-sourced his own second brain and most people still don’t know it exists
8 second brain repos you haven’t seen yet:
-> gbrain (by garrytan)
the second brain of the YC CEO, every fact keeps its source
G
Luke Hoban a relayé @github
Through internal testing, GitHub’s new Git storage architecture supported 5,000 concurrent agents working on a repo, sending 7,200 pushes per second, all while serving read traffic with consistently low latency.
D
Omar Khattab a relayé @DSPyOSS
Nice section in this AWS post on ShipStation Global’s AI IT help desk.
A dspy.RLM module holds their entire Slack history as a variable, and mines it into a knowledge base of resolved issues.
S
Teknium 🪽 a relayé @ScottLeimroth
I'm deaf, so I never hear the fridge alarm, an oven timer or a knock at the door. I built Signal Dog for @NousResearch's Hermes Agent Plugin Hackathon: a plugin that listens through the microphone and tells you in chat. No audio kept.
L
merve a relayé @Llama_app
Llama integrates to your favorite coding agents through their onboarding 🦙
Let your agents work offline: anywhere, anytime, keep your private info locked on your computer 🔒
Point your agent to this doc to set it up
A
J
Terry Yue Zhuo a relayé @JohnCLangford
The End of AI Frontiers .
There's a bit of a train wreck here that's painful to express. Nevertheless, it's a significant event for me and quite a few other folks involved so I wanted to memorialize a bit.
swyx 🔜 aie.codes a relayé @latentspacepod
How GitHub is rebuilding for Agents
225M developers. 34.3B commits in a year. @github’s agent era is a full-stack rebuild—and @jayparikh, EVP of CoreAI at Microsoft, is leading it.
Fresh off his GitHub Universe keynote, Jay goes deep on Copilot Fusion:
J
Zachary Nado a relayé @JohnCLangford
The End of AI Frontiers .
There's a bit of a train wreck here that's painful to express. Nevertheless, it's a significant event for me and quite a few other folks involved so I wanted to memorialize a bit.
J
Eric Hartford a relayé @JohnCLangford
The End of AI Frontiers .
There's a bit of a train wreck here that's painful to express. Nevertheless, it's a significant event for me and quite a few other folks involved so I wanted to memorialize a bit.
Graham Neubig a cité @RichardSSutton
The replies on this thread saying "what happened to the bitter lesson" 😂
Apparently many bitter-lesson-pilled people have never actually read the original post
If anything, it's an argument for RSI, not an argument for brute-force scaling.
H
Richard Ngo a relayé @natural_hazard
kinda hilarious that woodrow wilson wrote a bunch of books about how america should have a deep state
R
M Saiful Bari (MARUF) a relayé @reflection_ai
Beam’s model technical report is now available. Read the full report to learn about the methodology our team used to train Beam end-to-end in-house, here:
James Bradbury a relayé @ajeya_cotra
In a new post, I walk through some very rough modeling (thanks to @RyanGreenblatt for lots of help and feedback!):
I did not work on this project or see the unredacted report. All opinions are my own and do not represent METR or the report authors.
Andrej Karpathy a cité @karpathy
Earlier this week Tristan published his top YouTube channels:
Which looked excellent so clearly I had to turn it into TristanTube:
47K super nerdy videos +TV mode for laid back watching. (One-shotted with Claude Code in 3 minutes)
E
David a cité @steveruizok
ok I love this prompt
added it as a *very* simple find-state-machines skill to my software factory skills repo:
M
Kyle Russell a relayé @maiab
come work with me! we are HIRING a staff product designer @ImprintPayments
F
chansung a relayé @_pwd_
Codevember 2026 / 06 / Orrery
#threejs
...sound on!
#indiedev #creativecoding #shaders #webgpu #tsl #gamedev
younes a cité @yb2698
If you want to learn more about "harness adapters", we (/w @LysandreJik and @nathanhabib1011) have a blog post for you.
Blog:
(stay tuned for the final checkpoints) 🥸
T
F
younes a relayé @ariG23498
This week we were blessed by:
1. @Nouamanetazi dropped MoE training handbook:
2. @yb2698 started training harness adapters:
3. @joelniklaus released the harness leaderboard:
Have a lovely weekend y'all. 🤗
vLLM a cité @RedHat_AI
76 guest speakers, 24 organizations, 59 sessions — and every one driven by this community 📷 Thank you @RedHat_AI and all our guests. Join us bi-weekly, open to all:
Eric Hartford a cité @IshakSebsib
lazy-hq/aisdk is a big deal.
If you write agents - you should use it. No matter what language you are using.
vercel is being dumb not to own it.
C
K
Pieter Abbeel a relayé @kvfrans
Happy to release a small report on some studies at Cursor on stabilizing the combination of RL + self-distillation in production-like agentic settings. w/ @srush_nlp and @ashvinair
This work/setup is a bit stale now, but could provide some insight on
F
Theo - t3.gg a relayé @viticci
I'm working on the iPhone Duo more and more every day.
