AI High Signal

Ce qui vient de franchir un palier de reprise, le plus récent en tête.

Depuis 2 h 6 h 24 h 3 j
À partir de Tous 2 relayeurs 3 relayeurs 4 relayeurs 5 relayeurs Tout remettre
  • 14
    de 5 à 14 il y a 22 min
    Richard Sutton @RichardSSutton
    Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
    incompleteideas.net https://incompleteideas.net/IncIdeas/BitterLesson.html
    Qui l'a repris
    Stanford NLP Group chansung elie” François Fleuret” M M Saiful Bari (MARUF)” Denny Zhou”
    Ce qu'ils en ont dit
    elie @eliebakouch a cité · 08/11 09:58
    i disagree with the first point, there is nothing that says that scaling up current transformer recipe is not enough for "intelligence". you take kimi K3 or your favorite open arch + some stability fixes, you scale it to 50T parameters, you do classical training end to end, you
    François Fleuret @francoisfleuret a cité · 08/11 07:44
    We need a model that learns *in spite* of its inductive bias, not thanks to it. We have to stop being cross-entropy fetishists.
    M
    M Saiful Bari (MARUF) @sbmaruf a cité · 08/11 07:02
    True! This is the era of the RL objective. The biggest economy on this earth (US stock market) runs on subscriptions. Subscriptions require vendor lock-in. Currently, the frontier is pushing “Variance” learning from customer data, like how Google won the search market. But I
    Denny Zhou @denny_zhou a cité · 08/11 06:26
    The Bitter Lesson 2.0
  • 11
    de 5 à 11 il y a 4 h
    J
    Justin Lo @loresjpg
    Nice DeepSWE score! Does it survive a lowercase filename? Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement. Models clustered around 74% on DeepSWE end up 23 pp apart. Good
    Qui l'a repris
    J Julian Schrittwieser” Hamel Husain” Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)” Ofir Press” Lisan al Gaib Mat Velloso”
    Ce qu'ils en ont dit
    J
    Julian Schrittwieser @Mononofu a cité · 01:37
    Making novel variations of evals is very useful, public evals tend to get memorized!
    Hamel Husain @HamelHusain a cité · 01:16
    I see some are drawing the wrong inference from this study: they are focusing on this meta benchmark as yet another benchmark for model picking, i.e. open vs closed models, model families etc. That’s too narrow of an interpretation! You need to do your own private evals on
    Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex a cité · 08/11 22:00
    holy shit, an art form of scientific grift next METR/Epoch/AA employee right here you all need to work on your epistemic hygiene
    Ofir Press @OfirPress a cité · 08/11 21:26
    GPT-5.6 Sol and Luna were built by roughly the same people using roughly the same infra, I don't think we can blame the 20 point difference in Luna just on memorization. There's something else being tracked here in addition to memorization. Maybe MaxxBench questions are slightly
    Mat Velloso @matvelloso a cité · 07/11 21:20
    I'm both surprised and also not at all
  • 23
    il y a 9 h
    R
    Reflection @reflection_ai
    Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active. - Frontier reasoning efficiency - Advances the Western open frontier on coding & agentic tasks - Trained end-to-end from scratch Full weights release this month. Learn more about
    huggingface.co reflection (Reflection) Org profile for Reflection on Hugging Face, the AI community building the future.
    luma.com Open by Design: NYC Developer Meetup | Together AI × NVIDIA × Reflection · Luma Open by Design NYC: Open Models in Production What does it take to move open models from experimentation to production? Join Together AI, NVIDIA, and…
    Qui l'a repris
    Aaron Defazio” Andreas Kirsch 🇺🇦” Ben Burtenshaw” younes Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)” Together AI”
    Ce qu'ils en ont dit
    Aaron Defazio @aaron_defazio a cité · 08/11 20:11
    Reflection's Beam was trained with my AdamC algorithm! Cool... Thinking Machine's blog post also mentioned using it, so it's seeing a lot of use. It really does fix late-stage gradient norm issues.
    Andreas Kirsch 🇺🇦 @BlackHC a cité · 06/10 14:33
    Congrats to everyone there!
    Ben Burtenshaw @ben_burtenshaw a cité · 06/10 13:00
    No one (actually) working on open source is surprised by this release. For months, @reflection_ai engineers have been down in the trenches contributing to projects like transformers, vLLM, SGLang, TRL, and OpenEnv. Go check the commits. I'm super excited to see an open lab
    Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex a cité · 06/10 03:22
    interesting fact about Beam: 23B*23,8T/37B*14,8T≈0,9996! It's an iso-flop replication of DeepSeek V3, 2 years later. Not of V3's efficiency, though. in tender loving Chinese quant hands, one gimped H100 is > one GB300 in NVL72.
