J'ai lancé un LLM open source dans un avion : résultats choquants

J'ai lancé un LLM open source dans un avion : résultats choquants

I launched an open-source LLM on an airplane: shocking results

Title history
  • J'ai lancé un LLM open source dans un avion : résultats choquants — observed on 20261112-0057 UTC (YouTube)

by Thomas Berton · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Thomas Berton 👥 48K 📅 May 15, 2026 ⏱ 15 min 👁 2K 📄 tutorial 🧭 2026-11-12
Available in: English (current) Français

Keywords

Ollamalocal LLMClaude Codeoffline AIQwen 2.5 Coder

Summary

In this video, Thomas Berton attempts to run a local LLM in an airplane without internet to continue coding with Claude Code. He begins by explaining how to identify the best local model for one’s PC using a prompt that analyzes system specs. He then introduces Ollama, a platform for downloading open-source models, and installs it. After downloading Qwen 2.5 Coder (7B and 14B), he tests the models locally and gets quick responses. Next, he tries to connect Ollama to Claude Code via a terminal command. Despite initial success on the ground, once in the air, Claude Code fails to launch, likely due to missing offline capabilities or resource constraints. He concludes that using Claude Code with a local LLM requires a very powerful machine and is not practical for typical users. He also notes that the experiment highlights the heavy resource consumption of Claude Code. He suggests alternative approaches like using cloud-based LLMs with Claude Code for lower cost. The video ends with an honest debrief and an invitation to join his community for detailed guides.

177 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable practical insights into running local LLMs and connecting them to Claude Code. It demonstrates the steps clearly, including the command to fetch system info, using a prompt to select a model, installing Ollama, downloading models, and testing them. The argumentation is based on a real-world experiment, which adds authenticity. However, the reasoning for the failure is not deeply explored; the creator assumes it’s due to offline limitations without thorough debugging. The value lies in the realistic assessment of performance and resource usage, which is often missing in promotional content. The argumentation is straightforward but lacks rigorous analysis of the underlying causes.

Scientific Rigor, Source Quality, Title Accuracy

The video is scientifically rigorous in its approach: it uses a systematic method to choose a model, follows official instructions, and reports the outcome honestly. The sources are mainly official tools (Ollama, Claude Code) and the creator’s own community resources. No external scientific studies are cited. The title is somewhat sensationalist (‘shocking results’) but the actual results are modest, which could be seen as a mismatch. However, the video does deliver unexpected findings about the impracticality of local LLMs with Claude Code, so the title is acceptable. Overall, the rigor is moderate, as it relies on a single anecdotal test rather than comprehensive benchmarking.

224 words

Title / Content Match

The title suggests surprising results, and the video does deliver an honest account of the challenges, though the 'shocking' part is that it doesn't work well. The title is slightly exaggerated but not misleading.

Quality & Reliability

7/10

The video provides a clear step-by-step tutorial on setting up Ollama with local LLMs and attempting to integrate with Claude Code. The experiment is honest, including the failure to use Claude Code offline due to resource constraints. However, the technical depth is moderate and the conclusions rely on personal experience rather than systematic testing.

Chapters

Cited Sources

  • Ollama — Official website for downloading Ollama and accessing the model library.
  • Claude Code — Referral link for Claude Code, used for the integration attempt.
  • VibeSocials — AI-powered social media app mentioned as the project the creator works on.
  • AIOS — AI operating system service presented as an alternative in the description.
  • Free IA community — Free community for learning AI, mentioned for additional resources.
  • Crash course guide — Link to a detailed guide associated with the video's content.
  • Custom AI project service — Service for tailored AI projects, mentioned as a commercial offer.

Concurring Sources

External References

Contribution & Novelties

The video makes a practical contribution by testing a common workflow (running a local LLM with Claude Code) in a real offline scenario. It provides a concrete example of the challenges, resource consumption, and performance limitations. The originality lies in the honest debrief that the setup is not viable for typical users, which is more informative than many promotional tutorials. It also highlights the significant overhead of Claude Code itself.

Pour aller plus loin :

  • Ollama — The official tool for running local LLMs; essential for understanding how to manage models locally.
  • Qwen 2.5 Coder on Ollama — The specific model used in the video; useful for exploring its properties and variants.
  • Claude Code — Anthropic’s coding assistant that can be connected to local models. While not used here, its architecture and requirements are relevant to the discussion. (No URL due to uncertainty.)

143 words

Radar Profile

The scores show a balanced profile: quantity and quality are moderate-high, but technical level is slightly lower, indicating the content is accessible yet informative. Reliability is the weakest point, reflecting the single-experiment nature of the test. Overall, the video stands as a practical guide with honest limitations.

Reliability 6/10