Research findings and how we learn to work with AI.
16 fiches · 47 entities · Updated
What can models and agents actually do, and how do we learn to work with them? Evidence answers the first, pedagogy the second, and both are collected here. Empirical studies, benchmarks, and experiments measure capability — while the design of a benchmark often quietly decides its result, a caveat several entries return to. Beside them sit analyses of how developers learn to direct agents, which expertise stays essential, and how teaching adapts once generation is cheap. Mentoring programs and hands-on, time-boxed assessments appear alongside the studies. Understanding, not product, is the priority: each fiche records what was tested, what the numbers support, and what remains open.
Key figures
12 videos
MCP for Beginners · stated in source
16,8% of work-related messages and 43,5% of occupation-specific messages concern a task from another occupation
Research post by Anthropic's **Frontier Red Team** (authors: **Andrew Fasano**, **Marius Fleischer**, Cole McFaul, Robert Xiao, Tripp Gallagher), published on **September 29, 2026**; approximately **2,300 words**, six figures. It analyzes **GLM-5.3** from **Zhipu AI (Z.ai)**, an open-weight model, five months after the announcement of Claude Mythos Preview. **(A)** Capability: on ExploitBench (41 Chrome V8 bugs), GLM-5.3 builds a complete exploit in **50 out of 410 trials**, versus **56 out of 410** for Mythos Preview; on the internal memory-corruption benchmark, control-flow hijacking in **4%** of trials versus **6%**; Claude Opus 4.6 and GLM-5.2: 0. **(B)** Trials with human experts: a browser 0-day vulnerability chain culminating in reading the visitor's files; with **GLM-5.3-Flash**, an N-day exploit chain (CVE-2026-11645) bypassing PAC on ARM64, **20 minutes** of human attention, eight hours of compute, **$20.40** at API prices. **(C)** Safeguards: bypassed in **64%** of cases via a false red-team context, **92%** via reasoning prefill, **100%** via *abliteration* (about 2,200 GPU-hours, $4,400); the Claude models tested remain at 0%. **(D)** Recommendations: broaden defenders' access to frontier models, independent government testing. Cites NIST's **CAISI** evaluation (September 17): "the most capable open-weight model in cyber released to date," roughly four months behind the US frontier. Extends the corpus's notes on Mythos and Glasswing.
#GLM-5.3#Zhipu AI#Z.ai
Andrew Fasano · Marius Fleischer · Cole McFaul · Robert Xiao · Tripp Gallagher ; Frontier Red Team · Anthropic (billet de recherche du site anthropic.com).
X post by **Yann LeCun** (founder and president of **AMI Labs**, professor at NYU, former Chief AI Scientist at Meta, Turing Award), September 25, 2026, ~110 words. He replies to **Noah Smith**, who deemed a video *« already falsified »*, in which LeCun calls the idea of a *« country of geniuses in a data center »* *« complete BS »*. (A) Four test questions: where is the domestic robot, the voiture autonome de **niveau 5**, the car that learns to drive *« in a few hours of practice like any 17 year old »*, the system that understands the real world and learns fast *« like a house cat »*? (B) A concession: AI will match humans in every domain, and the technology **already surpasses them on a growing number of tasks**. (C) Two caveats: *« We're still far from that »*, and it won't be built on **LLMs**, confined to a text-interface role. A second post points back to his September 20 formulation: autoregressive LLMs, *« in and of themselves »*, will not lead to human-level AI. Until now, the corpus held LeCun only as a reference cited by others ([[deanwball-open-weights-decelerationnistes-kimi-2026-07-17]]), not as a source; the text reads alongside the skepticism in [[luc-julia-ai-controversy-statements-media-2025-08-22]].
#Yann LeCun#AMI Labs#NYU
Yann LeCun — fondateur et président d'AMI Labs · professeur à NYU · ex-Chief AI Scientist de Meta · prix Turing. Post X.
Product communication from **Anthropic** published on **September 1, 2026** on anthropic.com (~4,000 words, six sections, 22 testimonials from early-access partners). It announces **Claude Fable 5.1** (general availability) and **Claude Mythos 5.1** (verified access): *the same model, but with different levels of safeguards*.