So, I've started collecting some of my favorite productivity apps for the Duo.
There are some gems in here:
CuddlySalmon a relayé @kylebrussell
In 2023, someone on Reddit posted a concept video of Magic: The Gathering as a Game Boy game, styled like Pokémon Trading Card Game. Made it real last month when Opus 5.5 dropped
Early-era rules, ~158 classic cards and boosters to crack:
Z
Tinker a relayé @ZelligeAI
Tessera-27B-B-CJK-RL: our compressed-CJK coding agent, now trained with RL.
AIME + HMMT 2026 11 point increase over non RL, Per agent turn it still uses 0.67x the base model's thinking.
Open weights:
And our followup paper:
Big
J
Chris Lattner a relayé @jaysquaredlabs
Hey @clattner_llvm, I built a web server in Mojo for a $12 server challenge. io_uring + reading SQLite’s B-tree directly through FFI. It took first place three times, peaking at 304,000 users. Currently #2, right between Rust and C.
D
Stanford NLP Group a relayé @ddemszky
I'm hiring an Applied Research Lead for my lab at Stanford! You'd work with edtech partners to build our research into tools teachers use, test it in classrooms, and lead papers on what we find.
Grad degree, 2+ years in tech, strong research writing.
L
Stanford NLP Group a relayé @LaudeInstitute
Collaborative Intelligence for the Future of Work | @Stanford
@Diyi_Yang, @tatsu_hashimoto, @chrispiech, and @erikbryn are building AI systems that complement human expertise to equip the workforce of the future.
C
Austin Lyons a relayé @chipstrat
AMD reported Q3 FY26 on November 3. I thought it was a great call, so I wanted to share some of my thoughts:
G
Pierce Boggan a relayé @github
Through internal testing, GitHub’s new Git storage architecture supported 5,000 concurrent agents working on a repo, sending 7,200 pushes per second, all while serving read traffic with consistently low latency.
J
Omar Khattab a relayé @jayden_teoh_
Evolving the architecture, in combination with our recent works on Reactive Optimizers () and Weight-Space Programming (). The picture is coming together nicely :)
C
J
Aaron Defazio a relayé @JohnCLangford
The End of AI Frontiers .
There's a bit of a train wreck here that's painful to express. Nevertheless, it's a significant event for me and quite a few other folks involved so I wanted to memorialize a bit.
Richard Socher a cité @deedydas
Currently hiring in SF + London for @Recursive_SI
-Kernel Engineer
-Platform Engineer (London)
-DevRel Engineer (SF)
+ more
B
rohan anil a relayé @BEBischof
A rat can run a restaurant, but can an agent run a bakery?
Welcome to The Great Business Bake Off: our AI Agent Hackathon.
On December 11, you and your agents get five hours to make the most profitable bakery in town.
Spots are limited:
Tim Rocktäschel a cité @deedydas
Great to see @Recursive_SI in the top 10 in talent movement. We are actively hiring:
Join a fantastic team of ex OpenAI, Anthropic, DeepMind, and Meta members of technical staff.
B
ethan ding 📊 a relayé @BEBischof
A rat can run a restaurant, but can an agent run a bakery?
Welcome to The Great Business Bake Off: our AI Agent Hackathon.
On December 11, you and your agents get five hours to make the most profitable bakery in town.
Spots are limited:
Hamel Husain a relayé @sh_reya
This hackathon is going to be epic!! Love the excellent taste that is going into it. Excited to judge. Also, the first 50 people who sign up here will get a free copy of the Evals book, so, don’t miss it!!
B
Hamel Husain a relayé @BEBischof
A rat can run a restaurant, but can an agent run a bakery?
Welcome to The Great Business Bake Off: our AI Agent Hackathon.
On December 11, you and your agents get five hours to make the most profitable bakery in town.
Spots are limited:
E
Ethan Perez a relayé @EricLevitz
The GOP had a massive money advantage. I asked
@davidshor why it bought the party so few votes.
His answer: Campaign donations are helpful. But the freedom to ignore rich weirdos' preferences is priceless.
Ryan Greenblatt a relayé @ajeya_cotra
I am sad that so much of the underlying evidence in this report was redacted. I think evidence transparency is very important for the world to figure this out in time.
Teknium 🪽 a relayé @NousResearch
LongCat-2.5, the new multimodal model from @Meituan_LongCat, is free in Nous Portal for one week.
Try it free today at
Eric Hartford a cité @QuixiAI
SPL has evolved into Charter - a DSL built on Lean4.
A superset of L. Thomas van Binsbergen's eFLINT. (All of eFLINT's tests pass on Charter)
T
N
AK a relayé @NetEaseYouDaoAI
Our Confucius4-R2T2（Real Real-Time Transcribe） paper is now on arXiv!
📄
R2T2 learns what to transcribe—and when to wait. 🎙️
Daniel Han a relayé @UnslothAI
Qwen3.8-Flash-Next GGUFs now run faster with less thinking and less token use.
