Episode des news de la semaine précédentes avec beaucoup de sujets notamment la suite de l'intervention de Arthur Mensch devant la commission de l'assemblée nationale!

Arthur Mensch a été convoqué devant les députes pour une commission d'enquête sur la souveraineté. Son intervention est géniale et on y réagit.

- The distillation panic https://substack.com/home/post/p-196156125

# SpaceX rachète Cursor pour $60b? - Show Notes

# SpaceX rachète Cursor pour $60b?

## Episode Summary

In this news-packed episode, the hosts dive into the shocking $60 billion acquisition offer from SpaceX for Cursor, exploring Elon Musk's strategy to dominate AI compute. They discuss Anthropic's strategic moves including Cloud Design, OpenAI's legal battles, the explosion of open-source models, and the evolving role of data teams in an AI-first world. The conversation covers everything from GPU wars to the gamification of software development.

## Key Topics Discussed

- **SpaceX's $60B offer for Cursor** - Elon Musk's strategy to control AI compute through vertical integration

- **Anthropic's expansion** - Recruiting CTOs from major tech companies, launching Cloud Design, and limiting API access

- **OpenAI vs. Elon Musk lawsuit** - Legal battle over OpenAI's transition from non-profit to for-profit

- **Open source model explosion** - QwQ, DeepSeek's 1.6 trillion parameter model competing with Opus and GPT-4.5

- **Google's $70B investment in Anthropic** - $40B upfront plus $30B in milestones, including TPU access

- **Cloud Design release** - Anthropic's Figma competitor showcasing chat-first interface philosophy

- **Dynamic software interfaces** - YC's call for startups building AI-personalized user experiences

- **The future of data teams** - Shift from coding dashboards to writing agent guidelines

- **Agent Craft** - RTS game-style interface for managing software development agents

## Notable Quotes

> "Imagine if we had said 10 years ago that a company who makes spaceships would buy a company who makes an IDE... and that IDE is just a fork of Microsoft's IDE."

> "The problem with skills is that now everyone wants to have a skill for everything... people love having lots of skills, it gives the impression of having control, even though the information could be much more compact."

> "Between now and a year ago, for the same task I don't think quality changed much. But what really changed is the runtime - today you can tell an agent 'run until fixed' and it will actually do it."

## Resources/Links Mentioned

- **Cloud Design** - Anthropic's new design tool with chat-first interface

- **Agent Craft** - RTS-style interface for managing development agents

- **Upskill.ai** - Platform for selling AI skills (mentioned as project concept)

- **YC's Request for Startups** - Dynamic software interfaces theme

- **Basti UI** - YouTube designer's review of Cloud Design

- **AI Engineering Europe** - Conference where Agent Craft was presented

- **Project Deal** - Anthropic's Craigslist-style marketplace experiment with AI agents

## Guest Information

Regular co-hosts Julian and colleague discussing AI engineering news and trends. No external guest featured in this episode.

---

*This episode offers essential insights for anyone tracking the rapidly evolving AI tooling landscape, particularly around LLM applications, agent workflows, and the competitive dynamics between major AI players.*

# Le plus petit agent IA possible — pi Agent (l'agent derrière OpenClaw) - Show Notes

# Show Notes: Le plus petit agent IA possible — pi Agent (l'agent derrière OpenClaw)

## Episode Summary

Dans cet épisode nocturne du Engineering Podcast, Julien et son co-animateur décortiquent une conférence fascinante de Mario Schnecner (créateur de Pi Agent) donnée à AI Engineer Europe. Ils explorent comment construire l'agent de code le plus minimaliste possible avec seulement 4 outils de base, tout en discutant des dérives de Cloud Code, de la philosophie open-source face aux "clankers" (agents spammeurs), et des meilleures pratiques pour travailler efficacement avec les coding agents sans se faire submerger.

