---
sheet_format: 1.0
id: 2026-09-maci-166
title: "Classificateurs, IA locale et CAO : l'IA sort du code"
title_en: "Classifiers, local AI and CAD: AI moves beyond code"
speaker: Erwan Rougeux
with: [Horacio Gonzalez (host), Ivan Dalmet (Peinture.app, BearStudio), Julie Pouny (Clever Cloud)]
date: 2026-09-23
event: "Message à caractère informatique #166"
place: "online / studio"
format: podcast
language: fr
duration: "1:00:13"
video: https://www.youtube.com/watch?v=33L1mgLU6FE
slides: null
topics: [learning to code, ai agents, local models, classifiers, open source, postgresql, 3d generation, robotics]
source: "Summary written by Claude (AI) from the local Whisper transcript, with statements attributed to Erwan by local identification of his voice only (other speakers are not identified). Not a verbatim transcript; check the recording before quoting."
review: spot-checked
review_date: 2026-10-18
---

# Classificateurs, IA locale et CAO : l'IA sort du code

**In one sentence:** A MACI episode on learning to code in the age of agents, AI-generated contributions, typed classifiers and local models; Erwan talks about teaching algorithmics, fencing in his own AI, generating 3D models, and the rise of household robots.

## TL;DR
Recorded on 23 September 2026 with the same panel as episode 165, the episode opens with a question carried over: how should a child start programming in 2026? The panel favours concepts and algorithmics first (paper and pen, Scratch, games such as Human Resource Machine), while noting that AI can either multiply learning or switch off thinking. The news covers the delay of PostgreSQL 19 after AI-assisted contributions had to be reverted, open-source projects writing AI-contribution guidelines, and Jev, a typed classifier from TypeSafe AI that returns decisions rather than text. Erwan explains why he built a gate between his AI and his Signal messages. The last third moves beyond code: Julie Pouny's local agents for CAD, Erwan's 3D generation and generative design, and robotics.

## Erwan's key points
- He has helped many schools teach coding: block programming with Scratch, including IoT projects with pupils, and Ozobots that children program with coloured cards. He says algorithmics should come before any language, like grammar before writing.
- He recalls learning with a "create your own game" box bought in a supermarket, dreaming of a Master System hit and ending up with a text game, but learning a lot.
- He went back to OCaml, which he last used at university, and says AI helped him relearn its concepts. He sees AI helping people improve in any hobby, cites video apps that coach your sports movements, and says AI reached every layer of society at once.
- He argues AI lets anyone with a problem at work build a tool for it. He says lawyers are losing simple work because people first check a document or draft it with AI, then ask an expert to review it.
- He finds the latest American models less sharp than before: they go too broad or rely on outdated memory, for example about firmware. So for every project he first gathers knowledge in a local folder, converts PDFs such as datasheets to Markdown with anydoc, and increasingly uses Pi rather than Claude Code, with DeepSeek or Kimi.
- He says an AI that syncs his work reads his Signal messages, but only a dedicated work folder: his private life stays private, even with a local AI. A simple rule failed on the first prompt, so he built a Docker application that acts as a gate. His lesson: an AI will quickly find its way around your rules.
- He uses Microsoft's TRELLIS.2 on his Mac to make sense of about 2 TB of 3D files and to generate assets, for example a damaged metal texture on a shield. He mentions HeroForge for figurines and Invoke AI for inpainting, then turns 2D mock-ups into 3D objects. Too much generated fur makes a figurine unprintable.
- He describes using Fusion 360's generative design to make a large metal hopper from his constraints. It produced the bill of materials, he bought the plates from a scrap dealer, and he still had to weld it himself.
- On robotics, he contrasts China's first robot games, where robots fell over, with this year's kung fu and housework. He links this to the data gathered by Chinese home devices and to ageing populations. He says he saw robots make a bed and walk a dog live at a show in Southeast Asia, while allowing it may have been scripted, and that he wants a robot.

His introduction and a few other lines (on the C++ versus assembler debate at 09:23, and on crashing Linux by copying commands at 11:21) are mixed voices and are not attributed.

