Post by **Ethan Mollick** published on **October 1, 2026** on *One Useful Thing* (~2,000 words), subtitled *Benefitting from the Bitter Lesson*. Mollick revisits an anticipation he now considers mistaken: he thought that getting agents to work in groups would require lengthy human design, "akin to building a company."
#Bitter Lesson#bitter lesson#agent management
Ethan Mollick — professeur à la Wharton School (University of Pennsylvania) · auteur de *One Useful Thing* (Substack).
Article by **Asutosh Padhi**, **Brian Elliott**, **Daniel Swan**, **Kim Baroudy** and **Vik Sohoni** (senior partners, McKinsey), presented as expressing the views of **QuantumBlack, AI by McKinsey**, published in **October 2026** (exact day unspecified, dated the 1st by convention); 12 pages, 2 exhibits, a table, an architecture sidebar and two cases.
#AI operating system#AI OS#second brain
Asutosh Padhi · Vik Sohoni (Chicago) · Brian Elliott (Carolinas) · Daniel Swan (Connecticut) · Kim Baroudy (Copenhague) — senior partners · McKinsey (QuantumBlack).
A collective article by seven authors (**Zach Davis**, **Michelle Gill**, **Elisabeth Hendrickson**, **Angie Jones**, **Sha Ma**, **Randy Shoup**, **James Wickett**), published in the *Enterprise Technology Leadership Journal* by **IT Revolution** (fall 2026, vol. 3, no. 2), under a CC BY-NC-SA license. About **7,500 words**, titled *« Toward a Pattern Language for Code Reviews in the Age of Agentic AI »*. **(A)** Framing: code review is defined as an examination by a non-author entity leading to an action; empirical studies (Microsoft 2013, Mäntylä 2009, Beller 2014) show that its value lies mainly in maintainability and knowledge sharing, less in functional defect detection. **(B)** Three conditions for taking the human out of the critical path: maximal deterministic validation, **risk-graded** review, a **durable trace** of every agent decision; *« no human in the loop »* is a configuration choice, not a default. **(C)** Nine composable patterns, presented as a toolbox rather than a maturity ladder: small changes, Shift Left, design for verification, deterministic engineering practices, context engineering, continuous quality loop, adversarial LLM review, multi-agent reviews, specification-driven development. **(D)** Three case studies: **Block**, **Cloudflare** (131,246 reviews across 48,095 PRs in one month, 1.2 findings per review), **DryRun Security**. Extends the corpus's notes on context engineering and agent-assisted code review.
#agentic code review#code review#pattern language
Zach Davis (LaunchDarkly) · Michelle Gill (Engineered by AI) · Elisabeth Hendrickson (Curious Duck) · Angie Jones (Agentic AI Foundation) · Sha Ma (Topogy) · Randy Shoup (CircleCI) · James Wickett (DryRun Security) ; article collectif · IT Revolution.
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).
Long thread published on X on **September 29, 2026** by **Ryan Sarver** (@rsarver), titled *"Personal AI Should Actually Be Personal"*; roughly **1,700 words**, 91.5k views at time of collection. Starting point: **Amazon** blocked **Muse** (**Meta**) from its shopping site, and users lost access to their shopping agent, *"a side effect of a dispute I had no say in."* **(A)** Definition: a personal AI must be **owned** (open source, data in controlled storage, portability), have **a single principal**, offer a readable contract and inspectable trade-offs, run wherever one chooses, and be pleasant to use. **(B)** Observation on the current landscape: Muse opts users into training by default, **Instinct** carries a *"perpetual and irrevocable"* license; estimated cost gap of **$3,000 to $7,000 per user per year** against $20 a month. **(C)** A reversal: the owned stack can end up better, because the user gives it *"every corner of your life."* **(D)** A five-layer stack (clients, harness and connectors, memory, two-tier models, machine) and a call to build a productized version of **OpenClaw**. Responds to **Brad Gerstner** on trust; extends the body of work on shopping agents and markdown memory.
#personal AI#personal AI#personal agent
Ryan Sarver (@rsarver) · auteur du fil sur X ; fonction non indiquée dans le texte.