## Key Topics Discussed

- **Pi Agent** : Un harness d'agent de code ultra-minimaliste (40k+ stars GitHub) avec seulement 4 outils de base (read, write, edit, bash)

- **Evolution et critiques de Cloud Code** : Comment l'ajout rapide de features a créé des bugs et fait perdre la vision initiale du produit

- **Extensibilité vs complexité** : La philosophie de Pi qui privilégie un core minimal et des extensions TypeScript que l'agent peut auto-générer

- **Le problème des "clankers"** : Les agents IA qui spamment les repos open-source avec des PRs de mauvaise qualité

- **Stratégies de maintenance OSS** : La politique innovante de Mario (issues uniquement, validation humaine avant PR)

- **Bonnes pratiques pour travailler avec les agents** : Scope limité, code modulaire, review critique du code généré

## Notable Quotes

> **[Timestamp non précisé]** "We need better ways to fuck around. And for me, that means self-modifying mailable agents. Things that the agent itself can modify, and I can modify, depending on my workflow." - Mario Schnecner

> **[Timestamp non précisé]** "I used to work at construction sites and if my hammer breaks every day, I'm getting really mad. And if my development tools break every day, I'm also getting mad." - Mario Schnecner sur la stabilité des outils de dev

> **[Discussion podcast]** "It's crazy. When you look at the screen, it's in 2025... In the prompt system, there's an A in the middle [dans Cloud Code]. It's crazy to see a product like that."

## Resources/Links Mentioned

- **Pi Agent (Pi-mono)** : [GitHub repo](https://github.com/pi-ai/pi-mono) - 40k+ stars

- **AI Engineer Conference** : AI Engineer Europe (vidéos récentes sur YouTube)

- **Cloud Code** : Anthropic's coding agent

- **OpenClaw** : Utilise Pi Agent comme backend

- **Alternatives mentionnées** : AMP, FactoryDroid, Cursor, Codex

- **TerminalBench** : Benchmark pour coding agent harnesses

- **A.I. Documentary** : Documentaire à venir sur l'IA avec Sam Altman, Amodei, Demis Hassabis

- **Boring UI** : Projet à venir de l'un des hosts (open-source prévu)

- **Nao IDE** : Projet des animateurs inspiré de Cursor

## Guest Bio

**Mario Schnecner** - Développeur autrichien, ancien game developer et créateur de Pi Agent, le harness d'agent de code minimaliste utilisé par OpenClaw. Connu pour son approche pragmatique et son engagement envers l'open-source de qualité. Récemment rejoint Y Combinator comme fondateur. Maintient une politique stricte anti-spam sur ses repos GitHub et pratique des "maintenance holidays" pour gérer son projet open-source de manière soutenable.

---

*Note : Cet épisode est particulièrement recommandé pour les développeurs qui utilisent des coding agents au quotidien et cherchent à comprendre les fondamentaux d'un bon harness sans la complexité superflue.*

# Que devient dbt quand l’IA écrit le code ? (and news) - Show Notes

# Show Notes: Que devient dbt quand l'IA écrit le code ? (and news)

## Episode Summary

In this two-part episode, we explore how AI is transforming analytics engineering with Benoît Périgaud from dbt Labs, who shares insights on AI-powered code generation, the future of dbt, and whether traditional data skills will remain relevant. The second half covers the latest AI and data industry news, including Anthropic's Mythos release, managed agents, and the convergence of agent architectures.

## Key Topics Discussed

### Part 1: AI & dbt with Benoît Périgaud

- **AI integration at dbt Labs**: Skills engine, MCP server, and dbt Index for agent optimization

- **Token efficiency challenges**: Managing context windows for large dbt projects (1000+ models)

- **The future of analytics engineering**: Will agents replace the need to learn dbt?