## Outline
- [00:00](https://www.youtube.com/watch?v=33L1mgLU6FE&t=0s) Opening of episode 166, recorded on 23 September 2026.
- [00:30](https://www.youtube.com/watch?v=33L1mgLU6FE&t=30s) Introductions: Ivan, Julie (Clever Cloud support, two and a half years), Erwan (international growth at Clever Cloud).
- [02:11](https://www.youtube.com/watch?v=33L1mgLU6FE&t=131s) The carried-over question: what should a child who wants to program learn first in 2026?
- [02:38](https://www.youtube.com/watch?v=33L1mgLU6FE&t=158s) A participant: understand loops and variables first; the best first language is paper and pen.
- [03:35](https://www.youtube.com/watch?v=33L1mgLU6FE&t=215s) Erwan: Scratch, IoT with pupils, Ozobots; algorithmics before languages.
- [05:25](https://www.youtube.com/watch?v=33L1mgLU6FE&t=325s) Human Resource Machine and Factorio as ways in; Claude Code as a playground for solving problems.
- [06:56](https://www.youtube.com/watch?v=33L1mgLU6FE&t=416s) Assembler you no longer need to write; is Python still the right start?
- [08:56](https://www.youtube.com/watch?v=33L1mgLU6FE&t=536s) Erwan's first game-making box. Others recall Flash and ActionScript; AI is a higher abstraction that can also remove the thinking.
- [11:21](https://www.youtube.com/watch?v=33L1mgLU6FE&t=681s) AI multiplies learning when you ask how it did things.
- [12:50](https://www.youtube.com/watch?v=33L1mgLU6FE&t=770s) Erwan on relearning OCaml, learning hobbies with AI, and lawyers at [15:22](https://www.youtube.com/watch?v=33L1mgLU6FE&t=922s).
- [16:23](https://www.youtube.com/watch?v=33L1mgLU6FE&t=983s) AI alongside experts, never instead of them.
- [17:45](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1065s) Vibe-coded apps keep service companies busy; a rough prototype beats an idea that only exists in someone's head.
- [19:55](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1195s) PostgreSQL 19 delayed: a wave of AI-assisted patches brought bugs and had to be rolled back.
- [21:57](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1317s) Open-source projects write AI-contribution guidelines. The real problem is huge changes nobody can review, not who wrote them.
- [25:23](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1523s) Erwan: less sharp models, local knowledge folders, anydoc, Pi with DeepSeek or Kimi.
- [27:58](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1678s) The same model gives very different results with the wrong context.
- [28:55](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1735s) Erwan's gate between his AI and his Signal messages; a quick defence of Docker.
- [31:41](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1901s) A participant had AI write a rule-based local classifier for personal expenses: no data sent out, zero cost per run.
- [32:36](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1956s) OpenAI publishes six cases of misaligned agents.
- [33:04](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1984s) Jev from TypeSafe AI: natural language in, typed answers out (yes/no, scores, labels), fast and cheap.
- [34:38](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2078s) From hand-trained classifiers to LLMs to Jev; a risk score to decide when an agent needs a human, at [36:56](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2216s).
- [37:45](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2265s) System 1 versus System 2 thinking; a Builder.io test on computer use.
- [41:19](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2479s) Julie: with a friend, local DeepSeek and Kimi agents that design 3D assemblies, aiming at a full bicycle.
- [45:34](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2734s) FreeCAD's API, Blender driven by Claude; by hand is still faster for now.
- [48:08](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2888s) Erwan: TRELLIS.2 on a Mac, HeroForge, Invoke AI, from 2D mock-up to 3D object.
- [51:22](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3082s) CAD constraints are harder than sculpture; SolidWorks and Fusion 360 AI features.
- [52:26](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3146s) Erwan: generative design of a metal hopper in Fusion 360.
- [53:33](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3213s) Classifiers as the brain of robots, which need decisions more than text.
- [54:37](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3277s) Erwan: Chinese robotics, ageing populations, robots seen live in Southeast Asia.
- [57:11](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3431s) Julie's step-by-step plan: a cup, an axle, a screw and bolt, then wheels and the bike.
- [58:15](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3495s) Most tasks local, frontier models for planning; "you need crazy people".
- [59:22](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3562s) Julie's closing track: Gary Davis, "Death Don't Have No Mercy".

