In order to collaborate with me, Fabien Benetou, professionally or personally, with any tool relying on "artificial intelligence" or AI for short you must :
Note that this about AI but it could as well be about XR, 3D printing or basically any other emerging technology. In fact it should be about any technology, not just emerging. AI just happens to be extremely popular when writing this. It is extremely popular compared to other emerging technologies in the past so this is also a potential "shield" for whatever next hype cycle will bring.
It is about avoiding abuse and consider genuine empowerment, not AI. If AI though is challenging that, then its usage must be reconsidered.
- insure the model runs offline, no API call, on an open system, e.g. GNU Linux
- show that no specific hardware is required, solely CPU and GPU generic computer resources are required
- provide all the source code any model used
- provide all the training data the model used for training
- provide all the processes used for training, in particular but not limited to data annotation
- if any data annotation, e.g. labeling, was used during the training, you must provide a certificate for a third party authority proving that the labor was respectful of labor law in EU
- the model and all the tools required to train it are under open source (e.g. MIT) or free software licences (e.g. GPL, aGPL, etc)
- the model card must provide the CO2eq or another metric facilitating comparisons
- in a format that is computationally parseable,
- provide at least 1 analogy for a non professional,
- includes all training runs, not solely inference
- any generated content resulting from the model must be labeled as so, maintaining provenance at all time, respecting the license that all authors included in the dataset stipulated
- no anthropomorphisation can be used either anywhere, neither in the tool itself, the way to rely on the model (e.g. prompt or documentation) nor any communication
- if any content, e.g. text, video, code, etc is generated from using a model it must be labeled as generated and remained as such, always, including in future usages, and include a reproducible setup, including at least the model name, version, prompt, heat parameter and anything else required to get the exact same result again
- all tools included in the collaboration must respect those constraints. This is not limited to programming or prototyping work.
If at any point during the collaboration any of those point is not validated the collaboration will cease immediately including on the broader scope of the project.
You must read, understand and agree with my own work on the topic, namely but not limited to
Model verified so far
Model partly verified so far yet still in use
Model suggested so far which didn't pass and why
See also
Fabien Benetou's PIM