- **Framework advantages**: Why agents prefer using frameworks like dbt over writing raw SQL

- **Benchmarking AI performance**: The 8Bench Analytics Engineering Benchmark and its limitations

- **Business context problem**: The missing piece in agent-generated analytics code

### Part 2: Industry News

- **Anthropic's Mythos model**: $25/million tokens, 5x price increase, cybersecurity focus, and controversy

- **Managed Agents**: Anthropic's new architecture for hosting agents with virtualized components

- **MCP vs CLI debate**: Token optimization strategies and implementation best practices

- **Cloudflare's approach**: Search and Execute tools instead of 2,600 API endpoints

- **Sandbox evolution**: The concept of user-specific runtimes embedded in applications

- **DuckLake V1 release** and other data ecosystem updates

## Notable Quotes

> "Most frontend models know how to write dbt... But what they know less is the process—how to think about creating a model." - Benoît Périgaud

> "Ruby on Rails is doing a revival because agents are really good at using frameworks... it uses a lot less tokens than doing it from scratch." - Benoît on why dbt will remain relevant

> "The intelligence wasn't scaffolding a fresh app—it was understanding an environment that is already stateful and using code as a mechanism to interact with it." - Discussing Cloudflare's sandbox philosophy

## Resources & Links Mentioned

- **dbt Tools**: dbt Skills, MCP Server, dbt Index (upcoming)

- **Benchmarks**: [8Bench](https://8bench.com) - Analytics Engineering Benchmark

- **Cloudflare Talk**: Sunil Pai's presentation at AI Engineer Europe on MCP optimization

- **Anthropic releases**: Mythos model, Advisor tool, Managed Agents, auto-approval mode

- **Google Gemma 2**: New open-weight multimodal model for edge devices

- **DuckLake V1**: Stable release with in-memory caching

- **State of Analytics Engineering**: Annual report from dbt Labs (released during recording)

## Guest Bio

**Benoît Périgaud** works on the Developer Experience team at dbt Labs, where he's spent four years building AI integrations for the platform. His recent projects include the dbt Skills engine, MCP server implementation, semantic layer benchmarks, and the upcoming dbt Index feature designed to optimize agent interactions with large dbt projects.

---

*Episode features two segments: AI/dbt discussion (beginning) and industry news roundup (second half). Use chapter markers to skip to your preferred section.*

# Claude Code leaked (on explique tout) - Show Notes

# Claude Code Leaked (on explique tout) - Show Notes

## Episode Summary

In this explosive episode, hosts Julien and guest discuss the shocking leak of Anthropic's Claude Code source code via an NPM registry mishap. The leak exposed 512,000 lines of frontend code, triggering a cascade of events including GitHub DMCA takedowns, a viral Rust rewrite that became the fastest-starred repo in GitHub history, and revealing insights into how AI coding assistants actually work. They break down the timeline, analyze what the code reveals about Claude's architecture, and explore the ironic copyright implications for a company built on training AI with internet data.

## Key Topics Discussed

- **The Claude Code leak timeline**: From the 4am NPM publication mishap to Anthropic's DMCA response

- **Ultraworker/CloudCode**: The viral Rust rewrite using Codex that gained 141,000+ GitHub stars in days

- **Code quality debate**: Boris Cherny's defense of "messy" AI-generated code and whether traditional code quality standards still matter

- **Feature flags and hidden capabilities**: Anthropic "ant" user types getting enhanced features and verification

- **Technical architecture**: Tool system breakdown showing surprisingly few tools (read/write, search, bash, MCP, sub-agents)

- **Token efficiency vs quality**: Why users switch between Claude Code and Codex based on limits and output

- **The copyright irony**: Companies trained on stolen internet data filing copyright claims

- **Supply chain security**: Axios attack and the continuing vulnerability of the open source ecosystem

## Notable Quotes

> **On the leak's impact**: "The thing I think is interesting is that the agents are super simple and the harnesses are all the same... It's the models that make the difference." - [~45:00]

> **On code quality evolution**: "Maybe the good practices or the good way of writing code will completely change now since it's not us who have to maintain it... this paradigm is changing so maybe we don't need to have such a clean code because it's not us who are going to modify it." - [~32:00]

> **On the copyright irony**: "Companies that stole the internet to train models complain when someone takes back their stuff that they made a mistake on... the notion of copyright on AI it only goes in one direction - we steal and then we privatize." - [~35:00]

## Resources/Links Mentioned

- **CloudCode NPM leak**: Version 2.1.88 (now deleted) contained full source in CLI.map.js file

- **GitHub DMCA repo**: 95 repositories received takedown notices from Anthropic

- **Ultraworker/CloudCode**: The viral Rust rewrite (currently ownership transfer in progress)

- **CloudCode Unpacked website**: Visual tool explorer showing architecture, patterns, and tool systems

- **Wall Street Journal article**: Coverage of Anthropic's response to the leak

- **Axios supply chain attack**: Another major NPM security incident this week

- **Boris Cherny's response**: Lead of Claude Code's post-mortem on the incident

## Guest Bio

This episode features regular co-host Julien discussing the week's developments in AI engineering with the show's primary host.