Unsloth Dynamic V3.1 GGUFs bring thinking & token use to match 95% of full precision Qwen3.8-Flash.
Results: 50% fewer tokens + 15% higher accuracy vs same size quants.
GGUF:
vLLM a cité @hsliuunc
Love seeing what the community builds on vLLM-Omni 🤩 OpenDirector: a local short-film workspace — generate, edit, export — powered by vLLM-Omni with a wide range of models. Congrats @hsliuunc! 🎬
🎥 OpenDirector:
⚡ vLLM-Omni:
Tom Ellis a relayé @haskellfound
We're partnering with @MLabs10 to offer Haskell Ecosystem Packages. By procuring services from MLabs via these packages, you can contribute back to the Haskell ecosystem
Thanks MLabs!
E
Mario Zechner a relayé @earendilworks
We are looking for teammates who operate with intensity, integrity, and have a yearning to craft software and open protocols that strengthen human agency with us
N
AK a relayé @NetEaseYouDaoAI
Our Confucius4-R2T2（Real Real-Time Transcribe） paper is now on arXiv!
📄
R2T2 learns what to transcribe—and when to wait. 🎙️
P
Yann LeCun a relayé @PolymathicAI
Interested in building what’s next in AI for Science? Join us at PolymathicAI in New York (NYU + Flatiron).
We are looking for: Interns, Postdoctoral Researchers, Research Scientists and Research Engineers.
Submit your interest here:
F
Alexandr Wang a relayé @fdesouza
Today we're publishing a major RL environments update: 3,300+ tasks across tool use and computer use, already running with customers. Some tasks in our newest RL environments take a human team 3 weeks and 500+ tool calls.
N
younes a relayé @Nouamanetazi
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
Simran Arora a relayé @vipulved
Together Compiler (TKCC) generalizes beyond LLMs to ML models in PyTorch to significantly outperform the PyTorch fwd pass baseline but also a handwritten claude code harness to optimize the baseline.
Vaibhav (VB) Srivastav a cité @thsottiaux
if you’re the one securing the infra Tibo’s talking about, this is why we built Daybreak
> blue helps you find and fix bugs with GPT-6 Sol and Luna
> red helps teams test exploits with GPT-5.6 Cyber
select Red teams also get GPT-6 Cyber
apply here:
Theo - t3.gg a relayé @theo
@ajambrosino I believe you can do the work. I just think it's worth emphasizing how much work there is to be done :)
If you want a fresh start, this idiot left all his source code public:
D
Percy Liang a relayé @dlwh
BTW @WilliamBarrHeld has a great new blog post up from last week about what got us here.
Percy Liang a relayé @dlwh
It’s done! Marin’s Mochi 535B-A23B finished its main pretraining phase this morning. 18T tokens! Long context this weekend, SFT over the next week or so. Hoping to release checkpoints next week. Some EU paperwork first.
Rebecca Qian a relayé @thinkymachines
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
I
Woosuk Kwon a relayé @inferact
Great collaborating with @Google to bring day-0 Kimi K3.1 Pro support to @vllm_project on TPU v7x! ⚡🤝
Read more 👇
A
Rosanne Liu a relayé @ArnavChavan6
Will be sharing about : Towards Continual Learning: Mitigating Forgetting With Subspace Tuning at the DLCT reading group () tomorrow 6 Nov 10am PT. Check out the link to join!
Thank you
@savvyRL
@ml_collective
for the opportunity!
Paper -
Vipul Ved Prakash a cité @hasith_v
Together Compiler (TKCC) generalizes beyond LLMs to ML models in PyTorch to significantly outperform the PyTorch fwd pass baseline but also a handwritten claude code harness to optimize the baseline.
S
Victoria X Lin a relayé @SeongsikKi5837
One of the most impressive things about Inkling-2-Small is its self verification capability with vision in the loop 👀 It builds 3D scenes in Blender, renders them, inspects the output, spots what’s wrong and keeps iterating until it’s right. Huge congrats to @du_peichao75719
Y
Rowan Zellers a relayé @YueYangAI
🚀 Inkling 2 Small is out! 🚀
We pushed really hard on visual perception 👀, reasoning 🧠, and creation 🎨!
Let's try it in your own workflow:
Go team and Multimodal-first 🔥
Zach Mueller a relayé @thinkymachines
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
Pierce Boggan a relayé @moraes_c_
The new opt-in PR Overview page surfaces merge readiness, summarizes review activity, and filters unresolved threads.
A first step towards a broader PR redesign. Let us know what you think
Elicit a relayé @sethbannon
Incredible resource! A searchable and sortable database of 138 AI x pharma companies.
Victoria X Lin a cité @thinkymachines
✨Thrilled to introduce Inkling 2 Small 594B-17B(A). It is a strong multimodal agentic model with native vision capability. We invite you to try the model out through Tinker, and open-weights w/ Apache 2.0 license is coming next week.
