Spinoza’s portrait by Stable Diffusion
A few years back, I had devoured Frédéric Lordon’s book Capitalisme, Désir et Servitude. In particular, I’d been fascinated by the notion of conatus, the desire for power. Frédéric Lordon explained that Spinoza, as a mathematician, had defined conatus in the form of mathematical vectors, enabling him to transform his philosophical thinking into mathematical axioms and theorems.
And it was here that I was struck by the resemblance to transformers.
Indeed, a transformer takes as input a succession of tokens or words, which it replaces with their embedding. An embedding is a huge vector that has been assigned to each word or token, enabling them to be compared with each other. So we can, for example, perform mathematical operations such as “boy = woman – man + girl”.
The transformer then combines these vectors via a succession of mathematical transformations learned during training to generate a new vector V. This vector is then compared with the set of embeddings to identify the closest to generate the next word.
[E1 E2 E3 … En] –> TRANSFORM –> V –> next token
It thus takes as input a sequence of words for which it constructs a true interpretation in the form of a new semantic vector V. At the next iteration, it injects this word at the end of the context and resumes its generation until it ends up with the complete text.
I can’t help seeing the succession of V as a form of conatus. Indeed, a model learned with a reinforcement algorithm is intentioned, unlike other models, which are supervised. This intention is defined in particular by RLHF (Reinforcement learning from human feedback), which forces the model to stay in line, to align itself with human desires, their conatus according to Spinoza.
This process can be seen as a means of imposing a particular conatus on a model.
According to this interpretation, we are faced with a thinking object.
* Grands Modèles de Langage : Pourquoi les réseaux neuronaux ont-ils réussi là où la linguistique échouait ?, by Claude Roux, March 8th 2024
PJ TV – IA : Pourquoi la linguistique a-t-elle échoué ? a conversation with Claude Roux, March 15th 2024
Portrait of a thinking object according to DALL·E