catgary

I think you should just focus on the road because most of us are just trying to get home safely to our families. Some of us are even biking beside the road on a lightly-protected bike lane.

Yeah, because they aren’t in need of a social safety net.

I think you need to give some concrete examples, considering the US happily let its companies offshore a lot of work to China over the years, and Chinese funds own large chunks of American companies.

I’ve found they’re quite good when you’re higher in the compiler stack, where it’s essentially a game of translating MLIR dialects.

Because it doesn’t? Thats why the field of formal methods exists.

Because the code is the actual thing, tests can only show that the code fails in certain cases, they don’t actually prove the code is correct.

Generally speaking, I try to ensure that the LLM is using core abstractions throughout the codebase in a consistent manner. This makes it easier for me to review any changes it makes.

And even then - I still read the code it generates, and if I see a better way of doing something I just step in, write a partial solution, and then sketch out how the complete solution should work.

This line of thought is honestly a bit silly - uv is just a package manager that actually does its job for resolving dependencies. You’re talking about a completely orthogonal problem.

I think there are 5-7 thousand confirmed deaths by the UN, and medical reports in Iran estimated there could be 20,000+ casualties.

I see the value of the students, it just seems like an odd thing for a government to subsidize via NIH/NSF funding. We don’t really have anything analogous to that in Canada and it just seems awfully weird that it exists in the US without the “it’s older than the country” excuse that Oxford/Cambridge have.

EVs should do much better on brake dust thanks to regenerative braking, no?

Luckily, the LLM has the text to cite, it can be passed in at inference time, which is legally distinct from training on the data.

Because text analysis is substantially easier than video analysis?

Sure, in the sense that any belief about the law isn’t cut and dried until a judge has explicitly dismissed it in the court of law.

Are you implying that an LLM needs to be trained on a specific piece of text to answer questions about it?

You don’t need any rights to execute the feature. The user owns the book. The app lets the user feed the book into an LLM, as is absolutely their right, and asks questions.

I work on a much easier problem (physics-based character animation) after spending a few years in motion planning, and I haven’t really seen anything to suggest that the problem is going to be solved any time soon by collecting more data.

I think this is an interesting direction, but I think that step 2 of this would be to formulate some conjectures about the geometry of other LLMs, or testable hypotheses about how information flows wrt character counting. Even checking some intermediate training weights of Haiku would be interesting, so they’d still be working off of the same architecture.

The biology metaphor they make is interesting, because I think a biologist would be the first to tell you that you need more than one datapoint.

Did he get an actual type theorist for that part of the project?

These companies are also biased towards solutions that will more-or-less trap you in a heavily agent-based workflow.

I’m surprised/disappointed that I haven’t seen any papers out of the programming languages community about how to integrate agentic coding with compilers/type system features/etc. They really need to step up, otherwise there’s going to be a lot of unnecessary CO2 produced by tools like this.

Most of these C-suite executives have more than enough money to retire comfortably.

It’s kind of a bummer because this is the exact same playbook as DirectX, which ended up being a giant headache for the games industry, and now everyone is falling for it again.

You would need to identify specific far left views that would be comparably objectionable. I’m not going to be upset if someone has wonky ideas about free market systems or tax codes. DHH has said some objectively racist rhetoric on his blog and called Tommy Robinson’s recent march “heartwarming”.

Yeah, I think I’ve seen a statistic that getting hit by a car going 30km/hr, you’re probably going to be fine. If you get hit by a car going 50km/hr, you are most likely going to die. I can’t think of any reason to allow cars to travel over 30km/hour in urbanized areas, outside of designated arterial roads.

I agree with you that writing kernels isn’t necessarily the most important thing for most ML devs. I think an MLIR-first workflow with robust support for the StableHLO and LinAlg dialects is the best path forward for ML/array programming, so on one hand I do applaud what Mojo is doing.

But I’m much more interested in how MLIR opens the door to “JAX in <x>”. I think Julia is moving in that direction with Reactant.jl, and I think there’s a Rust project doing something similar (I think burn.dev may be using ONNX has an even higher-level IR). In my ideal world, I would be able to write an ML model and training loop in some highly verified language and call it from Python/Rust for training.

Maybe! I’ve found that people usually don’t do extra work if they don’t need to. The heavy notation in differential geometry, for example, can be awfully helpful when you’re actually trying to do Lagrangian mechanics on a Riemannian manifold. And superfluous bits of a definition might be kept around because going from the minimal definition to the one that is actually useful in practice can sometimes be non-trivial, so you’ll just keep the “superfluous” definition in your head.

I’m going to push back on this a bit. I think a simpler explanation (or at least one that doesn’t involve projecting one’s own insecurities onto the authors) is that the people who write these papers are generally comfortable enough with mathematics that they don’t believe anything has been obfuscated. ML is a mathematical science and many people in ML were trained as physicists or mathematicians (I’m one of them). People write things this way because it makes symbolic manipulations easier and you can keep the full expression in your head; what you’re proposing would actually make it significantly harder to verify results in papers.

I think it’s the move towards GPU-based computing is probably more significant - the constraints put in place by GPU programming (no branching, try not to update tensors in place, etc) sync up with the constraints put in place by differentiable programming.

Once people had a sufficiently compelling reason to write differentiable code, the frameworks around differentiable programming (theano, tensorflow, torch, JAX) picked up a lot of steam.

From what I can tell, NPUs are mostly being used by Microsoft to encourage vendor lock-in to the MicrosoftML/ONNX platform (similar to their DirectX playbook).