Long-form interview (**25-minute read**) with **David Heinemeier Hansson**, creator of **Ruby on Rails** and co-owner of **37signals**, by **Vikas Shah** in *Thought Economics*, published on **September 28, 2026**, five days after the opening keynote of **Rails World** in Austin where he announced that 37signals was going *"pencils down"* on hand-written code. **(A)** The reversal, dated: in summer 2025 he told **Lex Fridman** he would not let AI write his code; what changed his mind was the arrival of coding agents in late 2025, and *"the change has really become vertical, in my opinion, over the last three or four months"*. **(B)** The underlying argument is not technical but political: programmers had become *"a clerical class"* mediating access to the machine, and the agent plays the role of Luther — *"Agent Luther"* — by removing the intermediary. He draws a consequence for himself: his coding skills are no longer *"economically viable as an independent lever"*. **(C)** The sorting of business models: client-side is *"already gone"* (Adobe), client-server still holds because *"people have no desire to run their own servers"*, plumbing moats hold (Shopify, *"payment, shipping, and tax systems"*), and **Salesforce** and **SAP** are *"in serious trouble"* since *"there's no more value in the code, only in process intuitions"*. **(D)** **Omarchy**, his Linux distribution with built-in agents: the **Quattro** version written exclusively by agents, **200,000 downloads in 18 days**, **4,000 plugins** in four weeks, **$21.7M** in donation pledges. **(E)** The forward-looking close: AGI reached, *"pockets of ASI"*, robots, and *"what do you do when you have nothing left to do?"*. To be read against [[lecun-still-far-from-human-level-ai-llm-2026-09-25]] from the same month.
#end of hand-written code#pencils down#coding agents
David Heinemeier Hansson · créateur de Ruby on Rails et copropriétaire de 37signals · interrogé par Vikas Shah dans *Thought Economics*.
**Playbooks** article by **Lance Martin**, published on the claude.dev blog on **September 28, 2026** (12-minute read, nine figures). It presents two subcommands added to Claude Code's **claude-api** skill, `/claude-api build-eval` and `/claude-api hillclimb`, after laying out the principles they apply. Three contributions. **(A)** Four properties of a good evaluation: tasks representative of production, scores that rise with the model and the effort, headroom below **100%** at the top, low variance from one run to the next; plus *adversarial sampling*: a hard case is kept because a human can say why, not because today's model fails it. **(B)** Disciplined hillclimbing: one change per round, separate **train/test** set, patch reverted if train rises while test stays flat. **(C)** Two measured examples: a customer-support benchmark (**44** tickets) whose cost drops from about **4.6 to 1 cent** per ticket, and the evaluation of the claude-api skill itself, from **66%** to about **88%**. A line from the text: *"it's important to read a sample of scored transcripts before trusting your evaluator."* Related to the author's piece on prompt caching, [[lancemartin-anthropic-prompt-auto-caching-claude-2026-02]].
#evaluation#eval design#hillclimbing
Lance Martin (Anthropic) · blog claude.dev · avec remerciements à Misha Khalman pour le développement de la skill.
Technical documentation page from **GPT Researcher**, *"Context Filter"*, rewritten on **September 26, 2026** by **Assaf Elovic**, the project's creator, the day he made **Jev** (a model from **TypeSafe**) the default context filter. About **2,200 words**, including a reproducible benchmark. **(A)** The design thesis: the context filter is *"the biggest lever on report quality after the research itself"*: *"what it discards, the writer never sees"*. **(B)** The mechanism: each scraped page is split into **1,000-character** chunks. Jev scores each chunk on a four-level usefulness scale (0 to 3), and only chunks at **≥ 1.5** are kept, ten at most. Without an API key, a local **BM25** fallback with a relative threshold takes over. **(C)** The benchmark: 28 tasks replayed on the same pages. **73%** relevant passages for Jev versus **46%** for embeddings, **15-10-3** in head-to-head, at the same cost (~$0.115 per report). **(D)** The most transferable finding: *"the threshold matters more than the ranking"*. Without a threshold, Jev drops back to 50%. The gain comes from less noise, not more content. Extends, on the web-search side, the comparison of retrieval strategies from [[comparethemarket-context-retrieval-ai-code-review-gkg-rag-2026-03-06]].
#context filter#context filter#passage selection
Assaf Elovic · créateur de GPT Researcher · auteur de la page (documentation officielle docs.gptr.dev, non signée).
