Site powered by Jekyll using the No-Good-Very-Bad template.

Policy regarding LLMs

I follow the code of ethics of major conferences and journals stating that (1) researchers are ultimately responsible for their contributions and (2) use of LLMs should be disclosed (e.g. as a non-native english speaker, I use LLMs to improve my writing). Also: I believe LLMs should not be used to write reviews. See for example ICLR or Springer policies on LLMs on the topic.

As LLMs are built by harvesting enormous amounts of human data, I consider them to be transformations and recombinations of human production (including some copyrighted material which, I hope, has been duly paid for). As such, I consider LLMs to be a common public good, resulting from collective human production. I believe LLMs should therefore always be open, meaning that their weights and the software necessary to run them should be publicly available. This does not exclude companies from making money with LLMs by providing services such as the infrastructure for running them, or support and help with deployment, rather like companies such as Canonical (with Ubuntu) or Red Hat can build businesses around Linux without owning Linux itself. I use subscriptions to proprietary services such as ChatGPT and Claude AI, but I strongly support general-purpose LLMs being released as open models, and I use some of these locally on my computer, such as DeepSeek, Llama, and Qwen.

Sustainable research

I am committed to taking at most one international flight (whether for work or personal reasons) every two years (tracker: 1 international flight since 2020), and to replacing all regional flights with train travel. I do use LLMs, though I favor text-based interactions, balancing between advanced models and lightweight models running locally.

Food for thought (more of a collection of sticky notes for me to remember):

  • a Paris-Tokyo round trip is about 3 tons CO₂e, which corresponds (very approximately) to 45 million tokens from DeepSeek-R1 - around 1.5% of the size of the english Wikipedia (about 3b tokens), or more than 1000 times the length of Lewis Carroll’s Alice in Wonderland (assuming roughly 30k tokens, 0.75 token/word). 3 tons CO₂e is also roughly what is needed to build 12 laptops or 40 smartphones, or to use a laptop for 60 years or a smartphone for 250 years. Or to build 28 road bikes.
  • “Why is it so much cheaper to fly than take the train? A big reason is tax. Unlike drivers and train travellers, fliers do not have to pay tax on jet fuel, and plane tickets are mostly exempt from VAT”. Citation from an article in The Guardian. More on that in this Greenpeace 2025 report (citation: “The most expensive country for cross-border rail travel compared to flights is France, where 95% of all routes were more expensive by train on at least 6 out of 9 days”).
  • of course, CO₂e is not all there is. Cost in terms of water consumption should be taken into account too, as well as the exploitation of finite resources, which can be expressed in antimony equivalents (Sb eq).

Misc

Is there a life after the PhD? Here are the slides I made for the ISIR Young Scientists Day in 2026 (i.e. PhD students and young postdoctoral researchers working in a French CS/robotics academic research lab).

And now for something completely different: I sometimes draw weird stories and used to make strange videos.