Lukasz Stafiniak
lukstafi.bsky.social
Lukasz Stafiniak
@lukstafi.bsky.social
Currently I'm an independent wannabe researcher working on Machine Learning and AI in OCaml.
7. Consciousness and the Social Brain by Michael Graziano (2013)
8. The Centered Mind: What the Science of Working Memory Shows Us About the Nature of Human Thought by Peter Carruthers (2015)
9. Human and Animal Minds: The Consciousness Questions Laid to Rest by Peter Carruthers (2020)
November 24, 2024 at 1:39 PM
4. The Architecture of the Mind: Massive Modularity and the Flexibility of Thought by Peter Carruthers (2006)
5. How Can the Human Mind Occur in the Physical Universe? by John R. Anderson (2007)
6. The Conscious Brain: How Attention Engenders Experience by Jesse J. Prinz (2012)
November 24, 2024 at 1:38 PM
(BTW, "lowerer", while a bit awkward, is a good name for "converts into a low-level representation".)
November 23, 2024 at 12:13 PM
Thanks, that makes sense! I carried into this my definition of scheduling as ordering and/or synchronizing the computation of kernels, or materialized buffers, based on a topo-sort of inputs-outputs.
November 23, 2024 at 12:00 PM
But it might well be worth it! I'm warming up to things as I get more familiar with tinygrad. Maybe the steeper learning curve pays off and my original comment was unwarranted.
November 23, 2024 at 11:53 AM
(2) Before PR 7065, engine/schedule.py is preparing indexing aka. movement ops, is integrating indexing with computation. Why is the file called "schedule" when it doesn't do scheduling? I probably mis-generalized from this, to the feeling that it's surprising where things happen in tinygrad.
November 23, 2024 at 11:51 AM
(1) Axis representation combines user-facing buffer-specific semantics (index into a tensor), with operation-specific semantics (whether axis is reduced), with tunable computation details (global mem vs. local mem, whether computation is upcasted, when the reduction happens: local vs late).
November 23, 2024 at 11:32 AM
A steeper learning curve because often a single representation interacts with various stages of processing. Two examples below.
November 23, 2024 at 11:30 AM
So how to square this with Tesla's published statistics: FSD users have half the accidents of non FSD users? Until a few months ago -- certainly during the reporting here -- FSD was only granted to people with good safety score. That would cut inferring that FSD is good, but still a defeater.
November 22, 2024 at 11:55 AM
Mind mapping: one page of pithy notes organized into a graph.

Coding music sprint: put on music, start a timer, completed with a good line of code or an idea that moves me forward.
November 22, 2024 at 10:03 AM
My current daily stack:
* Exercise
* Mind mapping
* Coding music sprint
* Chores or a nap
* Work outside for a while (a park or a cafe)
* Meditation
* Stretching
November 22, 2024 at 10:01 AM