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.
Talk by **Eno Reyes** (co-founder and CTO of **Factory**, whose "droids" agents are described in the introduction as used by hundreds of thousands of developers and companies such as **Nvidia** and **MongoDB**, right after a **$150 million Series C**), published on **September 22, 2026** on the **FirstMark Capital** channel (25 min 51 s video, transcript of about 5,300 words). He presents the "software factory" as an agent-driven feedback loop, and the role of humans as stewards of this factory.
#software factory#software factory#Factory
Eno Reyes — cofondateur et CTO de Factory ; conférence vidéo (FirstMark Capital).
Article by **Beth Stackpole** in *Ideas Made to Matter* (**MIT Sloan**), published on **September 22, 2026** (~5 min), presenting the *research briefing* « Designing Decision Rights for AI » by **Ina Sebastian**, **Peter Weill**, **Thomas Haskamp** and **Jan vom Brocke** (**MIT CISR**), built from interviews with **30 executives**.
Beth Stackpole — rédactrice · *MIT Sloan Ideas Made to Matter* ; recherche de Ina Sebastian · Peter Weill · Thomas Haskamp · Jan vom Brocke (MIT CISR).
Article by **Jens Kengelbach**, **Daniel Friedman**, **Lianne Pot**, **Georg Keienburg** and **Tobias Söllner** (**Boston Consulting Group**), published on **September 21, 2026** in the series *The 2026 M&A Report* (~12 min). The displayed title is "How Non-AI Companies Win at AI Dealmaking"; the page and URL title is "How Non-Tech Companies Should Buy AI in 2026."
#AI M&A#M&A#AI dealmaking
Jens Kengelbach · Daniel Friedman · Lianne Pot · Georg Keienburg · Tobias Söllner — Boston Consulting Group ; série *The 2026 M&A Report*.
Radio interview of **54 minutes** with **Philippe Corrot**, co-founder and co-CEO of **Mirakl**, on *Les Pirates de l'IA* hosted by **Éric Hazan** (RCJ), the season-opening episode broadcast on **September 20, 2026** (volume 10), in three parts separated by musical breaks. **(A)** The career path, dated: **Keyrus** in 1997 (web agency, sold to **Progiware** in 2000), **Mayrev** in fashion accessories, then **Splitgames** — a video-game marketplace — bought by **Fnac** in 2008, whose two partners ran the marketplace for three years before founding **Mirakl** with **Adrien Nussenbaum**. **(B)** The vendor's exposure to what he calls the *"SaaSpocalypse,"* broken down into two mechanisms: **per-seat** pricing, from which Mirakl is exempt since it takes a percentage **on transactions** — *"whether it's an AI buying or a human customer, it makes no difference to us"* — and the rewriting of a piece of software by a model, which he considers unlikely in the short term because *"Mirakl isn't just lines of code"* but an infrastructure connecting nearly **100,000 sellers**. **(C)** The internal transformation, with a method starting from **deliverables** and working back up the chain, the merger of the documentation, support, and training teams, and the **OneHelp** agent: at constant headcount, customer satisfaction **×2** and support users **×10**. **(D)** The forward-looking part, where he answers *"yes"* to the question of a *job apocalypse*, treats **universal basic income** as a question worth asking, and blames French labor law and the lack of a French vision. Caveat set by the guest himself: *"what I'm telling you is true for the next six months."* Continues on the vendor side with [[bain-ai-rule-of-40-headwinds-tailwinds-saas-2026-04]].
#digital marketplace#marketplace#platform model
Philippe Corrot · cofondateur et co-CEO de Mirakl · interrogé par Éric Hazan dans *Les Pirates de l'IA* (RCJ).