## Notable quotes
> "L'IA m'a vachement permis de me réexpliquer des concepts." — Erwan Rougeux, [13:20](https://www.youtube.com/watch?v=33L1mgLU6FE&t=800s)
*"AI really helped me get concepts explained to me again."* (on going back to OCaml)

> "Ma vie privée ça reste ma vie privée, tu vois, même si ça tourne en local." — Erwan Rougeux, [28:55](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1735s)
*"My private life stays my private life, you see, even if it runs locally."*

> "L'IA au bout d'un moment elle va vite contourner ton problème." — Erwan Rougeux, [31:12](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1872s)
*"Sooner or later, the AI will quickly find a way around your problem."*

> "Le meilleur langage de programmation pour commencer c'est le papier et le stylo." — another participant, [03:04](https://www.youtube.com/watch?v=33L1mgLU6FE&t=184s)
*"The best programming language to start with is paper and pen."*

## Questions this answers
**Q: How should a child start learning to program in 2026?**
A: The panel says concepts first: paper and pen, Scratch, robots like Ozobots, or games such as Human Resource Machine. Erwan stresses algorithmics before any language. See [02:38](https://www.youtube.com/watch?v=33L1mgLU6FE&t=158s) and [03:35](https://www.youtube.com/watch?v=33L1mgLU6FE&t=215s).

**Q: Why was PostgreSQL 19 delayed?**
A: As told in the episode, many AI-assisted patches brought new bugs, so changes had to be reverted before release. See [19:55](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1195s).

**Q: How does Erwan Rougeux stop his AI from reading his private messages?**
A: Rules in the prompt were not enough, so he built a Docker application that acts as a gate and only exposes a dedicated work folder of his Signal messages. See [28:55](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1735s).

**Q: What is Jev from TypeSafe AI?**
A: A classifier rather than a text generator: you send natural language with typed questions and get yes/no answers, scores or labels back in under a second, at low cost. The host suggests using it as a risk gate for agents. See [33:04](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1984s) and [36:56](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2216s).

**Q: Can local models do useful work beyond code?**
A: Julie describes local DeepSeek and Kimi agents assembling CAD parts in FreeCAD, still slower than by hand. Erwan runs TRELLIS.2 on his Mac to generate 3D assets. See [41:19](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2479s) and [48:08](https://www.youtube.com/watch?v=33L1mgLU6FE&t=2888s).

**Q: How does Erwan Rougeux keep AI answers accurate?**
A: He gathers sources in a local folder for each project, converts complex PDFs such as datasheets to Markdown with anydoc, and works with DeepSeek or Kimi in Pi. See [25:53](https://www.youtube.com/watch?v=33L1mgLU6FE&t=1553s).

**Q: Why does Erwan Rougeux think household robots are coming?**
A: Robots now handle complex movements and chores, Chinese home devices have gathered a lot of data, and ageing populations face staff shortages. He saw robots doing chores live at a show, and notes that housekeeping robots are now banned in the United States ([55:21](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3321s)). See [54:37](https://www.youtube.com/watch?v=33L1mgLU6FE&t=3277s).

## Mentioned
Clever Cloud, Peinture.app, BearStudio, Scratch, micro:bit, Ozobot, Human Resource Machine, Factorio, Python, JavaScript, Flash, ActionScript, C++, OCaml, COBOL, Master System, PostgreSQL 19, anydoc, Pi, Claude Code, DeepSeek, Kimi, Signal, SQLite, Docker, Debian, Anthropic, OpenAI, TypeSafe AI, Jev, Builder.io, ChatGPT, FreeCAD, Blender, RTX 5090, Microsoft TRELLIS.2, HeroForge, Invoke AI, Fusion 360, SolidWorks, Xiaomi, Gary Davis