zhuai

[Mistral AI for Developers @MistralDevs]

🔥 Meet Mistral Small 4: One model to do it all.
⚡ 128 experts, 119B total parameters, 256k context window
⚡ Configurable Reasoning
⚡ Apache 2.0
⚡ 40% faster, 3x more throughput

Our first model to unify the capabilities of our flagship models into a single, versatile model.

[Hodachi | LLM & AI Engineer @hokazuya]

キタコレ
Qwen3.5-397B-A17Bの日本語OCR精度、これはやばいといわせてほしい

CER（文字誤り率）もひっくるめて、ほとんどの項目でGPT-5.2、Claude-Opus 4.6より高性能という結果に👀

OCRで正確な日本語抽出が難しいとされている、文部科学省のSSH支援事業の一枚を渡して同じ指示で出力させた結果、OCRの一般的な測定項目。
これがApache2.0で商用利用可能なモデルってのが激熱すぎる

もちろん、これだけで判定はできないけど、このレベルの画像に対する精度結果でこれだけの優位性はやばいでしょ

オンプレでB200購入して完全閉鎖空間でしかやれないPSの企業とちょうどOCRもローカルでという会話していたからありがたすぎる

テスト画像の出展：
文部科学省 - 令和6年度スーパーサイエンスハイスクール生徒研究発表会表彰校を決定しました - 参考1：スーパーサイエンスハイスクール概要（PDF:589KB）

[シェイン・グウ @shanegJP]

モデルもプロダクトも作らずに論文を書いていたわけでなく、本を書いていた様です。私も暇な時間に本を書いて4000億円の市場価値を狙いにいきます。

[シェイン・グウ @shanegJP]

この２年間大したモデルも作れていないし、プロダクトもない。論文のほとんどはインターン任せで、創業メンバーの論文も「AICudaEngineer」など炎上。シリコンバレー基準なら、この実績でシリーズB評価をここまで積むこと自体かなり不自然で、日本の盲目的な受託頼みで延命している印象が強い。そして今回のシリーズB、実は一部創業側のセカンダリー（キャッシュアウト）が入っていてもおかしくないと思う。

[Susan Zhang @suchenzang]

anyone outside of modal hosting this 9B mamba ssm right now?

separately... has anyone looked at the tokenizer for this model? something looks... off...

would also like to find this source, if it exists.

[シェイン・グウ @shanegJP]

日本AI市場は歪んでいます。論文だけでユニコーンになれたのは日本だけです。堅実にプロダクトやモデルをちゃんと作っている会社達を世界基準で評価し、一歩一歩競争を作らないと歪みは加速します。日本は表の批判が無いので。日本には根本的に強くなって欲しいです。https://t.co/xj91yxuceV

[CLS @ChengleiSi]

Are AI scientists already better than human researchers?

We recruited 43 PhD students to spend 3 months executing research ideas proposed by an LLM agent vs human experts.

Main finding: LLM ideas result in worse projects than human ideas.

[Alexandr Wang @alexandr_wang]

very excited to be working with @shengjia_zhao, @jhyuxm, @ren_hongyu, and @shuchaobi 😊

towards superintelligence 🚀

[Lucas Beyer (bl16) @giffmana]

hey all, couple quick notes:
1) yes, we will be joining Meta.
2) no, we did not get 100M sign-on, that's fake news.

Excited about what's ahead though, will share more in due time!

cc @__kolesnikov__ and @XiaohuaZhai.

[Binyuan Hui @huybery]

We’ve released the Qwen3 technical report, with model details and comprehensive evaluations. Enjoy! 👇🏻
https://t.co/hoVoUjaNYb

[Wenhu Chen @WenhuChen]

https://t.co/CnF3eyDXRs is quite interesting. It standardizes the evaluation of all the existing math reasoning models and re-evaluate these models.

Takeaway 1: Most RL-trained variants of the DeepSeek R1-Distill model do not yield meaningful performance improvements (except DeepscaleR), suggesting that a reliable and scalable RL training recipes are still lacking.

Takeaway 2 While RL-trained methods can often substantially improve base model performance, instruction tuning remains superior (except Open Reasoner Zero), suggesting again that a reliable and scalable RL training recipes are still lacking.

They propose to maintain a third-party evaluation of math reasoning models at https://t.co/ZlJc35RXsI. This effort is really applaudable.

