Son M. Nguyen ⌬

AI for Science - Drug Discovery and Development.
Undergraduate Student.
Graduate Teaching Assistant.
VNU - Hanoi University of Science

[藤木稟 @fujiki_rin]

『記憶を消去するとき、必ず熱が発生する』と物理学が証明しました。
物理学者ロルフ・ランダウアーが提唱した原理です。計算を行うことや情報を生み出すこと自体はエネルギーを消費しないが、1ビットの情報を消去する・忘れるときには、必ず一定量の熱エネルギーが周囲に放出されるというものです。
つまり、人間の脳もハードディスクも、何かを「忘却」するときに宇宙のエントロピーを増やし、物理的な熱を生み出している。情報に質量があるとは言いませんが、情報の消去には熱力学的な代償があるという訳です。
「忘れる」という一見ふわっとした現象が、実は宇宙に熱として支払われ、その帳簿にちゃんと記入されるのです。
これは脳科学の話であり、コンピュータの限界の話であり、宇宙論の話になります。

[唐华斑竹🦅 @uniswap12]

微软开源了一个语音 AI，60 分钟长音频一次转写，4 个人同时说话都能搞定

VibeVoice，微软开源，24.8k star，今天才知道这个。录音一键转文字这件事，我之前一直用 Whisper，但它处理长会议录音经常超时，多人说话识别错得也挺厉害的。

VibeVoice 这个直接支持 60 分钟连续音频，自带说话人分离和时间戳，四个人同时说话的场景也能分清楚谁说了什么。最让我意外的是 TTS 这边，4 个角色同时合成，90 分钟连贯输出，声音全程不跑偏。

想做有声书或者播客内容的同学应该会很感兴趣，以前多角色合成经常前后声音不一致，这个解决了。底层是 Qwen2.5 加了专门的连续语音 tokenizer，还有个 0.5B 的轻量版本，300ms 延迟，可以直接接进对话 AI 做实时语音交互，不用再单独接第三方 TTS 服务了。

正在想把 ASR 这块接进自己的会议记录工具里，如果真能稳定跑，一个会下来自动生成带发言人标注的纪要，那效率真的拉满了。

开源地址： https://t.co/XKcsCWGmcE

#AI #AIAgent

[antisense. @razoralign]

CellClick: an interactive platform for adjustable and accurate cell type annotation in single-cell and spatial omics data https://t.co/Oxd6CcdcEs

[Bodhi @BodhiCogSci]

As a PhD student myself and interacting with a lot others, my genuine takeaway is that most of us in cogsci don't have strong shared basics! As a TA I often refer to this playlist by @langofmind who did a great job at explaining cogsci with rigor & clarity https://t.co/Z5zQAac8CW

[恒星 @vintcessun]

写论文换投期刊时，最耗时的不是改内容，是改写法——不同期刊对贡献定位、结果展开、讨论范围有隐性规矩，硬套容易吃拒信。

这个项目把问题拆成两层：静态层（通用学术写作规则，可复用）和动态层（从目标期刊论文里实际抽取风格偏好）。优先级系统保证事实不可动，期刊模式优先，语料不够才回退通用规则。

生成的修订框架可审计、可人工干预，不是黑盒改稿。本质上是个解耦的“写作适配器”，扩展性好，已有静态技能也能接入。

https://t.co/1zCDoqztr6

[恒星 @vintcessun]

这篇论文终于把为什么AI学东西比人慢的原因讲透了：问题不在数据量，而在学习目标。它从样本复杂度理论出发，证明预测自身的隐表示（latent）比预测原始token在数据效率上有指数级优势——PCFG数据上，token级SSL需要Ω(exp(L))样本，latent预测仅需O(log L)。这首次从理论上解释了data2vec、JEPA等隐空间方法为何高效，也暗示了H-JEPA那种显式多尺度堆叠可能是冗余的。不过理论局限在组合结构数据，对无结构或非层次数据仍需验证。

https://t.co/RgcdRUAstg

[Masashi Misawa MD PhD @masa_ai_med]

PubMedをAIに検索させるなら、私のPubMedMCPがおすすめ。Opus 4.8なら50本以上検索させてもハルシネーションはありません。やり方は↓

[東大AI研究会 @todai_aiken]

【Google ColabだけでLLMを0から作る】
「EveryonesLLM」を公開しました🔥
Google Colabだけで、0.5Bサイズの本格的なLLM/SLMをフルスクラッチで作る教材です。

600問以上の穴埋めを解きながら、実装して、学習して、最後は会話できるモデルまで育てます。

ちょっと大変ですが、自分が作ったAIと話す朝は格別です😁
https://t.co/hEgqkto3dx

<謝辞>
本プロジェクト「EveryonesLLM」は、株式会社NTTデータMSE様のご支援のもと、学生団体「東大AI 研究会」によるOSSプロジェクトとして完成しました。 本教材の作成にあたり、温かいご支援とご協力をいただいた株式会社NTTデータMSE様に、心より感謝申し上げます。

[向阳乔木 @vista8]

最近两天爆火的写论文Skill

看来太刚需了，没几天都1.6w Star了。

安装命令：
/plugin marketplace add Imbad0202/academic-research-skills

/plugin install academic-research-skills

开源地址：https://t.co/MEfeImSCsD

[Math Cafe @Riazi_Cafe_en]

UC Berkeley's "Introduction to Probability at an Advanced Level"

PDF: https://t.co/cibGWtpeo6

[Hao Yin @HaoYin20]

