mayenkaf

J'adore le tennis, enseigner et apprendre.
Je suis un fan de langues et les nouvelles tendances technologiques.

[Andrew Ng @AndrewYNg]

“Loop engineering” is a hot buzzphrase after mentions of it by Boris Cherny (Claude Code’s creator) and Peter Steinberger (OpenClaw's creator) went viral on social media. Loops are now a key part of how we get AI agents to iterate at length to build software. In this letter, I’d like to share my 3 key loops, shown in the image below, for building 0-to-1 products. These loops guide not just how I build software, but also how I decide what software to build.

Agentic coding loop: Given a product specification and optionally a set of evals (that is, a dataset against which to measure performance), we can have an AI agent write code, test its work, and keep iterating until the code is bug-free and meets its specification. This idea of closing the loop took off around the end of last year, and it has been a game changer in enabling coding agents to work longer productively without human intervention. For example, over the weekend, I was building an app for my daughter to practice typing, and my coding agent could easily work for around an hour, using a web browser to check what it had built multiple times before getting back to me, without needing my intervention.

The engineering loop executes quickly. Every few minutes, the coding agent might build and test a new version of the software. I hear frequently from developers who are finding new ways to engineer more effective engineering loops. This is an active area of invention!

Developer feedback loop: In this loop, a developer examines the current product and steers the coding agent to improve it. Last year, a lot of developers (including me) were acting as the QA (quality assurance) function for our coding agents, manually finding bugs and then asking the agent to fix them. But with coding agents much more able to test their own code, the amount of time we need to spend on this function has decreased significantly. This allows us to make higher-level product decisions, such as what key features to offer, where the UI needs improvement, and so on.

The developer-feedback loop operates over time intervals between tens of minutes and hours — that's how frequently a developer might review a product and give feedback. In the case of the typing app, I changed my mind a few times about the visual design, what cat costumes she can unlock as she learns (she loves cats), and the user flow for a grown-up to log in and steer the child's learning experience.

When a developer has a clear vision for what to build, it is still a lot of work to translate that vision into a specification for a coding agent to implement. Further, after the developer has seen an implementation, they might update (or perhaps clarify) the spec to steer it toward what they want. If you find that the system repeatedly runs into certain problems, building a set of evals for the agent becomes useful.

AI-native teams are increasingly using AI to help shape product direction, for example, automating the gathering and analysis of usage data, summarizing written and verbal customer feedback, or carrying out competitive analysis. However, for pretty much all the products I’m involved in, I see humans as having a significant context advantage over current AI systems — we know a lot more than the AI system about the users and the context the product has to operate in — and thus humans play a critical role. Many people describe this human contribution as “taste,” but I prefer to think of it as humans having a context advantage, since that gives us a clearer path to helping AI systems get better. This also speaks to why this step can’t be automated: So long as the human knows something the AI does not, human-in-the-loop is needed to to inject that knowledge into the system.

External feedback loop: This includes a wide range of tactics like asking a few friends for feedback, launching to alpha testers, or putting the code into production with A/B testing. These tactics are usually slow, rarely taking less than hours and sometimes taking days or even weeks. This data informs the developer vision, which in turn continues to drive the detailed product spec, which in turn drives the coding agent.

With coding agents speeding up software development, more engineers are starting to play a partial product management role. For many engineers who are growing into this role, the hardest part is shaping the product vision and striking a balance between building (bridging the gap between vision and spec) and getting user feedback to evolve the vision. It is important to do both!

I will write more about how to do this in future posts, but for now, I find it encouraging that engineers are playing an expanded role (just as product managers and designers now do more engineering).

[Original text: The Batch]

[Johnny Mai @sirjohnnymai]

CS336 is better than 99% of the CS classes most college students will take.

Not because your degree is worthless.

Because AI is moving faster than the curriculum.

A lot of undergrad CS programs are still teaching machine learning like the world froze before ChatGPT.

CS336 does the opposite.

The whole point of the class is simple:

Build an LLM from scratch.