[ちょっと昔の日本の景色bot @jaapn80to00pn]

1994年、奈良市・パラディ学園前

[Jürgen Schmidhuber @SchmidhuberAI]

The #NobelPrizeinPhysics2024 for Hopfield & Hinton rewards plagiarism and incorrect attribution in computer science. It's mostly about Amari's "Hopfield network" and the "Boltzmann Machine."

1. The Lenz-Ising recurrent architecture with neuron-like elements was published in 1925 [L20][I24][I25]. In 1972, Shun-Ichi Amari made it adaptive such that it could learn to associate input patterns with output patterns by changing its connection weights [AMH1]. However, Amari is only briefly cited in the "Scientific Background to the Nobel Prize in Physics 2024." Unfortunately, Amari's net was later called the "Hopfield network." Hopfield republished it 10 years later [AMH2], without citing Amari, not even in later papers.

2. The related Boltzmann Machine paper by Ackley, Hinton, and Sejnowski (1985) [BM] was about learning internal representations in hidden units of neural networks (NNs) [S20]. It didn't cite the first working algorithm for deep learning of internal representations by Ivakhnenko & Lapa (Ukraine, 1965)[DEEP1-2][HIN]. It didn't cite Amari's separate work (1967-68)[GD1-2] on learning internal representations in deep NNs end-to-end through stochastic gradient descent (SGD). Not even the later surveys by the authors [S20][DL3][DLP] nor the "Scientific Background to the Nobel Prize in Physics 2024" mention these origins of deep learning. ([BM] also did not cite relevant prior work by Sherrington & Kirkpatrick [SK75] & Glauber [G63].)

3. The Nobel Committee also lauds Hinton et al.'s 2006 method for layer-wise pretraining of deep NNs (2006) [UN4]. However, this work neither cited the original layer-wise training of deep NNs by Ivakhnenko & Lapa (1965)[DEEP1-2] nor the original work on unsupervised pretraining of deep NNs (1991) [UN0-1][DLP].

4. The "Popular information" says: “At the end of the 1960s, some discouraging theoretical results caused many researchers to suspect that these neural networks would never be of any real use." However, deep learning research was obviously alive and kicking in the 1960s-70s, especially outside of the Anglosphere [DEEP1-2][GD1-3][CNN1][DL1-2][DLP][DLH].

5. Many additional cases of plagiarism and incorrect attribution can be found in the following reference [DLP], which also contains the other references above. One can start with Sec. 3:

[DLP] J. Schmidhuber (2023). How 3 Turing awardees republished key methods and ideas whose creators they failed to credit. Technical Report IDSIA-23-23, Swiss AI Lab IDSIA, 14 Dec 2023. https://t.co/Nz0fjc6kyx

See also the following reference [DLH] for a history of the field:

[DLH] J. Schmidhuber (2022). Annotated History of Modern AI and Deep Learning. Technical Report IDSIA-22-22, IDSIA, Lugano, Switzerland, 2022. Preprint arXiv:2212.11279. https://t.co/Ys0dw5hkF4 (This extends the 2015 award-winning survey https://t.co/7goTtI5Uwv)

[zhuai @guo0914]

@prajdabre1 @_arohan_ UL2 is in ICLR2023

[Noam Shazeer @NoamShazeer]

Character AI is serving 20,000 QPS. Here are the technologies we use to serve hyper-efficiently. [https://t.co/R14Jt9Z5yo ]

[zhuai @guo0914]

@prajdabre1 People prefer Neurips? And COLM attracted some submissions?

[zhuai @guo0914]

@ShiweiLiu9 @_EldarKurtic @neuralmagic Thank you very much!

[zhuai @guo0914]

@prajdabre1 @AlhamFikri For several years, the ICLR review form had a score on this. The RoBERTa was rejected due to the lack of technical novelty.

[zhuai @guo0914]

@ShiweiLiu9 @_EldarKurtic @neuralmagic It will be very helpful for the pruning community to open your code for evaluating the speedup. Reviewers who are not familiar with this field often have such concerns.

[zhuai @guo0914]

@ShiweiLiu9 @_EldarKurtic @neuralmagic In nm-vllm https://t.co/z6PxLaVKFN , it achieves around 1.5x inference speed-up in A10 GPU using 50% unstructured sparsity (not 2:4). Have you checked nm-vllm?