#Claude Fable 5.1#Claude Mythos 5.1#foundation model
Anthropic — communication produit publiée sur anthropic.com · sans signature individuelle.
Entry by **Jeremy Norman** on **HistoryofInformation.com** (timeline "Exploring the History of Information and Media"), dated "circa **370 BC**" for the commented text, site last updated **August 27, 2026** (~600 words, including a long excerpt). It reproduces the passage from **Platon**'s *Phèdre* (274c-275b) where **Socrate** recounts the Egyptian myth of **Theuth** and King **Thamus** on the invention of writing.
#Socrate#Platon#Phèdre
Jeremy Norman — historien et éditeur de HistoryofInformation.com ; notice de frise chronologique citant Platon (*Phèdre*, v. 370 av. J.-C.).
X post by **Andrew Ng** from **August 14, 2026** (16:29 UTC), reprising the "Dear friends" letter from ***The Batch* #366** (DeepLearning.AI, same date), ~900 words. Ng presents **The AI Engineering Skills Map** and publishes **four skills** held to be the most important. **(1) Building and deploying AI applications** — the specificity is named: *« The key difference between AI and non-AI applications is that the former has unpredictable outputs »*, hence the emphasis on *evals* and error-analysis loops. **(2) Software engineering fundamentals**, because *« Understanding software fundamentals allows you to recognize what tradeoffs even exist »* — the inexperienced developer fails *« because they don't know what context to give their coding agent »*, hence the goal of *« steering coding agents using the precise language of software engineering »*. **(3) Using coding agents**, in an operational formulation: *« help the agent autonomously close loops by providing verifiers or evals »*, and *« knowing how much to intervene and how much to leave them alone »*. **(4) *Shaping the build***: *« Given a clear spec, coding agents are rapidly improving at delivering to it. Thus, our work as engineers is shifting toward deciding what should be in the spec »*, paired with *« Engineers should no longer expect to be given a pixel-perfect design and asked only to implement it. »* A **terminology note** carries most of the framing: Ng talks about **skills** in AI engineering and **not the role** "AI Engineer", with an explicit analogy — *« All developers today should know how to work with the cloud, and only a smaller number have a "Cloud engineer" title. »* The whole is backed by *« an analysis of more than 10,000 job postings, dozens of structured interviews with experts, hiring managers, and recruiters, surveys, and other online data »*, of which **no numeric results are published**: Ng describes his process as *« informally… akin to running clustering »* and announces a detailed map in future posts. He states the interest in the second-to-last sentence: *« DeepLearning.AI's principal focus is to help developers gain these AI engineering skills. »*
#AI Engineering Skills Map#skills map#Andrew Ng
**Andrew Ng** — fondateur de **DeepLearning.AI** · general partner d'**AI Fund** · cofondateur de **Coursera** et de **Google Brain** · ancien chief scientist de Baidu. Texte signé · à la première personne · écrit *« with my team »* sans qu'aucun collaborateur soit nommé. Publié le **14 août 2026** sur X et dans ***The Batch* n°366** — même texte aux deux endroits ; préférer *The Batch* pour toute citation durable. Quatrième fiche Ng du corpus · après les lettres n°350 (24 avril) · n°352 (8 mai) et n°359 (26 juin).