Essay by **Asma Mhalla**, political scientist, published on **September 20, 2026** on her Substack newsletter *Dangerous Ideas* (~2,300 words, five parts titled in English), written on returning from a European tour — Madrid, Milan, Brussels — where within the same hour she was discussing AI, Meloni, the AfD, and sovereignty. She argues that September's AI sequence and the rise of nationalist parties are **two faces of the same anxiety**: the fear of becoming superfluous. **(A)** The cross-cutting observation: the viral resignation of **Jacob Coxon** (ex-OpenAI/Anthropic) on September 9, **Hinton** estimating that there is *"maybe a year"* left to put guardrails in place, divided executives (Amodei, Altman, Huang) — and, at the ballot box, **AfD at 43.8%** in Saxony-Anhalt, Vannacci's **Futuro Nazionale** around **7.5%**, Le Pen at **34-36%** in the first round. She withholds judgment on the probability of extinction and defers to the rapport international de sécurité de l'IA. **(B)** The mechanism: the word *replacement*, which first denotes a sensation — *"something could carry on without me"* —, and **retrofuturism**, read through **Jeffrey Herf**'s *Reactionary Modernism* (1984): the authoritarian right never fought technology, it detached it from the market and handed it to the nation. *"Technology is never the enemy. It is spoils."* When AI's producers say they no longer know how to govern it, they supply the vocabulary (control, frontier, sovereignty) to those who promise restoration. The thesis is not an electoral causality but a **narrative compatibility**. On the European side, extends [[wallace-wells-nyt-magazine-ai-populism-altman-backlash-no-one-ready-2026-05-08]].
#fear of becoming superfluous#replacement#narrative compatibility
Announcement by **Diogo Almeida**, founder of **TypeSafe AI** (formerly of OpenAI, where he worked on instruction-following training methods), published on **September 15, 2026** on the company blog (~2,500 words). The text introduces a new class of models, **System One Models**, and their first model, **Jev**, in early access.
#System One#decision model#Jev
Diogo Almeida — fondateur de TypeSafe AI · ancien chercheur chez OpenAI (billet du blog d'entreprise).
Long-form article published on **September 14, 2026** on the X account **@GoogleCloudTech** (~2,000 words, four steps, six blocks of commands and configuration), signed by **Alan Blount**, Technical Solutions Consultant at Google. Starting point: a trivial question, handed to a research agent, billed at **$38**. The article proposes running **Claude Fable 5.1** and **Gemini 3.8 Flash** side by side behind the same **Gemini Enterprise Agent Platform** API (formerly Vertex AI). **(1) Configuration**: access to Fable is subject to Google Cloud's *Advanced AI Safety Addendum*, which requires a `setPublisherModelConfig` call with prompt-response sharing enabled for Anthropic, on pain of a 403; Flash works as soon as it is enabled. **(2) In-house benchmark** with promptfoo driving opencode on two tasks: on a git branch state, Flash answers in **1.1 s** for under a tenth of a cent while Fable costs **23 times more**; on a Postgres-to-Spanner migration, Fable produces in **12 s** a formal DAG with clock-drift risks and a rollback gate, whereas Flash delivers a linear sequence in 3 s. **(3) Usage**: *« asymmetric coordination »* — the fast model as first-line coordinator, the deep planner at checkpoints, an `ask_for_help` tool for escalation, and a warning against dynamic routers: *« Keep it boring »*. **(4) Token ROI**: the value of human time, shipped features, and incidents avoided, minus the cost of tokens and the team's ramp-up. The same task-boundary routing move as [[mazmanov-portal-spotify-shunt-claude-code-tokens-2026-09-03]], within the FinOps framework set out by [[rafal-wenvision-tokenomics-foundation-finops-ia-2026-06-04]].
#model routing#asymmetric coordination#tokenomics
Alan Blount (Technical Solutions Consultant, Google Cloud, @zeroasterisk) · publié sur le compte X @GoogleCloudTech.
Essay by **Donn Felker** published on **September 14, 2026** on his Substack (~4,400 words, subtitle *"Your identity was built on a bottleneck"*). The author, AI platform lead at **Polygon** and formerly at **Tinder**, an Android specialist for over ten years, dates his shift to **July 2025** based on three facts: a SaaS built without writing a single line, a threading library migration at Tinder cut from **6-8 hours** to **10 minutes** by a **2,000-word** prompt, and non-technical people (a CFO, a gym manager) shipping complete products. **(A)** He draws from this a psychological framework — *occupational categorization*, *professional identity centrality* — to explain engineers' anxiety: *"It was never about coding, it's about the ability to execute"*. **(B)** He proposes an image of a **leverage pendulum** between product and engineering: AI shifts the initial leverage toward product, complexity and scale pull it back toward engineering, and each oscillation dampens as guardrails and architecture get codified. **(C)** He describes a hybrid role, the **agentic builder**, and a two-step path (codifying one's knowledge into skills, shifting from specialist thinking to systems thinking) with a stated horizon of **six months**. Extends the shift to *agentic engineering* described in [[karpathy-vibe-coding-agentic-engineering-software-3-0-2026-04-29]] and the collapse of boundaries between professions from [[sfeir-ia-frontieres-metiers-skill-based-organisation-2026-08-01]].