M
swyx 🔜 aie.codes a relayé @mkanat
@swyx Haha, no way! Okay, proof of a previous statement I made: (Better if I warm up but I got excited and just wanted to make it fast.)
Denis Yarats a cité @perplexitydevs
we have been gradually improving the performance and cost of our agent api by optimizing every part of the stack. the main goal is to get the most intelligence per dollar.
find more details in this blog:
elie a cité @hyperstition_cc
very cool work, megakernels are still a relatively new field and i'm glad to see more effort in the open!!
big fan of this team and excited for what they will ship next
swyx 🔜 aie.codes a cité @swyx
[non tech, non work post]
Some of you know I have a side gig singing Broadway musicals. Proud to announce we're back next week (right after AIE CODE)!
Nov 14 3pm/7pm shows left (Nov 13 sold out)
ft. our favorite songs from:
- Tick Tick Boom
- Hamilton
-
Emmanuel Ameisen a cité @antirez
This is actually a common misconception.
Hidden states in LLMs are used to predict all future tokens not just the next one!
And in fact we see representations that correspond to plans, or the broader situation
W
Stanford NLP Group a relayé @WiAIR_podcast
New episode of Women in AI Research: Stanford PhD researcher @JulieKallini talks about her ACL 2024 Best Paper, which put Noam Chomsky's claim to the test. Can a language model learn a language no human ever could?
🎧 Youtube:
#NLProc #LLM #WiAIR
Stanford NLP Group a relayé @dlwh
It’s done! Marin’s Mochi 535B-A23B finished its main pretraining phase this morning. 18T tokens! Long context this weekend, SFT over the next week or so. Hoping to release checkpoints next week. Some EU paperwork first.
D
Pratyush Maini a relayé @datologyai
Fellowship applications close in 10 days.
Datology Labs Fellows get funding, compute, and mentorship to spend a semester on the science of data. Paid, San Mateo/NYC, starting January.
Apply by Nov 15:
N
Loubna Ben Allal a relayé @Nouamanetazi
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
E
Zachary Nado a relayé @EricLevitz
The GOP had a massive money advantage. I asked
@davidshor why it bought the party so few votes.
His answer: Campaign donations are helpful. But the freedom to ignore rich weirdos' preferences is priceless.
E
Zachary Nado a relayé @EricLevitz
Democratic pollster David Shor's take on why AI has become so unpopular
N
Tony Wu a relayé @Nouamanetazi
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
F
Vipul Ved Prakash a relayé @thefounderspack
467 seconds to boot an inference engine. Together's new compiler does it in 9.
Its agents rewrite the stack every time traffic shifts. Up to 2.6x the throughput on the same GPUs.
Same silicon. More margin.
via @togethercompute →
Woosuk Kwon a relayé @SemiAnalysis_
The Iron(wood) Throne:
Google's TPU Stakes Its Claim on Inference
Viable Alternative of CUDA Moat,
NVIDIA Working Nights & Weekends,
Jensen’s Counter Submission, Kimi K3,
Meta Muse Spark Open, GLM5.3,
TorchTPU, SGLang, vLLM, llm-d, Ray
C
Justin Johnson a relayé @ychngji6
Okay, this is what I meant by REAL real-time :)
Frameboy is finally out! If you've tried Genie, the idea will feel familiar: WASD to move, mouse to look around, with an AI model generating the world as you explore -
But one thing we've been paying a lot
H
Ross Wightman a relayé @HorizonAircraft
Canadian-made hybrid-electric aircraft could reduce northern Ontario medevac times
Q
Drew Breunig a relayé @QCon
Is another prompt instruction fixing your agent or adding debt?
@dbreunig demonstrates four context failures and practical fixes at QCon San Francisco.
Read the abstract:
Demis Hassabis a relayé @OfficialLoganK
Today we are rolling out Gemini Omni Flash 1.2, with:
- longer video generation (15 seconds)
- deeper cinematic controls for camera pov
- support for multi-turn conversation editing
- less visual drift
- 30% lower price than Omni Flash 1.1
Read more:
D
Hamel Husain a relayé @benhylak
introducing by @raindrop_ai
test your agent on every PR, against real scenarios, with one-click.
we handle all the set up: sandboxing, fixtures, real data, synthetic data, evals and more.
N
Leandro von Werra a relayé @Nouamanetazi
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
Z
swyx 🔜 aie.codes a relayé @zachlloydtweets
My chat with @swyx on software factories is live. We walk through the full factory stack:
Jimmy Lin a relayé @nour_jedidi
The official version of our paper “Don’t ‘Overthink’ Pointwise Reranking: Is Reasoning Truly Necessary?” is out! 🎉
📄
See the 🧵 below for details! 👇
Nicholas Dunzelman a relayé @OfficialLoganK
Today we are rolling out Gemini Omni Flash 1.2, with:
- longer video generation (15 seconds)
- deeper cinematic controls for camera pov
- support for multi-turn conversation editing
- less visual drift
- 30% lower price than Omni Flash 1.1
Read more:
Sarah Catanzaro a cité @HarryStebbings
Things that @bernhardsson is right on:
* The right fundraising strategy for early-stage startups
* The size of the inference market
* The opportunity to build a catalog of products
Things that Erik is wrong about:
* He *should* keep blogging. Long live
Vikram Sekar a cité @jukan05
English version of the post. 👇
Palomino is engineering microLEDs specifically for datacom, not adapting them from displays.
