Q6T46nT668w6i3m

I love MLIR. The modularity and friendly abstractions make it incredibly flexible. I've now used it to write _multiple_ domain-specific optimizations and transformations for some of my recent research! It truly bridges the gap between different devices (CPUs, GPUs, TPUs, etc.). I pray more people adopt it so it doesn't end up abandoned!

Truly! RIP brave Kerbins.

The Times summary is solid:

> SpaceX’s Starship Rocket Explodes After Launch

> The most powerful rocket ever built got off the launchpad in South Texas, but did not achieve its most ambitious goals on Thursday.

Sure, but, nevertheless, I’d like to know why it exploded before declaring this a success!

Only?

You’re misremembering? Everyone wanted a web-browsing smartphone that didn’t skimp on performance. There was insane demand and Apple delivered.

I wrote this part of the code. The features are computed on the server.

I love Dick Cavett too.

Terry Gross is, of course, fantastic and there’s nearly fifty years of Fresh Air interviews. I think we sometimes ignore Fresh Air because it’s so ubiquitous. Elvis Mitchell is another obvious example.

I think what separates them from, for example, Lex Friedman, is taste.

NumPy doesn’t provide automatic differentiation or heterogenous compute.

I love JAX … but JAX, frustratingly, doesn’t obey NumPy promotion semantics! Enforcing double precision is understandable from a performance perspective but it breaks NumPy compatibility and frequently introduces subtle and difficult to debug numerical bugs. Nevertheless, JAX has encouraged so much progress in ML software engineering, e.g., I’ve been cribbing from diffrax and torchopt for my PyTorch work.

I recently watched a Tesla drive straight into a barrier at 10-15 MPH. It was bizarre. The driver was fine but the car was totaled. Driver said he was using autopilot and wasn’t paying attention.

Meta?

LLMs are capable, at best, of very limited forms of reasoning. The absence of abstract latents precludes multiple inferences and the optimized action to perform a task.

Yes. This is an addition to that property. And, just to be clear, it isn’t Stanford’s.

If you’re one of the few foolish enough to climb Denali in the winter, you can find some comfort in the countless stashes from the summer.

So … Unreal? ;)

“Integrity” is a great description of Richard Posner’s writing but I’d, politely, claim “vehemently” is a bad description! Posner, to his credit, frequently refutes his past writings! A trademark of his writing is his weariness about immutable philosophy.

If you believe that restorative justice is an effective way to reduce recidivism, it’s essential that a court act as a meeting place between a defendant and their community. For this reason, you want a defendant’s legal representation to be a peer so there is a shared understanding of the particularities of their community.

I don’t know whether you’re being coy but this is rumored to be his strategy.

It depends what you mean by “open source.” You can find countless open source implementations. The issue is training. Meta has released their LLM weights. I believe Google has too.

DeepMind has an excellent track record of helping other researchers reproduce their work. OpenAI does not.

They stopped.

It’s great and everyone should use it ;)

(I’m a co-author)

My issue isn’t the scale itself, my issue is the scale used for _this_ data.

I don’t know exactly what kind of reply you are expecting to your second sentence.

Log transforming this data isn’t far off.

Is this an attempt at trolling statisticians?

Isn’t this textbook yellow journalism?

I think you misunderstand the situation. The transponder transmits the data. This isn’t anything like attaching a GPS device to someone’s car.

You think we should enshrine property (e.g., aircraft) with a right to privacy?