PROBLEM:
TRUTH

Last year, we explained model collapse not as a physical, but as an energetic consequence of altering the context of your AI model at scale (1). Eventually, all manufacturers have silently added measures that aim to keep that context fully referable and aligned with model’s pre-, and post training references.

This works well until it doesn’t, so the model producer knows they have put in a best effort, and leave the AI disclaimer to do the rest. They almost stop short of saying “if you want more accuracy, that’s on you”, because they know the work of correcting their AI systems drift goes up with the complexity of their models — and they are already covering implementation costs for keeping that complexity afloat.

For reference: what was necessary to keep the context of a GPT-4 sanitised was a computational overhead that is now as big as the model itself; sometimes bigger (2).

Some AI harnesses pick this up to added post AI compute cost and speed, but the problem never fully goes away.

For reasons well known to you and your organisation, your own references need updating. Company procedures, methodologies, protocols, tacit knowledge, they all have versions that reflect your current company policies «going forward». They all have versions that were in effect until a certain date, and might recur again in the future. They also have people that were and are responsible for their changing.

This is all part of a normal cycle of knowledge in an institution, and every institution has its own way of keeping up with it. From company lore preserving excel documents in sacrosanct folders to more modern and integrated approaches, this internal office drift is mitigated in ways that - once again: work well, until they don’t.

Also, "Final_FINAL_last_v3.mp4" (or .docx, .psd, .key) is the universal digital anthem of anyone who has ever worked in a creative or professional field. It is the classic lie we tell ourselves (and our clients!) when we are deeply hoping a project is finally over — and it’s not.

The final degree of coherence is given by the systems you peg your activity to.

In your daily work, it is most likely that your sources of truth are connected to realities outside your systems (AI systems, knowledge systems, operational etc). Those ones add one more constraint to your operation : they’re out of your reach for control so you can’t know when they are correct or not.

At this point, the AI inputs of your company can, and most likely will, form a three-body chaos problem

— at a cost that’s always coming too late.