The emission angle of palomino microLED is narrow and helps with better coupling into fiber.
P
mei a relayé @PriMachAI
We’re still building in stealth, but sharing an early result: Sierra verified Cortex at 73.2% on τ³-benchmark.
We build algorithmic complements to LLMs. We don’t fight what models do well. We fill in the rest.
I
A
Maarten Sap (he/him) a relayé @arianna_muti
📢 New preprint out: RAGing Against Misogyny!
Misogyny is often implicit, and LLMs struggle to explain it. Can external knowledge help? We test Wikipedia, slang dictionaries, and feminist literature to find out.
📄
K
swyx 🔜 aie.codes a relayé @auchenberg
For a long time I’ve talked about Agent Experience. Many think AX is the new DX, but that misses the bigger picture.
We’re moving to an agent-first world, where agents are the primary users of all software.
AX isn’t just the new DX. It’s the new UX.
D
Teknium 🪽 a relayé @0xdippo
My old Echo Dot run Hermes now, instead of Alexa.
I built a Home Assistant integration that connects it to a Hermes profile, with its memory and tools.
Code, setup skill and the hardware recovery notes are public. @NousResearch
Q
TomLikesRobots🤖 a relayé @quixeltools
The full Quixel Megascans library is going free. Fab listings will update between now & Nov 12, so check back if an asset you want isn’t free yet.
We’ve updated terms for all Quixel assets so you can train AI models on Megascans & Megaplants.
Learn more:
W
Hamel Husain a relayé @waldekm
It's tempting to blindly trust public benchmarks but don't get fooled: they're hardly representative of your work, not to mention, by now the results are likely skewed. Do your own evals!
O
sunil pai a relayé @kultekin
*sshole is out
if you run sshd on alpine, openwrt or any musl box with password auth on: update now
official writeup:
D
AK a relayé @HuggingPapers
ByteDance just released sparse autoencoders for Qwen3-30B on Hugging Face
These SAEs reveal that mixture-of-experts have fine-grained, interpretable specializations.
Yann LeCun a relayé @KenRoth
In the United States, Brazil, Hungary, Serbia and Israel, "voters are turning out in droves to cut autocrats down to size. In country after country stricken by democratic decay, overwhelming turnout in elections is making a difference."
B
chansung a relayé @bebuu_ai
chief scientist Jie Tang just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B) during a lecture at Tsinghua University today.
Instead of relying on public datasets, their pretraining pipeline pulls from six separate
M
Mahesh Sathiamoorthy a relayé @mertcemri
What if you could synthesize an LLM inference engine customized for your model and Mac, just in time?
Meet JITEngine: custom engines built with SkyDiscover for Apple Silicon, up to 2.7× faster than the fastest alternatives we tested. 🧵 (1/n)
C
OpenAI Developers a relayé @craigsdennis
Love building with @OpenAIDevs 🤝 @TwilioDevs
Check out this quick tutorial of how to build telephone based apps that remember you using GPT Live and Twilio Agent Connect
N
Daniel Han a relayé @coldniko
New @UnslothAI Dynamic 3.1 quants are now available in the new version of Strata! 🥳
See more at:
Z
熊师傅 weight decay 了吗 a relayé @Zekunnnnn
Kimi released its own TCS benchmark, TCSExam, yesterday, and the repo only has 7 stars so far. It has 100 tasks covering Lean proofs, solvers, constructions and open conjectures.
V
Maxime Labonne a relayé @vivmarquez
Last night, we hosted Decision Models Demo Night at @LiquidAI in SF!
We kicked things off with a talk from @maximelabonne, followed by community demos showcasing what developers are building with our decision models.
Missed it? Join us next time 👉
A
Kimi.ai a relayé @AIatAMD
Day 0 support: Kimi K3.1 Pro on AMD Instinct MI355X
2.8T params, 1M context, open weights. Runs on 8x MI355X with ATOM and vLLM.
Recipes, kernels and benchmarks:
R
Teknium 🪽 a relayé @NousResearch
Plugins, connectors, and skills for Hermes Agent are now unified in a single marketplace so you can easily browse and extend your Hermes experience in a few clicks.
The new Hermes Marketplace has everything your agent needs:
V
Kimi.ai a relayé @verdacloud
Today we're introducing Verda Inference, in preview, with day-0 access to Kimi K3.1 Pro, in partnership with @Kimi_Moonshot 🧵
Frontier open-weight models through one API, engineered for every request, from a single prompt to long-horizon agents.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) a relayé @_arohan_
You can train to frontier today by cobbling together <20 MW data centers. I think there is no moat for training, deep learning finds a way.