Integrated Single-Cell Transcriptomic Atlas of Human #PulmonaryHypertension

10 iPAH + 20 SSc-PAH + 4 chronic thromboembolic PH + 30 healthy Ctrl

5 Smooth muscle cell states (Contractile; VCAN-/APOE-hi; FLNA-/XIST-hi; ATF3-/SOCS3-hi; Cox4I2-/TMSB4X-hi), but their heterogeneity is much smaller than 7 Fibroblast states

Modified STARsolo
Uniform raw read processing
https://t.co/SYqBfMqgnJ

ScibPlot
Benchmarking plots for scRNAseq integration methods
https://t.co/gIsyqCjV70

PH-MAP
A hierarchical cell type classification
https://t.co/RWFRFS1hBC

@atvbahajournals 2026
https://t.co/z0UaPXXUlS

[NetSci 2026 @NetSciConf]

We’re pleased to welcome 𝗖𝗮𝗺𝗯𝗿𝗶𝗱𝗴𝗲 𝗨𝗻𝗶𝘃𝗲𝗿𝘀𝗶𝘁𝘆 𝗣𝗿𝗲𝘀𝘀 as an exhibitor at NetSci 2026 📚 Since 1534, Cambridge has published influential works spanning graph theory, network analysis, dynamics, and spreading processes. 🔗 https://t.co/Edh3paWNNZ

[Jorge Bravo Abad @bravo_abad]

When does a chemistry ML model need to know the electrons, and when is geometry enough?

Machine-learning models for molecular properties come in two flavors. Some take only nuclear charges and 3D coordinates, through descriptors like SLATM, FCHL, SOAP, or geometric deep learning models such as MACE. Others (the SPAHM family, OrbNet, MOB-ML, OrbitAll) explicitly inject electronic information from a guess Hamiltonian or localized orbital analysis. For neutral closed-shell organic molecules both classes typically work well. For transition metal complexes, where the same nuclear arrangement can carry very different charge and spin states, the question is sharper: when does electronic information actually pay off?

Yuri Cho and coauthors run that benchmark cleanly. They evaluate KRR with the descriptor zoo above, plus MACE and 3DMol (a molecular variant of their 3DReact model) with optional charge and spin embeddings, on three TM complex datasets: TM-GSspin+ (3d metals, charge -5 to +4, ground-state spin), tmPHOTO (3d-4d-5d, mostly neutral, singlet-only), and Octa-MK (octahedral 3d, paired LS/HS geometries). Targets are spin-splitting energies, HOMO, LUMO, the gap, and dipole magnitudes.

The result splits cleanly. For spin-splitting energies and frontier orbital energies, whose distributions shift strongly with charge and spin, electronically informed models win. MACE-QS reaches 0.21-0.35 eV MAE for HOMO across the three datasets, and 3-SPAHM is the best KRR descriptor for HOMO.

For the HOMO-LUMO gap and dipole magnitude, whose distributions are roughly insensitive to charge state, structure-only models pull ahead. SLATM benefits from strong cancellation between correlated HOMO and LUMO errors (R^2 ~0.82 in TM-GSspin+), and AtomicDipolesMACE wins on dipoles by a wide margin by predicting the full vector then taking its magnitude.

A practical takeaway: 3DMol with charge and spin embeddings hits accuracy close to MACE-QS but trains in ~40 seconds versus 4000-15000 for MACE on the same subset, useful when iteration speed beats the last few meV.

For homogeneous catalysis, photoredox, or magnetic materials: check your property's dependence on charge and spin before picking the model. If those electronic states drive the distribution, pay for an electronically informed architecture; if not, a fast structure-only model will likely match or beat it.

Paper: Cho et al., Digital Discovery (2026) — CC BY 3.0 | https://t.co/Md9LteLPrv

[pen(Taka) @iwatobipen]

Integration QM and RDKit #cheminformatics #orca https://t.co/PyehIJJH0k

[Caglar @caglar_ee]

Video lectures, SUSTech MEE 5114 Advanced Control for Robotics ( Screw Theory, Nonlinear Control and Optimal Control ) spring 2022, by Wei Zhang
https://t.co/wyuqsFap60
https://t.co/SgKRmfsGJh
.

[Rishabh Agarwal @agarwl_]

I gave a talk at ICLR 2026 about how we are scaling RL on frontier LLMs with 1T+ parameters, on experimental data from our physical lab at Periodic!

Here's a rough recording of the talk:

[Sebastien Bubeck @SebastienBubeck]

From "Mathematical theory of deep learning: Can we do it? Should we do it?" to "There Will Be a Scientific Theory of Deep Learning".

It's respectively the title of a talk I gave four years ago, and the title of an arxiv paper from four days ago.

I really like the "learning mechanics" perspective (think of it as a continuation of "statistical mechanics", "quantum mechanics", and so on). Several of my last academic papers can be viewed under that lens (e.g. Learning threshold neurons via the “edge of stability”; or LEGO). I'm not as optimistic as the authors of the recent arxiv paper that we will EVER be able to reach what the "physics mechanics" field have achieved, but it's certainly worth trying.

Talk: https://t.co/QsBPHP0fDm
Paper: https://t.co/N4JSj3AkYZ

[QDくん⚡️AI関連の無料教材紹介 @developer_quant]

旧東工大が無料公開しているPython解説サイト
https://t.co/Tz5VtwZQN2

初心者がハマりやすい落とし穴にも言及されており、かゆいところに手が届く教材。Google Colab起動ボタンですぐ動かせるほか、Jupyter Notebook形式でダウンロード可能。

・基本文法
・データ構造
・ファイル入出力
・オブジェクト指向
・numpyの使い方
・matplotlibの使い方