the Machine Learning Compute Protocol that unites all of the world’s compute into a global supercluster, accessible by anyone at any time

Why the world needs a Machine Learning Compute Protocol (i.e., @gensyn) 🧵

Investing in Decentralized Social at Collider Ventures. Hosting @thepodcaster. 

Machine Learning is an inevitable tech, essentially because of how effective it is in multiple areas. The internet as we know it today will broadly be replaced by machine learning models.

In the future, apps and websites will replace imperative code execution with probabilistic ML models, and we will get to a point where we will have intelligent agents acting on our behalf. This will obviously significantly increase the demand for compute power.

So, why do we want to decentralize this space? What do we get out of this?

3 things - (i) lower prices, (ii) increased scale, (iii) auditability.

(i) Lower prices – If we create an open market for ML compute (i.e., enable anyone to provide CPU/GPUs for the training/inference of these models), the increase in supply will lower costs and solve the huge logistics burden we currently face.

That radically changes the sort of ML models we can create. Your compute power is no longer limited to Google data centers but potentially includes every phone and compute-having device.

(iii) Auditability – Open sourcing ML models makes it much easier to understand why someone is building something and what they are trying to achieve with a specific model.

This means that a potentially dangerous model could be blocked at the application level (rather than much earlier, before the model is even created, as current regulations attempt to do).

Takeaways from an interview that @@benfielding from @gensyn recently gave. Listen to the full conversation here

Takeaways from an interview that @@benfielding from @gensyn recently gave. Listen to the full conversation here https://www.youtube.com/watch?v=g4jPsXMs7Gw

(ii) Increased scale – You essentially remove the scaling limit. If we can only run training/inference operations on data centers, we face huge constraints in terms of the compute power available.