    Together AI @togethercompute a cité · 06/10 01:46
    Congrats to the Reflection team on Beam 🎉 Be sure to join us next week at our NYC meet-up with the Reflection team and NVIDIA
  • 5
    il y a 26 h
    J
    John Langford @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.
    hunch.net The End of AI Frontiers
    Qui l'a repris
    Terry Yue Zhuo Zachary Nado Eric Hartford rohan anil” Aaron Defazio
    Ce qu'ils en ont dit
    rohan anil @_arohan_ a cité · 06/11 20:57
    I really love the full bandwidth transformer research - One of the beautiful works of art of this year! “For example, the Bytedance experiments suggest a variant of the Full Bandwidth Transformer might be comparable to a model of twice the size, implying significant savings in
  • 5
    il y a 29 h
    Nous Research @NousResearch
    Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of
    Qui l'a repris
    Tanishq Mathew Abraham, Ph.D.” clem 🤗 Jacob Lee Teknium 🪽 Zach Mueller
    Ce qu'ils en ont dit
    Tanishq Mathew Abraham, Ph.D. @iScienceLuvr a cité · 07/11 23:52
    nous shoulda just called it a wasian finetune and sidestepped all the criticism
  • 11
    de 8 à 11 il y a 39 h
    Thomas Wolf @Thom_Wolf
    18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over. And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams. So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super
    Qui l'a repris
    Nando de Freitas Stanford NLP Group Lewis Tunstall merve M M Saiful Bari (MARUF)” Albert Gu
    Ce qu'ils en ont dit
    M
    M Saiful Bari (MARUF) @sbmaruf a cité · 06/11 04:26
    thom just took away karpathy’s mote. 🤍
  • 5
    il y a 2 j
    ARC Prize @arcprize
    Inkling 2 Small from @thinkymachines on ARC-AGI (Verified): - ARC-AGI-2: 81.7%, $0.27/task - ARC-AGI-1: 96.5%, $0.08/task It outperforms every open-weight model we've evaluated on both ARC-AGI-1 & ARC-AGI-2. Thinking Machines plans to release its full weights next week.
    Qui l'a repris
    Victoria X Lin Horace He Soumith Chintala Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)” M Martin Ziqiao Ma”
    Ce qu'ils en ont dit
    Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex a cité · 06/11 21:05
    wtf, this is excellent and actually somewhat scary. Thinky did this on a tiny compute and token budget, without any fancy architecture or RL tricks. This is the same level of achievement as 1017. Maybe better. This is FINALLY what can be expected of top-tier ex-Frontier staff.
    M
    Martin Ziqiao Ma @ziqiao_ma a cité · 06/11 20:41
    Thanks @arcprize! One of my favorite eval for monitoring RL health🫡
  • 10
    il y a 2 j
    Kimi.ai @Kimi_Moonshot
    Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at
    tools.simonwillison.net https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F9b7181492209387da1d4f53002f26663
    Qui l'a repris
    T Tinker” TDM (e/λ) (L8 vibe coder 💫)” Ofir Press” stochasm” 熊师傅 weight decay 了吗” Simon Willison”
    Ce qu'ils en ont dit
    T
    Tinker @tinkerapi a cité · 06/11 22:09
    Kimi K3.1 Pro from @Kimi_Moonshot is live on Tinker today with 256K context. K3.1 Pro is the strongest open-weights model on several benchmarks of agentic coding and knowledge work, and is the largest model on Tinker today.
    TDM (e/λ) (L8 vibe coder 💫) @cto_junior a cité · 24/10 09:09
    kimi is dropping open weights for a 3T param model and Mckinsey is still in month 7 of the copilot pilot
    Ofir Press @OfirPress a cité · 23/10 20:36
    Congrats on the ProgramBench (@jyangballin @KLieret) results!! This is our 'almost resolved' metric. The competition is really heating up.
    stochasm @stochasticchasm a cité · 23/10 20:24
    what a release, congrats to the kimi team!
    熊师傅 weight decay 了吗 @bigeagle_xd a cité · 23/10 20:02
    on 10/24/2026, a present for the last programmer's day of humans
    Simon Willison @simonw a cité · 23/10 19:58
    Pretty solid pelican
  • 7
    de 5 à 7 il y a 3 j
    Joël Niklaus @joelniklaus
    I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard. Which harness should you run your open model in? I ran 28,849 trials to find out. The leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness
    Qui l'a repris
    Zach Mueller” clem 🤗 Mario Zechner” Leandro von Werra Teknium 🪽” Ofir Press”
    Ce qu'ils en ont dit
    Zach Mueller @TheZachMueller a cité · 07/11 16:58
    Is this the largest ablation of harnesses and models ever performed?