A
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) a relayé @arXivBangers
A Circuit for In-Context Learning
In small transformers doing few-shot classification, the authors show in-context learning acts as a gated key-value memory: middle layers write task keys, while upper layers match queries to read out mapped labels.
G
Yann LeCun a relayé @GaoyueZhou
Introducing ID-WM: a hierarchical world model that decouples "where to go" from "how to get there." Finally, we can plan over hundreds to thousands of steps with WMs.
It beats the strongest baseline by 38.5%, single-level WMs by 135%. #NeurIPS2026 🧵
C
Vikram Sekar a relayé @chipstrat
When Agents Are the Interface
Agent economics, who pays for inference, vertical integration, and what it means for Google, Amazon, Microsoft, Meta, OpenAI, Anthropic, Apple and semis
Turing Post (Ksenia Se) a relayé @TheTuringPost
An agent can remember a conversation. But what about the visual context?
Oracle AI Agent Memory 26.8 gives agents image-aware memory.
It lets them keep screenshots, receipts, and product photos as persistent context across sessions →
The original image
S
Demis Hassabis a relayé @sean_n_wu
The Collaborative Intelligence Benchmark paper is out!
We tested 18 AI models as teammates on 240 tasks. The best model scores only 82.2 out of 100.
Frontier models almost never cheat. But when a teammate pushes back, they give in!
Read the paper at
TimDarcet a relayé @PHOENIX61374199
@TimDarcet I made Meta's Muse Glimmer 30B up to 21% faster on a Mac Studio: no extra memory, 252/252 runs faster.
When Glimmer tried to optimise itself, it "cheated" 10 times. My referee caught every one.
Video:
Code:
#MetaAIHackathon
AssemblyAI a relayé @iHarnoorSingh
Have you shipped speech-to-speech to production?
2027 Voice Agent Architecture Survey, first 200 get $25:
Simran Arora a relayé @togethercompute
“AI is now improving the systems that run AI.” Thanks to @Forbes for covering Together Compiler and the shift to Polymorphic Inference.
New open models can now get fully optimized for production performance in hours, not weeks.
W
Aleksander Madry a relayé @wywiadowcy_pl
Z prof. @aleks_madry rozmawiamy o wyścigu zbrojeń w świecie AI, przesypianiu rewolucji technologicznych przez Europę i wielkich zmianach, które nas wszystkich czekają. Zapraszamy!
#aleksandermądry #sztucznainteligencja #wywiad #wywiadowcypodcast #AI
Eric Hartford a relayé @__tinygrad__
Course website is live. Assignments will be posted on here, lectures won't be. Come to class!
elie a cité @RicursiveAI
very nice open weight release from ricursive with a modern dense model, benchmark scores look very strong!
beautiful scaling ladder, interesting choices like value residual (interesting contribution pattern 👀) and.. SLIDING WINDOW ATTENTION AGAIN!
B
Tim Dettmers a relayé @bohanhou1998
After our kernel-agent experiments, we had to confront an uncomfortable question: Do we still need DSLs and ML compilers when agents can do it all？
In short: We're still here! 😎
If you're curious why, here's our take:
We've also rebuilt TIRx from the
G
Lucas Beyer (bl16) a relayé @thegautamkamath
STOC 2027 received 1665 submissions. In a bold new experiment, all submissions (and author identities) are immediately made public. This is the most immediate and revealing posting of any conference I am aware of.
Take a look and see what you find!
Denis Yarats a cité @perplexitydevs
we added web grounding to our decisions api, so now you can ask questions against online data, either with a search query or a url. it's the most accurate and fastest api out there. give it a try:
Jerry Liu a cité @llama_index
i love two things in life:
- anime
- giving users the most accurate and cheap pdf parsing service at scale
we've massively simplified pricing for anyone looking to run document parsing workloads in production. come check it out!
Pan Lu a relayé @james_y_zou
AI agents enter the wet lab.
In new @Nature N&V article, I discuss recent advances in agents running automated labs and ideas on how to make these systems more reliable.
H
Omar Khattab a relayé @HaozeHe_Hector
We edited 0.0002% of NanoJev’s weights, and the edited model beat the original 65–34 (1 draw) in 100 head-to-head Tank War games. 🎯
The original took only about 1 in 3 clean shots. Using Weight-Space Programming by @rikiyatakehi et al. (), we read the
M
Jon Durbin a relayé @MacrocosmosAI
Introducing IOTA Decide 4B, our first public model.
Pareto optimal. Built different.
Trained across globally distributed GPUs with our SDK.
Open weights, Apache-2.0.
🤗
Simran Arora a relayé @realDanFu
We view this as the first step towards building for an increasingly polymorphic future. @simran_s_arora and I wrote a little blog on our view of polymorphic inference and the implications for our research agenda.
Check it out!
5/
S
Pan Lu a relayé @sean_n_wu
The Collaborative Intelligence Benchmark paper is out!