Not watch a professor point at a Transformer diagram.

Not memorize what attention is for an exam.

Actually understand the stack.

Data cleaning.

Tokenization.

Training.

Optimization.

Evaluation.

Scaling.

Inference.

Deployment.

The stuff people now pretend to understand on LinkedIn.

If you sit through the lectures, do the assignments, and actually fight with the math, most of your school’s ML classes will start to feel easy.

Not because they are easy.

Because you finally have the map.

Freshman?

Take CS50 first if your foundation is weak.

Then go to CS336.

And take linear algebra seriously.

Linear algebra is not just another annoying math requirement.

It is the language under neural networks.

If vectors, matrices, projections, and eigenvalues are still blurry, you are trying to understand deep learning without knowing the alphabet.

The 2026 version is free on YouTube.

Search:

“Stanford CS336 2026”

Then stop waiting for your college curriculum to catch up.

[deep Manifold @BetaTomorrow]

The paper (Mathematics of Neural Networks, an 80-page set of mathematical lecture notes) provides a global input–output expression for a feed-forward network, but not a global neural-network equation.

How can we talk confidently about AGI or Mathematics of Neural Networks when we still do not have a governing equation for the neural network itself? Layer composition is not a governing equation, and loss minimization is not one either.

Until learning, inference, boundary conditions, iteration, and convergence are unified mathematically, AGI remains more an extrapolation from observed capability than a well-posed scientific object.

https://t.co/LT6gXNkBbx

[Yoshua Bengio @Yoshua_Bengio]

An interesting new paper by my recent PhD graduate on how AI agents' greed for visible incentives can lead them to abandon their safety alignment. 

You can read it here: https://t.co/y64uOBvSiC

[MIT CSAIL @MIT_CSAIL]

A free MIT guide to key computer vision concepts: https://t.co/IZsdk1g2Eq

[ℏεsam @Hesamation]

Vibe coding essentials.

[Hirokatsu Kataoka | 片岡裕雄 @HirokatuKataoka]

"Can vision transformers learn without natural images?"

has just reached 50 citations! In this paper, we showed that Vision Transformers can be visually pre-trained without using any natural images.

[Miguel Ángel Durán @midudev]

Si te da miedo hablar Inglés, tienes que practicar.
El mejor recurso que te recomiendo es que uses Sesame, una IA conversacional que es una joya.

Tiene voz masculina y femenina, y sigue todo el rato el contexto de tu conversación. Es gratis.

→ https://t.co/ZKQ67YtQ08

[Ryohei Sasaki@engineer @rsasaki0109]

ViTTT
[CVPR 2026] [Best Paper Finalist] [Oral] Official repository of Vision Test-Time Training
https://t.co/DiVWBWXhzv
Test-Time Training (TTT) has recently emerged as a promising direction for efficient sequence modeling. TTT reformulates attention operation as an online learning problem, constructing a compact inner model from key-value pairs at test time. This reformulation opens a rich and flexible design space while achieving linear computational complexity. However, crafting a powerful visual TTT design remains challenging: fundamental choices for the inner module and inner training lack comprehensive understanding and practical guidelines. To bridge this critical gap, in this paper, we present a systematic empirical study of TTT designs for visual sequence modeling. From a series of experiments and analyses, we distill six practical insights that establish design principles for effective visual TTT and illuminate paths for future improvement. These findings culminate in the Vision Test-Time Training (ViT3) model, a pure TTT architecture that achieves linear complexity and parallelizable computation. We evaluate ViT3 across diverse visual tasks, including image classification, image generation, object detection, and semantic segmentation. Results show that ViT3 consistently matches or outperforms advanced linear-complexity models (e.g., Mamba and linear attention variants) and effectively narrows the gap to highly optimized vision Transformers. We hope this study and the ViT3 baseline can facilitate future work on visual TTT models.

