I prefer to listen to people who actually put AI/ML models into production, instead of random programmer / AI dilettante types.

You'll only hear one side of the coin. I would also want to hear from people who understand AI/ML and have decided not to use it in an application domain. That's probably the most valuable info, knowing when not to use a tool.

I understand ML reasonably well, and have decided to use classical AI instead, because I want real-time performance on low-end hardware (and I also want to be able to predict and extend the operating parameters). It's a lot harder, though, because I'll have to understand the problems at a fundamental (mathematical) level better than I currently do, and all the time I can hear a little voice saying “a neural network could do this in half an hour of work and a week of training”.

But this is part of the author's point: where are all the production reinforcement learning models?

As someone who's just now diving into low-level programming on unix, this article is just what I needed.

Saving this for future reference.

DeepMind, Hinton, et al are still convinced that the brain must be doing something like backprop.

See Lillicrap address all common objections to backprop in the brain: https://www.youtube.com/watch?v=vbvl0k-aUiE&ab_channel=ELSCV...

Also from their paper Backpropagation in the brain:

"It is not clear in detail what role feedback connections play in cortical computations, so we cannot say that the cortex employs backprop-like learning. However, if feedback connections modulate spiking, and spiking determines the adaptation of synapse strengths, the information carried by the feedback connections must clearly influence learning!"

Notice he won't commit to any quantifiable short-term predictions that could come back to bite him in the future. No references to any ML benchmarks at all. Just a lot of hand waving and vague arguments.

A has been that has nothing to contribute towards ML or AI. Ignore him.

FIFY

Operations ^H^H^H Functions A common function for applications is logging a user in. It’s actually two sub-functions composed together: first get the email address and password from the user, second load the “user” model from the database and check whether the password matches.

Functions are the doers of the MFVE world. They are responsible for making changes to your models, for showing the right views at the right time, and for responding to events triggered by user interactions. In a well factored application, each sub-function can be run independently of its parent; which is why in the diagram events flow upwards, and changes are pushed downwards.

What’s exciting about using functions in this way is that your entire application can itself be treated as an function that starts when the program boots. It spawns as many sub-functions as it needs, where each concurrently existing sub-function is run in parallel, and exits the program when they are all complete.

"What’s exciting about using functions in this way is that your entire application can itself be treated as an function that starts when the program boots."

HN, rediscovering main()