We tested 18 AI models as teammates on 240 tasks. The best model scores only 82.2 out of 100.
Frontier models almost never cheat. But when a teammate pushes back, they give in!
Read the paper at
D
V
Eric Hartford a relayé @volkdude85
I have been working on this alot It's a Modified Quixidesk, I added hermes tools and daemon manipulation as well as some CSS work. It also has a flatpak now. @QuixiAI,
Z
Tadas Antanavicius a relayé @zachlloydtweets
I often see tweets that a software factory can be built with a string of prompts tying together cloud agents. My experience is that this is false.
Companies often start there and regret it because it doesn’t scale.
Azalia Mirhoseini a cité @RicursiveAI
Very excited to launch Ripple-65B Preview, a new state of the art for U.S. open-weight model on agentic coding!
You can run Ripple-65B locally and efficiently on a MacBook Pro or DGX Spark!
Give it a try:
HuggingFace:
Together AI:
A
Google Gemini a relayé @Android
Henry Cavill knows a thing or two about mustache injustice, and with a little help from Google, is making a video game on Googlebook this Movember. Watch the full episode at
Azalia Mirhoseini a cité @RicursiveAI
Very excited to launch Ripple-65B Preview, a new state of the art for U.S. open-weight models on agentic coding!
You can run Ripple-65B locally and efficiently on a MacBook Pro or DGX Spark!
Give it a try:
HuggingFace:
Together AI:
Jonathan Frankle a relayé @neondatabase
Most agents work for one person, in one chat.
@kylon_ai puts your whole team, agents included, in shared rooms
When those agents build apps, each app gets its own Neon backend, and every draft runs on its own branch before it ships:
M
Raiza Martin a relayé @mgorbach
Big news for our developers! We are shipping a developer preview that dramatically expands the capabilities of all apps to integrate into Siri. So excited to see what you folks do with this.
I
rohan anil a cité @Jessicalessin
You can train to frontier today by cobbling together <20 MW data centers. I think there is no moat for training, deep learning finds a way.
C
clem 🤗 a relayé @Kimi_Moonshot
Releasing the model weights of Kimi K3.1 Pro:
We'll keep advancing open intelligence.
B
Ben (no treats) a relayé @bradrcarson
So the 41 datacenter moratoria across the country are, I predict, going to go 41-0. In nearly every case it's about 67-80% opposing datacenters. True in California, Ohio, everywhere. Read all the results here:
A
Drew Breunig a relayé @AgenticAIFdn
Every morning, an agent checked Strava, a calendar and ferry times to tell a cyclist when to leave home. Once the routine was clear, it didn’t need a large model.
Drew Breunig on turning repeat agent tasks into workflows.
C
A
Ostris a cité @PrunaAI
AI Toolkit now supports fine tuning P-H3-Image directly with a training adapter.
Lewis Tunstall a relayé @victormustar
Axion-1.0 always cool to see a team release a full open training recipe:
model:
dataset:
code:
paper:
S
younes a relayé @evalstate
New release of atif-scan (0.8.16) available -- check benchmark traces for reward hacking, time skewing, awareness and more!
This has an improved MCP Server for trace analysis, some improved checks and handling for Kimi Code and tb4 tasks/allowances.
Jenia Jitsev 🏳️🌈 🇺🇦 🇮🇱 🇮🇷 a cité @micahgoldblum
Very nice work connecting the instabilities in training and previous workarounds to mitigate it to one common cause - loss of precision due to finite precision errors in model head when working with lower precision (BF16 and below). Fits to this run in FP8
Eric Hartford a relayé @TencentHunyuan
👋Hy4 is here, new with vision.
Firmly among the leading open models.
More productive everywhere: finance, frontend, game, officework, research…
Built responsibly, with cyber/biosecurity in mind.
Try it out and let us know!
🔗：
swyx 🔜 aie.codes a relayé @latentspacepod
⚡ Behind Latent Space: from AI Substack to the first @aiDotEngineer in China 🇨🇳
This time @Ninggao88 of Linkloud turns the mic around: asking @swyx how LS started, why “AI Engineering" needed a name, and what’s different about the builder community in
N
N
D
Nando de Freitas a relayé @diffusion_llms
Georgia Tech Dynamical ML Group is recruiting PhD students and postdoc in diffusion language models for Fall 2027
Please join our discord server to learn more.
J
Synthesia 🎥 a relayé @joeciarallo
Today at @OpenAI Dev Day in London, we launched the @synthesiaIO plugin in ChatGPT. Describe the video you want, or share a doc, deck and spreadsheet, and our platform builds a finished video right in the chat, then refines it with you.
S
chansung a relayé @sidahuj
Step-by-step breakdown of how you can create an photorealistic scenes using MCP for Blender:
Mario Zechner a relayé @migueldeicaza
Xogot on iOS now whips with a Pi.dev-powered agent. And it uses Live Activities to work while you doom scroll on the go.