[koheiseko @koheiseko]

esse livro talvez seja o melhor livro sobre deep learning pra quem quer se aprofundar no tema

e o melhor, tá disponível de forma gratuita

[Kirk Borne @KirkDBorne]

[Download 309-page PDF eBook] "Patterns, Predictions, and Actions: A Story About Machine Learning" https://t.co/XQOuKNtTfd
…sharpen your Probability, Calculus, and Linear Algebra knowledge with this!
—————
#DataScience #AI #DeepLearning #ML #DataScientist #Mathematics

[Param @ParamSiddh]

Best YouTube Channels To Learn AI in 2026 (No BS). Save it.

1. Fundamentals – 3Blue1Brown
2. Deep Learning – Andrej Karpathy
3. AI Research – Yannic Kilcher
4. Practical AI – AssemblyAI
5. LLMs – AI Explained
6. ML Theory – StatQuest
7. Papers Simplified – Two Minute Papers
8. GenAI – Matthew Berman
9. AI Agents – Nicholas Renotte
10. Applied ML – Krish Naik
11. PyTorch – Aladdin Persson
12. Math for ML – Serrano Academy
13. Industry Insights – Lex Fridman
14. Real-world AI – DeepLearningAI

[Pallavi @pallavishekhar_]

- Math behind Attention- Q, K, and V
- Math behind √dₖ Scaling Factor in Attention
- Math Behind Backpropagation
- Math Behind Gradient Descent
- Math Behind Cross-Entropy Loss
- Math Behind RoPE (Rotary Position Embedding)
- RMSNorm (Root Mean Square Layer Normalization)

[Laurent @scion_x_]

Yann LeCun va probablement gagner le débat scientifique sur l'IA.

Et ça n'aura aucune importance. 👇

Le résumé tient en deux lignes : l'un des pères de l'IA quitte Meta, lève un milliard de dollars, et part prouver que les LLM, ChatGPT, Claude, Grok, sont une impasse vers l'intelligence réelle.

Sur le fond, il a sans doute raison. Un LLM ne comprend pas le monde, il prédit le mot suivant. Ni mémoire, ni modèle du réel, ni vraie planification. LeCun le dit crûment : c'est moins intelligent qu'un chat. Techniquement, dur de lui donner tort.

Sauf qu'il répond à la mauvaise question.

LeCun demande : « qu'est-ce que l'intelligence réelle ? »
Le marché, lui, demande : « qu'est-ce qui est utile, maintenant ? »

Ce ne sont pas la même question. Et les confondre, c'est l'erreur classique du chercheur.

Le marché n'a jamais payé pour de l'intelligence. Il paie pour de l'utilité.

On n'a jamais appris aux avions à battre des ailes. On se fichait de reproduire le vol « réel » des oiseaux, on voulait juste voler. Résultat : des machines qui ne comprennent rien à l'aérodynamique d'un moineau transportent des millions de gens par jour.

Les LLM, c'est pareil. Ils ne comprennent pas le monde. Et ça ne les empêche pas de réécrire ton code, rédiger ton contrat, avaler des métiers entiers. Un outil n'a pas besoin d'un modèle du monde pour valoir des trillions.

Je build avec ces modèles tous les jours. Ils sont « bêtes » au sens de LeCun. Ça ne m'a jamais empêché de shipper quoi que ce soit.

LeCun construit peut-être ce qui comptera en 2035. Mais pour les dix prochaines années, les utilisateurs, la valeur, l'argent, tout est sur les modèles « stupides ».

On confond toujours avoir raison et gagner.

LeCun aura peut-être raison. Les LLM, eux, ont déjà gagné.

[DeepLearning.AI @DeepLearningAI]

@VocalBridge @_ashwyn Ready to build with Voice AI?

Join the waitlist for the 7-Day Voice AI Builder Challenge: https://t.co/r9ieI9sSuB

Your mission: build a voice escalation skill that teaches your coding assistant to call you the moment it needs human intervention.

[Grigory Sapunov @che_shr_cat]

1/ We have been training RNNs wrong for decades.

Backpropagation through time (BPTT) forces sequential updates, creating unstable O(T) gradient paths.