---

*AI Engineering Podcast - Episode 10 | Subscribe on Spotify, Apple Podcasts, and YouTube*

# Une librairie LLM attaquée vole tous les credentials? - Show Notes

# Une librairie LLM attaquée vole tous les credentials?

## Episode Summary

Dans cet épisode, l'équipe discute des dernières actualités de l'écosystème AI avec Alexis, notamment la récente attaque par chaîne d'approvisionnement sur LightLLM qui a exfiltré les clés API des utilisateurs. Ils explorent également l'évolution rapide des fonctionnalités d'Anthropic Claude, les limites actuelles des agents IA en production, et débattent de l'avenir de l'orchestration autonome des agents. Une conversation franche sur les défis réels de mise en production des systèmes IA et les risques de sécurité émergents.

## Key Topics Discussed

- **Attaque de sécurité sur LightLLM** - Compromission du mainteneur et injection de code malveillant exfiltrant les variables d'environnement et clés API

- **Nouvelles fonctionnalités Anthropic Claude** - Scheduler pour Claude Code, automatisation de tâches récurrentes, remplacement d'infrastructures Airflow

- **Défis de production des agents IA** - Les 5-10% de cas limites difficiles à gérer, problèmes de debugging, edge cases complexes

- **ArcAGI 3 benchmark** - Nouvelle version de puzzles où les modèles IA obtiennent 0% vs 99% pour les humains, démontrant les limites du raisonnement

- **Orchestration et refresh rate des agents** - Questions autour de la fréquence d'exécution optimale des agents autonomes

- **Stratégie OpenAI vs Anthropic** - Analyse critique des multiples initiatives d'OpenAI et focus produit d'Anthropic

- **Fine-tuning et spécialisation** - Tendance des entreprises à créer des modèles spécialisés pour leurs cas d'usage (call centers, etc.)

- **Stack Overflow pour agents IA** - Proposition de protocole Mozilla pour partager les solutions entre agents

- **Risques de skill poisoning** - Dangers potentiels de partage de compétences malveillantes entre agents

## Notable Quotes

> "Opérationnellement, je pense qu'il n'y a pas un jour où il n'y a pas quelqu'un qui annonce quelque chose. Aujourd'hui, j'essaie de regarder moins ce qui se passe ailleurs et je reste principalement sur la roadmap d'Anthropic."

> "C'est l'impression que j'ai, que tout ce qui est hard-codé, ça a de moins en moins de sens... C'est un peu le N8N de 2026."

> "Je pense que c'est ce qui nous reste, ces 5-10%. C'est comprendre les choses. Il leur manque le bon sens. C'est tellement cool de voir ça."

> "Peut-être que l'hallucination, c'est pas une feature qu'on sous-estime? On essaie tellement de les concentrer, de les bloquer dans nos canvas de workflow..."

## Resources/Links Mentioned

- **Claude Code** - Scheduler et nouvelles fonctionnalités d'automatisation

- **LightLLM** - Wrapper Python pour providers LLM (victime d'attaque de sécurité)

- **ArcAGI 3** - Nouveau benchmark de raisonnement par François Chollet

- **Mistral VoxTail** - Modèle Text-to-Speech OpenWeight (4 milliards de paramètres)

- **Stack Overflow for AI Coding Agents** - Protocole proposé par Mozilla

- **DSPY** - Framework utilisant LightLLM

- **Artifactory** - Solution de gestion de dépendances pour entreprises

## Guest Bio

**Alexis** - Praticien de l'IA engineering qui a participé à plusieurs épisodes du podcast. Il travaille sur l'intégration de solutions IA en entreprise, notamment avec Claude Code, et a une expérience significative dans l'automatisation de workflows et la gestion d'agents en production. Il partage régulièrement ses retours d'expérience sur les défis pratiques de mise en production de systèmes IA.