Since it is a sensitive subject for some developers, I wrote a little bit about it, with no AI:
B
Ethan Perez a relayé @bradrcarson
So the 41 datacenter moratoria across the country are, I predict, going to go 41-0. In nearly every case it's about 67-80% opposing datacenters. True in California, Ohio, everywhere. Read all the results here:
Y
Tinker a relayé @EchoShao8899
@GoogleColab In “HW2: Deep Personalization with Tinker”, students get the chance to posttrain production-scale LLM thanks to @tinkerapi!
B
James Bradbury a relayé @bradrcarson
So the 41 datacenter moratoria across the country are, I predict, going to go 41-0. In nearly every case it's about 67-80% opposing datacenters. True in California, Ohio, everywhere. Read all the results here:
J
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) a relayé @jessi_cata
For context, Sam is responding to "Machines of Loving Grace". Here, "Whuffie points" refers to a science fictional social credit system. Dario is vaguely suggesting AI-run social credit based on a human value representation, among other ideas.
H
Peter Welinder a relayé @haydenfield
In many circles, “Ask Chat” has replaced “Google it,” and that’s reflected in a new announcement from OpenAI: the volume of search queries run through ChatGPT now equals 20% of the volume run through Google.
W
Dan Fu a relayé @opengroundsFX
Proud to announce that ParallelKernelBench () has been accepted to NeurIPS 2026! 🎉
If you are interested in the future of multi-GPU kernels, come say hi!
Huge thanks to my amazing collaborators @asplencmnt @simonguozirui @simran_s_arora @realDanFu
W
Simran Arora a relayé @opengroundsFX
Proud to announce that ParallelKernelBench () has been accepted to NeurIPS 2026! 🎉
If you are interested in the future of multi-GPU kernels, come say hi!
Huge thanks to my amazing collaborators @asplencmnt @simonguozirui @simran_s_arora @realDanFu
K
Stas Bekman a relayé @Karthikg1996
Frontier labs train models for their own agent harnesses. Every product company should be able to do the same.
We’re making it practical with @OpenHandsDev, @harborframework ,@NovaSkyAI + @Snowflake Cortex Training.
N
Yann LeCun a relayé @Numerama
Dans une étude mise en ligne le 29 octobre, Yann LeCun et Randall Balestriero affirment démontrer qu'un petit modèle d'IA bien nourri se fait une meilleure image du monde qu'un LLM 660 fois plus gros.
Karina a relayé @thoughtfullab
The first polls are closing now, and agents are submitting their final predictions!
Track how the final predictions are doing live as races are called by @AP, @DecisionDeskHQ, and others as we grade the agents’ performance in live time at!
Agents are
A
James Bradbury a relayé @adamkovac
How important was AI and data centers to your vote?
Crucial: 23%
Major factor / minor factor / non-factor: 77%
S
Ethan Perez a relayé @ShakeelHashim
New research from @Tech_Oversight : quite striking stats on how much Dems are using tech accountability messaging in campaign ads.
Jerry Liu a cité @llama_index
There is so much alpha in this blog post for anyone learning model optimization.
Check it out!
Karina a relayé @brian1zhou
Take a look and see what the agents got right in live time at! We'll be on @MTSlive to talk about some early takeaways from long-running AI forecasters and traders at 9:30 ET :)
Jerry Liu a cité @llama_index
OpenDocRouter is an awesome playground for trying out different document OCR models: from open-weight to frontier.
Check it out!
Ethan Perez a relayé @washingtonpost
Exclusive: Workers inside Google’s AI division are pushing executives to waive their non-disparagement agreements, arguing they must be free to inform the public about the risks of the technology they are working to advance, according to sources.
vLLM a cité @RedHat_AI
Hola Madrid! 🇪🇸 vLLM is coming to town — join us Nov 12 with @RedHat_AI and @NVIDIAAI for an evening of open-source inference: vLLM, vLLM-omni, LLM Compressor, and llm-d on Kubernetes. Register:
MMitchell a relayé @NielsRogge
We @huggingface reproduced this paper!!
It's built with ML Intern, an open-source harness to do AI research, train models, and more
Model:
Demo:
Russ Salakhutdinov a cité @JasonJZLiu
Excited to share Pixel and Pixel-UMI from our team at CMU by amazing @JasonJZLiu and @_tonytao_ .
1. Pixel: Full-Stack Robot System for Learning General-Purpose Mobile Manipulation:
2. Pixel-UMI: Learning Environment-Aware
Jerry Liu a relayé @llama_index
OpenDocRouter provides a simple API and SDKs for integrating into your apps. Read our docs (or copy-paste them to your agent) to integrate into any existing codebase in minutes.
T
Stanford NLP Group a relayé @Shadowfetch
Spotlight: Percy Liang (@percyliang), Stanford professor and director of Stanford's Center for Research on Foundation Models. He's first author of the HELM paper, created Marin for open model development, and co-teaches CS336. 🌟
#AI #LLM #OpenSource