    Mario Zechner @badlogicgames a cité · 06/11 11:40
    neat
    Teknium 🪽 @Teknium a cité · 06/11 08:20
    Hermes supremacy tweet #9138
    Ofir Press @OfirPress a cité · 05/11 19:05
    mini-swe-agent is 130 lines of code and it keeps on winning in harness comparisons.
  • 17
    de 16 à 17 il y a 3 j
    Thinking Machines @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.
    Thinking Machines Lab Introducing Inkling 2 Small Today we are introducing the second generation of Inkling models with the release of Inkling 2 Small.
    Qui l'a repris
    Zach Mueller” Chubby♨️” swyx 🔜 aie.codes” John Schulman” Rebecca Qian L Lili Yu”
    Ce qu'ils en ont dit
    Zach Mueller @TheZachMueller a cité · 07/11 02:17
    What an American open-weight frontier lab looks like. Inkling v1 -> v2 was a feat in itself. I can’t wait to see where v3 lands, and will be shoving 2 Small on the 6000’s once the weights drop
    Chubby♨️ @kimmonismus a cité · 06/11 11:32
    Inkling 2 small is an extremely impressive release, especially for its size. It is a Mixture-of-Experts model with 594B total parameters, 17B active, and supports image inputs and variable effort. In coding benchmarks, it easily keeps pace with significantly larger models. What
    swyx 🔜 aie.codes @swyx a cité · 06/11 10:50
    ok this is just ridiculous
    John Schulman @johnschulman2 a cité · 06/11 03:19
    Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process. To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and
    L
    Lili Yu @liliyu_lili a cité · 06/11 02:18
    Inkling 2 small is very strong on vision perception, and also good at planing/coding/acting.
  • 5
    il y a 3 j
    N
    Nouamane Tazi @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.
    huggingface.co The Ultimate MoE Training Handbook - a Hugging Face Space by huggingface A hands-on guide to training Mixture-of-Experts models
    Qui l'a repris
    younes Lewis Tunstall” L Loubna Ben Allal” Tony Wu Leandro von Werra”
    Ce qu'ils en ont dit
    Lewis Tunstall @_lewtun a cité · 05/11 23:34
    I haven’t seen Nouamane in person. for more than a year. Now I know why
    L
    Loubna Ben Allal @LoubnaBenAllal1 a cité · 05/11 22:32
    Nouamane cooked again 🔥
    Leandro von Werra @lvwerra a cité · 05/11 19:49
    After the Ultra Scale Playbook, Nouamane is back with this crazy guide to train MoEs! It's probably the most comprehensive resource if you want to get into large scale MoE training and an absolute gem packed with beautiful visualization!
  • 8
    il y a 3 j
    Simran Arora @simran_s_arora
    Introducing the Together Compiler: continual compilation for AI inference! I’m super proud to finally share this work from our research team! ✨ Today’s inference is “Polymorphic” or shape-shifting, it doesn’t sit still. On OpenRouter, 80% of traffic goes to models released in
    Qui l'a repris
    T Ted Zadouri T Tri Dao” Vipul Ved Prakash” D Dylan Patel” Dan Fu SemiAnalysis”
    Ce qu'ils en ont dit
    T
    Tri Dao @tri_dao a cité · 05/11 20:01
    Distilling most of our expertise in kernel and inference engine optimization into an agent loop that powers our production. A 2-year project that we're finally able to share
    Vipul Ved Prakash @vipulved a cité · 05/11 01:21
    Together Compiler discovers the most efficient inference runtimes with continuous compilation for any combination of workload, model and silicon. Excited to share the details about the research and engineering that went into it. Together Compiler (or TKCC) is a production
    D
    Dylan Patel @dylan522p a cité · 04/11 19:56
    Great work that pushes our understanding of how agents and experts can collaborate on optimizing system performance Beyond the strong results, the team also found some surprising cases where our human inductive biases end up limiting what agents can achieve
    SemiAnalysis @SemiAnalysis_ a cité · 04/11 19:35
    The Together Compiler (TKCC) from @simran_s_arora and her team is honestly one of the coolest things ever, and we have validated one of the InferenceX/AgentX results! The ability to continuously improve performance automatically while ensuring accuracy is great enough to quickly