From the paper: ... we employ a decoder-only language model GPT- 2 (Radford et al., 2019) with a configuration of 4 layers, 32 hidden dimensions, and 4 attention heads. https://arxiv.org/pdf/2508.011...

Nvidia is charging what the market will bear, which leaves plenty of room for slower but cheaper AI hardware. Intel could be (part of) that competition, but not with this attitude.

If you entrust your data to a company, they may be forced to disclose that data to someone they would rather not disclose it to. One example is a court case: when a claim is made, the court has to examine the facts (the data) to decide the claim, and as part of the process (called discovery), the opposing claimant's lawyers will get access to the data. They're not supposed to use that data for anything unrelated to the case, though. It's not supposed to be unreasonable: obviously there's no way to decide a case without access to the facts, and so discovery is intended to ensure that the court has that access. There's nothing special about AI here, this holds in general. In the case of AI, it's a good idea (if you're concerned about privacy) to bring the model to the data rather than the data to the model, by using an open-weights model on your own machines, rather than sending your data to some site in a cloud somewhere controlled by another company, where it may be subject to something like this if e.g. the other company ends up in court.

Has anyone considered the possibility that Amazon actually wants people to work in the office for real actual working reasons, for example so that they can communicate with each other in person rather than only online? Amazon management is allowed to try to create the sort of working environment they think is the best for the company: that's the job of management. It seems quite plausible to me that Amazon sees people quitting as a result of return-to-office as undesirable attrition, and their main motivation is the actual benefits of in-person working. I don't see why it's necessary to conclude that the quitting is the goal.

Bezos addresses this Blue Origin meeting in his article. Worth a read. My impression is that Bezos, as he says, is worried about the lack of trust that many people (especially on the right) have of mainstream media, and is trying to shift the Post into being something that can speak to more than one political side. I'm not sure how possible it is to transform the Post into something that could do this. I can't think of very many examples of major news media that hasn't picked a political side, and the Post most certainly has, for many years now. It will take more than this non-endorsement decision to transform the Post in the way Bezos says he has in mind. A shift to the political center will be costly, as the Post will hemorrhage readers who like where it is now on the political spectrum. In fact, this decision itself has started the hemorrhage.

I don't find it hard to believe that standing always at your desk might not be good for you. Standing sometimes, however, is unlikely to be a problem. I use a Sit/Stand desk at work and while I usually use it sitting, I occasionally use it standing. I find standing is better when I am using the whiteboard, so I can move back and forth without switching from standing to sitting and vice versa.

ARM requires translation to run x86 software (especially games) so it isn't going to work well in all cases. Qualcomm's latest snapdragon X ARM cpu offers better performance per watt than most x86 offerings at the moment, so that pays off in laptops as better battery life. A laptop with fantastic battery life that can run most software but not a lot of games? Sounds great for some people. Not so fine for others. Just know what you're buying and it'll be fine. It's not as if you can't buy x86 laptops anymore if you want excellent software compatibility at the expense of a bit of battery life. My recent Thinkpad T14 has a fast eight-core AMD CPU and surprisingly good graphics. The battery life is fine, as good as I need, though not quite as good as these new Snapdragon laptops. I don't see the snapdragon option as a problem, just as a choice that doesn't work for me but maybe works for others. If x86 laptops go away entirely, that'll be a different story, but I suspect that won't happen unless ARM gets quite a bit better than it is now.

I rarely regret not having mod points but I wish I could mod this up.

The trackpoint, the little red thing in the middle of the keyboard, is the reason that all my laptops over the years have been thinkpads. I still have and use a Thinkpad 600X made in 2000, and have had many thinkpads since. Various people say that they don't use it and why shouldn't Lenovo get rid of it? Well, because some people (like me) use it all the time and love it. I even bought lenovo USB keyboards that have trackpoints: I am typing this on one right now. The few years when my main machine was an iMac, I used a lenovo USB keyboard with trackpoint on my Mac.Yes, I realize not everyone loves the trackpoint. That's why modern thinkpads have touchpads too, and they're good touchpads, I'm told, if that's what you like. I like the trackpoint and don't want to ever have to give it up.

Mainframes are useful, expensive, powerful I/O machines. They were doing visualization properly before it was a thing anywhere else. They are superb for running old, working code for many decades. IBM has been busy adding new features so keep their existing customers, and it seems to be working: most of their existing customers are probably not going to get rid of them, despite the ongoing expense. The real question is whether IBM can attract new customers. What is the value proposition to get a mainframe for a customer who does not already have one? Unfortunately this story doesn't really answer the question. If IBM wants to preserve their mainframe business in the long term, this question will need a good answer.

Model collapse when using AI-synthesized data is a thing, but model collapse is not necessarily inevitable. Work is being done on how to prevent it. Here's something current: https://x.com/KempeLab/status/...

I see some claims here that an AI model learning from data is considered differently in law than a human being learning from the same data. It seems plausible to me that law might make that distinction but I would love to see some references. Does anyone have any to share?

CUDA was more mature than OpenCL in 2009 when we started using GPUs. By 2012, when Alex and Ilya wrote AlexNet, we had a fair bit of experience in writing in CUDA and so CUDA is what AlexNet used. While nobody uses Alexnet anymore, pytorch and tensorflow (which succeeded AlexNet) also work best in CUDA. This has made Nvidia very rich. Not that this gets us much of a discount buying Nvidia GPUs, unfortunately.

This is a restatement of a fairly typical argument against short-term thinking. The issue isn't "good enough" per se, it is good enough for what? Restating this argument, it's that a course of action that is sufficient for a short term goal but doesn't move one sufficiently towards a long-term goal isn't going to get you to that long-term goal. This is a truism of course, but still a useful reminder: don't be satisfied with short term goals alone, consider and pursue your long-term goals too.

You don't need as much expertise to rent GPUs in the cloud as it takes to run them yourself. But you need a lot of money to do it. There are a lot of people with money but not onprem computing expertise chasing AI right now, so there is plenty of market at the moment for GPU-renters. I'm not sure if building an alternate GPU-focused rental service is a viable long-term business though, because it's unclear how to create a business that way that can compete on anything other than cost. Once AWS, Azure, or GCP drops their price (which they will do when necessary to keep their resources busy), these new vendors will be squeezed.