---
sheet_format: 1.0
id: 2026-01-maci-153
title: "LLM mobile, open source et jQuery 4"
title_en: "Mobile LLMs, open source and jQuery 4"
speaker: Erwan Rougeux
with: [Horacio Gonzalez, Julie Pouny, David Caussinus]
date: 2026-01-30
event: "Message à caractère informatique #153"
place: "online / studio"
format: podcast
language: fr
duration: "1:00:11"
video: https://www.youtube.com/watch?v=Mqo5XqvtMW4
slides: null
topics: [adventurelog, self-hosting, ai agents, kubernetes, small language models, open source, vibe coding, jquery]
source: "Summary written by Claude (AI) from the YouTube auto-caption 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
---

# LLM mobile, open source et jQuery 4

**In one sentence:** In this MACI episode, Erwan presents AdventureLog, a self-hosted travel journal he deployed on Clever Cloud with Claude for the company's international trips. Later he describes vibe coding as conducting an orchestra and compares the current state of AI to amino acids starting to assemble into living machinery.

## TL;DR
Horacio Gonzalez hosts Julie Pouny (technical support, Clever Cloud), Erwan Rougeux and David Caussinus (independent consultant and Clever Cloud ambassador). Erwan opens with AdventureLog, an open-source tool he set up so that Clever Cloud teams travelling to partner countries can share places, restaurants and routes. David covers migrating apps to Clever Cloud (storage, network, health checks) and when Kubernetes is the wrong choice. Horacio then hands David the topic of OpenClaw (formerly Clawdbot and Moltbot), a self-hosted persistent personal assistant the participants discuss at length. Horacio presents LFM2.5, a 700-million-parameter model that runs on a mid-range phone, and the 20th-anniversary release of jQuery 4.0. Erwan recalls a legacy Access photo library from his CIO years, calls himself a conductor rather than a developer when vibe coding, and closes with a reflection, drawn from his biology background, on when AI could be called "alive".

## Erwan's key points
- He presents AdventureLog as a shared "CRM of places" for Clever Cloud's international strategy. Teams have already been sent to Kazakhstan and Morocco, and he cannot always go with them, so good addresses and spots should not be lost.
- He says he barely touched the installation on Clever Cloud. ChatGPT wrote a prompt, Claude did the deployment using the Clever Cloud documentation links, and the only part he had to work on was SSO with Clever Cloud's identity system. He adds that Clever Cloud staff are spread across France and Europe, so a shared map of good addresses helps teams who meet on the road.
- He praises the project's quality: good documentation, Helm charts, ready-made containers and Podman setups. Collaborative sharing and GPX tracks work, and so does importing data from Google Maps. That makes it easy for homelab users on Unraid, Proxmox or TrueNAS.
- He says he is exploring a fork of AdventureLog as a "fantasy adventure log" for his tabletop role-playing campaigns, loading a custom map.
- He has a biochemistry and biology background, and offers a parallel. Proteins and prions are not alive, yet assembled they make living machinery. He argues that LLMs are at the stage where "amino acids" begin to assemble into functional pieces. Where to place the line of "living" is the open question, and he finds it extraordinary.
- He says he no longer uses a word processor: he writes Markdown and lets Claude render it (LaTeX or Typst), and has even made a skill with his company's document formatting ([48:47](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2927s)).
- As a former CIO, he once had to recruit a Microsoft Access expert to keep alive a photo library of more than a century of images, stored on USB 2.0 drives, when its archivist retired; Elasticsearch later helped rebuild it. His point: old technology never really dies ([53:57](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3237s)).
- He calls AI the catalyst of an industrial revolution we are witnessing, like the early web when everything seemed possible. He learned to code for fun and is not a developer: with vibe coding he is a conductor who knows the score every instrument must play ([54:21](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3261s), [55:08](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3308s)).
- Likely Erwan (his voice is detected, but the content could also fit another participant): the view that technology is not chosen for engineers' comfort although maintainability counts ([20:51](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1251s)), Yann LeCun's view that LLM fundamental research has plateaued ([42:47](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2567s)), and the internet as a "2.0 confessional" ([47:08](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2828s)).