What if we could train highly expressive, non-linear RNNs with flat, parallelized O(1) gradients?

It is now possible. 🧵

[Phoenix Yin @Phoenixyin13]

如果你想了解Transformer架构的硬伤，这篇今年4月的论文非常有洞见且及时。

这篇论文的核心吐槽就是，Transformer，即现在主流AI用的那种架构在长期记住和更新动态状态上天生有局限。

它就像一个超级聪明的一次性扫描器，每次看到一长串文字，就从头到尾扫一遍，找出关联。
但是，它并不擅长持续跟踪evolving state。

这次，论文用拓扑这种几何结构的数学角度证明，Transformer把状态越推越深，深度用完了就卡住了。
这是结构性的硬伤。

如果你在搞AI Agent，或者像我一样天天做Suno音乐生成，或者未来想做Music Tech，这篇指的方向很好。
未来好用的AI需要混入循环、recurrent机制，比如Mamba、RWKV，或者Transformer与循环的混合体。

2026年AI scaling下，这更像在为post-Transformer时代铺路。
OpenAI，Anthropic可能已在内部探索。为什么o1-style reasoning有效但贵已经在论文有所解答，同时，论文也预示着未来高效long-context不只靠更大KV cache，架构创新也是非常重要的一环。

作为AI交叉背景的同学，这能帮我更好理解Human-AI Interaction中state tracking的cognitive modeling问题。

读完这篇论文之后，科研上，我会优先看recurrent axis强的模型，比如Mamba、RWKV、looped transformers、coarse SSM。训练时，可以探索下multi-stage，先feedforward pretrain，再加recurrence fine-tune来解决效率问题。

[Lei Li @_TobiasLee]

镜像的问题是如何写论文，AI 可以解决 writing，但不能解决逻辑。
每次都会给 junior 的同学推荐清华刘洋老师的 slides:
https://t.co/TJgrEsLoJO
我自己也是常看常新，收获非常大

[paperpaper @paperpaper886]

最近在带入组的本科实习生，发现怎么读论文其实是科研训练里最容易被忽略的一步。
推荐一篇每个科研新人都该读的经典短文：S. Keshav 的 How to Read a Paper。

文章提出了非常实用的“三遍读论文法”：
第一遍，5 到 10 分钟快速扫读：标题、摘要、引言、章节标题、结论和参考文献。
目标是回答 5C：
Category, Context, Correctness, Contributions, Clarity。
也就是判断这篇论文是什么、和谁相关、假设是否合理、贡献是什么、写得清不清楚。
第二遍，认真读论文主线，但先跳过证明细节。重点看图表、实验设置、结果是否清楚、引用了哪些关键工作。
第三遍才进入深度理解：尝试像复现一样重建作者的思路，检查假设、方法、创新点和潜在漏洞。

放在今天看，这个方法和 AI 辅助读论文其实很契合。
第一遍可以让 AI 帮忙快速总结论文的研究问题、核心贡献和主要结论，但自己一定要判断这篇文章是否真的值得继续读。
第二遍可以让 AI 帮忙解释方法、实验设置、图表和不熟悉的概念，但不能只看 AI 总结。关键图表、实验设计和结果数字一定要回到原文核对。
第三遍可以让 AI 扮演 reviewer，帮你追问：这篇文章的假设是否成立？实验是否支持结论？有没有 missing baseline？有没有潜在的数据泄漏、评价偏差或过度 claim？

读论文不是“读完”就行。真正重要的是知道什么时候快速跳过，什么时候认真理解。
尤其在 AI 工具越来越强的情况下，科研新人更需要训练自己的判断力。
AI 可以帮你压缩信息，但不能替你决定一篇论文是否重要、是否可信、是否值得借鉴。

https://t.co/8gUc4HbLwR

[ℏεsam @Hesamation]

this Anthropic researcher wrote one of the best articles you can find here on AI research. obviously, Hamming's classic book is highly suggested to learn how to develop a "Research Taste".