Post and report from **OpenAI Economic Research** published on **July 27, 2026**, the first installment in the **Work at the Frontier** series, analyzing **more than 800,000 messages from US ChatGPT users**. **Coined concept**: ***task crossover*** — *« work historically associated with one occupation appearing in the AI use of people in another »*. **The headline figure is actually two figures, and that's the point coverage loses**: **16.8% of work-related messages** concern tasks associated with another occupation, and **43.5% of occupation-specific messages**. The funnel explains the gap: **61.5% of usage is generic** (writing, summarizing, planning — too widely shared to count as evidence of crossover) and is excluded; of the **remaining 38.5%**, **43.5% fall outside the occupation** and 56.5% are *« inside **or near** »* — so the upper bound is calculated on a reduced base, while the lower bound is calculated on the entire professional usage. **By occupation** (share of occupation-specific messages pointing to an external task): customer experience **77%**, design **75%**, HR **69%**, legal **56%**, marketing **53%**, sales **40%**, finance **40%**, engineering **28%** — *« a majority in five of eight groups »*. **Two distinct directions of circulation**: design **imports** (35.2%) and **exports** almost nothing (1.7%); engineering does the opposite (imports 18.5%, exports 7.4%); **marketing does both** (imports 24.3%, exports **8.9%**, the highest outward share in the sample). **Two tasks appear in the top 3 of borrowings for the other seven groups**: **financial calculation** and **technology troubleshooting**. **The heatmap, absent from coverage, is the richest object**: it gives the full distribution of tasks by user occupation, and its diagonal is striking — engineering retains **53%** of its own work while customer experience retains only **11%**, HR **10%** and design **12%**. **Size effect**: the outside-occupation share drops from **18.9%** (2-5 employees) to **16.3%** (>100 employees) — **but only « among average users »**, OpenAI specifying that *« among the heaviest users, we do not see the same monotonic pattern »*, and concluding conditionally: *« AI **may be** especially useful as a generalist tool where specialist resources are scarce. »* **Claimed status**: an **early signal**, visible *« before firms rewrite job descriptions or create new job titles »*. **Structural caveat**: OpenAI measures OpenAI's own usage, on US ChatGPT users only, and presents this position as an asset — *« our unique window into how the world of work is changing »*.
#OpenAI Economic Research#Work at the Frontier#task crossover
**OpenAI Economic Research** — équipe de recherche économique d'OpenAI ; la page crédite simplement *« OpenAI »* et la classe sous les tags *Economic Research* et *2026*. Le billet est la porte d'entrée d'un **rapport PDF** (`work-at-the-frontier-report.pdf`) et s'adosse à un cadre antérieur de la même équipe · l'**AI Jobs Transition Framework** · dont il reprend la thèse que de nombreux métiers vont **se réorganiser** plutôt que disparaître.
A pedagogical article from **MindStudio Team** (the blog of the MindStudio platform, multi-model workflow orchestration) explaining **modèles de langage par diffusion** (*Diffusion Language Models*) through the case of **Diffusion Gemma**, Google's first **open weights** implementation (2B parameters, derived from Gemma 2). The thesis: where **autoregressive** models (GPT-4, Claude, standard Gemma) generate text **token by token, left to right** (causal attention, each token frozen once produced), **diffusion** models start from a **masked/noised** sequence and **refine it iteratively** (masked diffusion / *absorbing diffusion*), with **bidirectional attention**: the model can **revise any position at any step**. Consequences: high **parallelism** (a 500-token text would require 50-100 denoising steps instead of 500 sequential passes), natural **infilling** and **constrained generation** (template filling, code completion with surrounding context), and built-in **revision** capability. But at current scale (2B), Diffusion Gemma **does not match** the large autoregressive models (GPT-4o, Gemini 1.5 Pro) on reasoning, instruction-following, and general knowledge: the gap is "closing" without being closed. The inspiration comes from image generation (Stable Diffusion, DALL-E left autoregressive generation years ago); whether the same principle holds for text remains an open question. Diffusion Gemma is distributed on Hugging Face (Google DeepMind), AI Studio, and Vertex AI.
#modèles de langage par diffusion#Diffusion Gemma#Google DeepMind
Case study by **Surya Narreddi** (with **Cameron Franz** and **Alex Wang**) on his personal website, dated **March 2026** (day unspecified; ~1,500 words), titled *Training AI to Paint with Code*. The authors train a language model (**Qwen 3.5 35B**) via reinforcement learning (**GRPO**) to produce watercolor paintings as **p5.brush** JavaScript code, with the code itself serving as the editable artifact. The underlying question: how to do RL when the reward — aesthetic — is not verifiable.
#reinforcement learning#GRPO#reward function
Surya Narreddi — designer et développeur · auteur du projet (site personnel) ; collaborateurs Cameron Franz (infrastructure d'entraînement) et Alex Wang.
Wharton study (Generative AI Labs): expert personas don't improve LLM factual accuracy - GPQA Diamond and MMLU-Pro benchmarks - SSRN
#AI prompting#personas#LLM accuracy
Savir Basil · Ina Shapiro · Dan Shapiro · Ethan Mollick · Lilach Mollick · Lennart Meincke (Generative AI Labs, The Wharton School, University of Pennsylvania)