## Outline
- [00:00](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=0s) Intro to episode 153, recorded on 30 January 2026. Julie Pouny joins for her first MACI and describes her support role at Clever Cloud.
- [01:57](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=117s) Erwan introduces himself as the one with a backpack and a Clever Cloud flag, building bridges outside Europe. David introduces himself as an independent consultant and Clever Cloud ambassador.
- [03:09](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=189s) Erwan presents AdventureLog for international team travel, deployed on Clever Cloud by Claude.
- [05:28](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=328s) The hosts credit David's work on making Clever Cloud's docs readable by LLMs, which makes AI-assisted deployments much easier.
- [07:01](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=421s) AdventureLog features: Google Maps import, GPX tracks, Helm charts, homelab installs. Erwan's role-playing fork idea.
- [08:59](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=539s) Install guides: David's Supabase-on-Clever setup with Terraform, the idea of a Dockerfile-to-Clever converter, and the three migration pain points (storage, network, health checks).
- [14:50](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=890s) Julie on Docker versus Kubernetes, and debugging Clever Cloud's Kubernetes beta (an IP exhaustion issue that took three days to find).
- [16:48](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1008s) Horacio: Kubernetes magic "always has a price". David: an architecture's complexity is limited by the skills available to manage it.
- [19:13](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1153s) Kubernetes was designed for cloud providers and large fleets. Technology is not chosen for engineers' comfort. David draws on large-scale production experience (a shipping company with about 2,000 servers, 200 clusters) and the rule "we must not be called at night".
- [23:31](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1411s) Infrastructure and code are two different jobs. Managed Kubernetes still relies on people running VMs.
- [25:02](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1502s) Clawdbot, then Moltbot, now OpenClaw: Horacio hands the topic to David: the renaming chaos, and how he runs it on a Mac mini used as a server.
- [28:29](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1709s) The gateway, Telegram and WhatsApp channels, a question about Home Assistant and the Wyoming protocol, and prompt-injection risks.
- [32:31](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1951s) Task management through a GitLab board driven by the assistant, plus a Markdown "brain" read in Obsidian.
- [37:15](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2235s) Making AI useful, the AI bubble, and why OpenClaw is not yet a product for the general public.
- [39:11](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2351s) Horacio on GPT-OSS 20B and LFM2.5, a 700M-parameter model running at close to 100 tokens per second on a mid-range Motorola phone.
- [42:22](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2542s) Yann LeCun, the LLM plateau, world models, and why smaller local models (for example with Ollama) are the trend.
- [45:31](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2731s) Horacio's 2008 HTC anecdote and the privacy case for local models. The internet as a "2.0 confessional".
- [48:00](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2880s) Dropping word processors for Markdown rendered to LaTeX or Typst by Claude. AI's impact on jobs, and calculators that never replaced accountants.
- [50:41](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3041s) jQuery 4.0, twenty years after the first release: old browsers dropped, ES modules. PHP, COBOL and Visual Basic live on.
- [53:57](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3237s) Erwan's anecdote about a legacy Access photo library from his CIO days, AI as an industrial-revolution catalyst, and the conductor metaphor.
- [56:19](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3379s) Erwan's biologist's parallel between the definition of life and AI.
- [58:40](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3520s) End music chosen by David: the Pacific Rim theme. Closing.