Engineering post from **Anthropic** published on **September 14, 2026** on claude.com/blog (~1,400 words, five sections), authored by **Sachin Malhotra**. It recounts scaling the **test impact analysis** service — selecting which tests to run per change — under the pressure of agentic coding. The input numbers and their chain: engineers ship on average **8×** more code per quarter than in 2021-2025, **Claude writes 80% of it** and also carries significant weight in PR review and approval; the number of tests was multiplied by **10** with nearly constant headcount; hence **25× CI jobs in six months**. **(A)** The v0 architecture rests on two components that must stay synchronized — a *listener* that records the results of each run, a *selector* that decides which tests run on each PR — but as a **single process**, a per-test history assuming a single writer. **(B)** The failure mode is the listener falling behind the PR queue: *« 20 minutes of listener lag can translate into tens of thousands of test updates not being applied »*, with three named consequences — wasted investigations, *flaky reds* that block merges, a fixed or added test that doesn't run. **(C)** Three patches whose lifespan collapses — **70 days, 29 days, less than a day**: a bigger machine, sharding by package, daily restarts. Then the overhaul: an **in-memory** magasin de données en mémoire, **stateless** *listener* workers that are therefore horizontally scalable, a log aggregated every few seconds. **Three weeks for one engineer; a quarter, a year earlier.** Extends the 80% of [[clinton-anthropic-secure-ai-native-sdlc-2026-07-21]] through infrastructure.
#continuous integration#CI#test impact analysis
Sachin Malhotra (ingénieur, Anthropic) · sur le blog claude.com.
Threat intelligence report published by Anthropic's **Threat Intelligence** team on **September 10, 2026** on anthropic.com (~36,000 words including indicators of compromise, downloadable PDF, unsigned). It covers abuse disrupted between **December 2025 and August 2026** across **seven** domains: cyber, influence, surveillance, conventional weapons, biology, fraud, distillation. All cases involve Haiku, Sonnet, or Opus; none involve Fable or Mythos, except for one distillation attempt.
#threat intelligence#AI misuse#Generative Threat Group
Anthropic — équipe Threat Intelligence (rattachée aux Safeguards) · rapport institutionnel non signé publié sur anthropic.com.
First article in the **Never Send The Slides** series on martinfowler.com, published on **September 8, 2026** (~2,400 words, five sections and two sidebars), authored by **Sumeet Gayathri Moghe**, global head of culture and organisational design at **Thoughtworks**; **Martin Fowler** provided editorial review. The text first lays out a definition — *« a presentation is an act of storytelling in which a presenter orchestrates narrative, timing, and emotion to create meaning that neither visuals nor documents can achieve on their own »* — and its corollary: *« the slides are not the presentation »*. From this it derives a hierarchy of five artifacts, condensed at the end into a *cheat sheet* table. **(1)** The **linear document**, sufficient in most business situations where the only goal is to convey information. **(2)** The **infodeck** — the term is Fowler's — when the content has a long shelf life and deserves visual care. **(3)** **Applications web interactives**, to manage the reader's cognitive load. **(4)** **Recorded audio or video**, when the voice matters but not the interaction. **(5)** The **live presentation**, reserved for cases where synchronous presence serves to influence and address objections in real time. Two observations date the text to 2026: **coding harnesses** put the creation of interactive applications within reach of non-developers, and a linear document *« reduces the probability of hallucinations »* when a reader has an LLM summarize it. Same source as [[boeckeler-harness-engineering-coding-agents-2026-04-02]].
#presentation#slides#slideware
Sumeet Gayathri Moghe (global head of culture and organisational design, Thoughtworks) · publié sur martinfowler.com ; relecture éditoriale de Martin Fowler.