---

*Note: Cet épisode contient des informations critiques sur la sécurité des dépendances IA - essentiel pour tout développeur utilisant des librairies comme LightLLM.*

This week we are joined by Hussain Sultan, ex VP Engineering at Voltron Data (Ibis, Arrow, etc.)! Data engineering in the age of AI, xorq data pipelines using CLI.

# Cursor Composer 2, AgentMail, design bash tools et le futur des warehouses - Show Notes

# Episode: Cursor Composer 2, AgentMail, design bash tools et le futur des warehouses

## Episode Summary

In this news-focused episode, Julien and his co-host dive into the latest developments in AI engineering, from the Cursor Composer 2 controversy involving Kimi licensing to innovative approaches like bash-style interfaces for AI agents. They explore practical tools like AgentMail for isolating agent communications, discuss the shift from function calling to CLI-style commands, and debate whether warehouses like Snowflake and Databricks can maintain relevance in an agent-first world where OpenAI and Anthropic are positioning themselves as the new cloud platforms.

## Key Topics Discussed

- **Cursor Composer 2 licensing drama** - Controversy over potential unlicensed use of Kimi 4.2.5 model and performance comparisons with Claude Opus

- **Cursor's Instant Grep** - New search algorithm dramatically faster than ripgrep (13ms vs 16 seconds on large codebases)

- **AgentMail** - Temporary email addresses for AI agents to maintain isolation and security

- **Bash-style interfaces for agents** - Moving away from function calling to Unix-style commands for agent tools

- **Vercel's Bash library** - TypeScript implementation with multiple file system types (in-memory, overlay, read-write, mountable)

- **OpenAI acquiring Astral (UV/Ruff)** and Anthropic acquiring Bun - Big AI companies becoming the new cloud providers

- **The future of data warehouses** - Whether Snowflake, Databricks, and BigQuery can adapt to an agent-driven world

- **UI testing as the remaining bottleneck** - LLMs struggling with visual interface testing and location tasks

- **DBT Fusion going GA** - Static SQL validation for better agent loops in analytics

## Notable Quotes

> "I don't want to give access to my mailboxes to my agents. But the strategy to get around that is just to give temporary email addresses to agents for each project... I found the idea quite brilliant. It's really oriented for agents."

> "OpenAI and Anthropic become the new cloud of this age. I think it's them who are going to be the AWS, the GCP and the Azure of this age... They're basically creating the whole ecosystem so that you just have to do cloud code or co-work, deploy, and then it goes to Anthropic."

> "The more you talk to people, the more everyone is literally doing exactly [the same thing]. When we work on agents, everyone recreates the system, the tools, the file system, the orchestration, the browser screenshots, the memory, the prompts, the streaming, the test results, the logs."

## Resources/Links Mentioned

- **Cursor Composer 2** - Latest version with speed improvements

- **Cursor Instant Grep** - New fast search implementation

- **AgentMail** - Email inbox service for AI agents (free tier available, custom domains for developers)

- **Vercel Bash** - TypeScript-based bash implementation with isolated file systems

- **OpenClaude** - Slack/Telegram integration for Claude

- **UV/Ruff** - Python tooling from Astral (acquired by OpenAI)

- **Bun** - JavaScript runtime (acquired by Anthropic)

- **DBT Fusion** - SQL validation tool (GA in ~1 month)

- **Snowflake/Databricks** - Data warehouse platforms discussed

- **Manus** - Chinese cloud agent platform that pioneered bash-style interfaces

## Guest Bio

No guest in this episode - news discussion between regular co-hosts Julien and his fellow AI Engineering Podcast host.

---

*Note: This episode was conducted in French. Previous episode featured Hussein from XORC in English.*