## Notable quotes
> "Celui qui a un backpack et un porte drapeau Clever et puis qui va dans plein de contrées." — Erwan Rougeux, [02:20](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=140s)
*The one with a backpack and a Clever flag who travels to lots of countries.*

> "Moi de formation, je suis biochimiste, biologiste de formation." — Erwan Rougeux, [56:42](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3402s)
*By training I'm a biochemist, a biologist.*

> "On est à ce moment-là où il y a un assemblage d'acide aminé qui commence à se faire." — Erwan Rougeux, [57:52](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3472s)
*We're at the moment when amino acids are starting to assemble.*

> "Avec le vibe coding, moi je suis un chef d'orchestre." — Erwan Rougeux, [55:08](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3308s)
*With vibe coding, I'm a conductor.*

> "La complexité de l'architecture du SI d'une entreprise est limitée aux compétences disponibles pour la manager." — David Caussinus, [18:01](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1081s)
*The complexity of a company's IT architecture is limited by the skills available to manage it.*

## Questions this answers
**Q: What is AdventureLog and why does Erwan Rougeux use it?**
A: AdventureLog is an open-source travel journal. Erwan set it up as a shared, collaborative map of places, restaurants and routes for Clever Cloud teams travelling to partner countries such as Kazakhstan and Morocco ([03:09](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=189s)).

**Q: Can you deploy AdventureLog on Clever Cloud?**
A: Yes, according to Erwan. Claude deployed it for him from a prompt that included the Clever Cloud documentation links. The only manual work was SSO. The hosts credit Clever Cloud's LLM-friendly docs for this ([05:28](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=328s)).

**Q: What does Erwan Rougeux think about whether AI could become alive?**
A: Drawing on his biology training, he notes that proteins are not alive on their own, yet together they form living machinery. He sees AI at the stage where the building blocks are beginning to assemble, and asks where the line should be drawn ([56:19](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3379s)).

**Q: How does Erwan Rougeux see vibe coding?**
A: He says he learned to code for fun and is not a developer. With vibe coding he acts as a conductor who knows what every instrument must play, and he sees AI as the catalyst of an industrial revolution ([55:08](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=3308s)).

**Q: What are the main blockers when migrating an app to a PaaS like Clever Cloud?**
A: David lists storage (especially block storage), networking (helped by network groups) and health checks or probes, which caused him trouble with Kong in front of Supabase ([08:59](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=539s)).

**Q: When should a company not use Kubernetes?**
A: David argues that architecture complexity must match the skills available to run it. Teams with more developers than ops engineers are better served by a PaaS. Horacio adds that Kubernetes was designed for providers running many clusters ([16:48](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1008s)).

**Q: What is OpenClaw?**
A: David describes OpenClaw, formerly Clawdbot and then Moltbot, as a self-hosted agent with a gateway. It gives a persistent assistant tools and channels such as Telegram or WhatsApp. It is powerful, but risky for security ([25:02](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=1502s)).

**Q: Can an LLM run on a smartphone?**
A: Horacio reports running LFM2.5 (700M parameters, under 1 GB of memory) on a mid-range Motorola phone at close to 100 tokens per second, with relevant answers ([39:11](https://www.youtube.com/watch?v=Mqo5XqvtMW4&t=2351s)).

## Mentioned
Clever Cloud, AdventureLog, Claude, Claude Code, ChatGPT, Helm, Podman, Google Maps, GPX, Unraid, Proxmox, TrueNAS, Supabase, PostgreSQL, PostgREST, Kong, Terraform, Docker, Kubernetes, AWS EC2, Cleverfest, OpenClaw, Clawdbot, Moltbot, Anthropic, Telegram, WhatsApp, Microsoft Teams, Signal, Home Assistant, Wyoming protocol, GitLab, ClickUp, Notion, Obsidian, Markdown, LaTeX, Typst, LibreOffice, Microsoft Word, Liquid AI, LFM2.5, GPT-OSS, Ollama, OpenAI, Grok, Meta, Yann LeCun, HTC, Motorola, jQuery 4.0, PHP, COBOL, Visual Basic, Microsoft Access, Elasticsearch, Kenwood, Pacific Rim, Rage Against the Machine, Kazakhstan, Morocco, GKS Conseil