catgary

9 ms·

121.
by catgary 2y ago
It was infuriating joining a robotics company with a ROS based stack. A lot of the original engineers had started at the company in undergrad as interns, and I swear it was impossible to make anything better because they didn’t know what be
122.
by catgary 2y ago
The weights are like jpegs, the model itself is still a piece of software that executes code.
123.
by catgary 2y ago
Speaking as someone who deploys ML models on edge devices - having the MLIR and being able to make binaries for different platforms is terrifically useful! You’re often supporting a range of devices (e.g. video game consoles, phones, embedd
124.
by catgary 2y ago
I didn’t mention tensorflow or TF Lite? IREE is a different technology, it is essentially a compiler/VM for MLIR. Since JAX exports MLIR, you can run JAX on any platform supported by IREE (which is basically any platform from embedded
125.
by catgary 2y ago
I don’t think that’s a terribly fair description of Google - AWS’s chips (inferon and trainium) both have robust XLA/JAX support. Plus JAX now exports MLIR, so there is a really compelling JAX -> IREE pipeline so JAX models can more
126.
by catgary 2y ago
But electric bike usage has demonstrably lowered car usage in provinces/states that have subsidized them. They make 5-15 km commutes super achievable.
127.
by catgary 2y ago
That’s spot on. It makes that insane compensation package he tried to set up for himself look like a pump and dump scheme.
128.
by catgary 2y ago
This something I’ve thought about in the past, since I messed around with trying to add monads to JAX - I think you made the right call with effect handlers. You might want to take a look at what Koka does, that was the best implementation
129.
by catgary 2y ago
I’ve only been reading through the docs for a few moments, but I’m pleasantly surprised to find they the authors are using effect handlers to handle effectful computations in ML models. I was in the process of translating a model from torch
130.
by catgary 2y ago
Yeah, I’m currently an ML scientist but I regularly run into shit where I have no clue how you’d solve the problem without having a compilers/functional programming background.
131.
by catgary 3y ago
I’d probably say Karras has written the best papers on how to think about diffusion models.
132.
by catgary 3y ago
Have you looked at burn.dev? It’s kind of a neat approach where they backend to torch/candle/other things and are also building a pure Rust backend that compiles to webgpu shaders.
133.
by catgary 3y ago
Ehh, there’s a lot that goes into it. Just because a physics engine is differentiable doesn’t mean that its gradients going to be useful. For example, if you look at Brax/Mujoco in JAX, the gradients generated by Mujoco are absolute g
134.
by catgary 3y ago
There’s been loads of work in this area (differentiable simulators) in JAX. I think Google/Deepmind are porting mujoco to JAX ( https://mujoco.readthedocs.io/en/stable/mjx.html ), and for fluid dynamics or defo
135.
by catgary 3y ago
Like dentists who are against fluoride in water, I guess.
136.
by catgary 3y ago
He was reminded every day as his home canton still had high rates of gout and children being born deaf/mute.
137.
by catgary 3y ago
That was a fascinating read - there’s even a great villain in Eugene Bircher (not to get into politics, but he definitely seems to have trailblazer the “right wing populist attacks successful public health measures” strategy).
138.
by catgary 3y ago
We really have replaced “oh no, you put too much knowledge into a LISP program and it’s sentient!” with “oh no, you’ve stacked too many transformers and now it’s sentient!”
139.
by catgary 3y ago
I know Koka is more of a research project than anything else, but I think it’s by far the most interesting. Moving to effect handlers, the Perseus ARC algorithm, and identifying “functional-but-in-place” algorithms all feel like game change
140.
by catgary 3y ago
Yeah, maybe I’m a gatekeeper, but he just seems to fundamentally lack the credentials I would expect for the CTO of a self-driving car company. I know, startup culture, hackers, blah blah blah. But seriously - had he ever worked in the auto
141.
by catgary 3y ago
I would say that at that point you’re probably just digging in your heels - that’s a major group that does serious work. You can also look at Aaron Ames’s PhD thesis and how that work in hybrid dynamical systems, driven by category theory,
142.
by catgary 3y ago
If they are being paid a generous pension, I can’t see why it’s unreasonable we expect them to simply be retired.
143.
by catgary 3y ago
Taking ownership of Amazon seems crazy, but I do wonder if there’s some logic to the idea that it might be time to consider “public utility” versions of apps. Like a crown-corporation cloud services provider or a municipality-operated DoorD
144.
by catgary 3y ago
It’s basically just saying notions of compositionality from software doesn’t really translate over to analogue computers, and by extension robots that interact with their environment. It seems like they’re being careful to make atomic state
145.
by catgary 3y ago
There is an ‘unroll’ parameter in scan that lets you control how many iterations of the loop are fused into a single kernel.
146.
by catgary 3y ago
So, just to offer a few counterpoints there is Tom Leinster’s work on diversity/entropy [1], where people working in mathematical ecology have been happy with his contributions. I’ve also see the Penn robotics group start to use catego
147.
by catgary 3y ago
It’s a bit tough, on one hand John Baez is a great messenger because he’s such a fantastic communicator. But on the other hand, he’s a pretty horrible spokesperson for ACT because the man did theoretical physics and higher category theory f
148.
by catgary 3y ago
It’s a section/retract, definitely not an isomorphism. You’ll need to carry around the index data, and any sort of array update will incur a crazy upfront cost.
149.
by catgary 3y ago
Yeah I imagine an array write that changes the type of an index would be insanely expensive…
150.
by catgary 3y ago
Honestly, the roller bottle approach seems pretty reasonable when they’re making sure that the process is making good chicken. They’ve at least reduced the problem to making giant sheets of chicken-paper.

More ›