The profzi scheme is a famous comic from Jorge Cham. It compares academia to a Ponzi scheme: many students work for a professor to produce more papers. The professor takes credit for all the paper, but few of the advisee will have a great academic career because... the whole thing is pyramidal... the numbers don't add up. The whole operation is largely sterile and only works due to the pretence that producing peer reviewed research papers is equivalent (synonymous) to producing new, valuable, knowledge. What happens if AI can produce the papers? The scheme is exposed. We see it in mathematics right now. The AIs sidestep the entire peer-review process... and provide correct proofs to what was considered by the math establishment as the important problems. These people are very angry. It is the great AI debunking. So what is the solution? Go back to what should always have been. The purpose of research should never have been to produce peer-reviewed papers. This was always a perversion. (For people who do not read me regularly, the peer-review model became dominant in the 1970, it is a recent addition.) The purpose, of course, is to solve problems that people care about. That's much more difficult than producing papers... so it puts the bar much higher for scientists and researchers. But the bar should never have been lowered in the first place. What happens to the 'profzi scheme'? I predict, I hope, that it will fall. Not all of a sudden, but once the system come to the realization that the production of papers is a solved problem, AI can do it better than us... then it won't be decades before students start worrying about where the train is headed. (The comic attached is AI generated. Obviously.)

Most people who do machine learning with Python use PyTorch. Amazon has its own hardware for machine learning: trainium. On AWS, trainium instances are roughly half the price of Nvidia instances... although you get somewhat less compute. Overall, AWS claims that trainium is 30% cheaper than Nvidia. Amazon just made it much easier to run PyTorch on its own chips. TorchNeuron is a native PyTorch backend. Existing PyTorch code can move to Trainium by changing the device string: model = MyModel().to('neuron') Trainium chips are available on AWS EC2. This is still in beta. Code at github.com/aws-neuron/torch-…

What fraction of your code is typed by hand (no AI)

Intel is making AI accelerator cards. Intel’s GPU D270 (Crescent Island) is a PCIe 5.0 card built for inference. It is more or less a variant of a graphics card, engineered for AI. It uses 32 Xe cores. Each one mixes ordinary vector engines with specialized matrix engines called XMX. Memory is LPDDR5X. Intel’s card has about 160 GB. It can go up to 480 GB. I am curious about the price tag.

The Financial Times reports maybe only a third of all large AI infrastructure projects are realistic. Issues include grid connections, transformer shortages, liquid-cooling bottlenecks, and local opposition. If so, this would make Space Data Centres even more attractive. The next few years will be interesting.

What do you do if someone offends you? Adult human beings tend to be pretty good at the right answer: do not react. At least at first. For a tribe or a civilization, this non-reaction is obviously useful. Fighting is expensive for the group and best be avoided. But it is also often the best strategy for the individual. For three main reasons: 1. We are frequently wrong about what happened. Sometimes people are just careless. Or there is a misunderstanding that can be resolved. It happens often enough. 2. Provocations often see to extract status. Someone who offends might want to establish their dominance. By ignoring them, we are fighting back. It takes power to be able to ignore someone. If you fight back, you signal that you might be backed into a corner. You have to fight back because the other people do have some power over you. Not reacting as expected is a sign of power, your own power. 3. Your attention and energy are finite and scarce. Spending them on an offense is a tax that you are paying. The absence of reaction is not the same as passivity or weakness. Because if you refuse to react, you still get to play. Yet you play your own game, on your own terms. To win. piped.video/watch?v=qX-KZfoZ…

We often convert binary data to base64 (crypto, images). The Python standard library is ridiculously slow. Thankfully, you can just use the fastpybase64 module and get crazy speeds... See my article for details.

The next Redis release should have native support for Roaring bitmaps. I ran some benchmarks. In some instances, this could drastically improve your performance. This could similarly be added to Valkey soon if people want, there is an issue and code proposal. Check my article for more details. (Such as 'what are Roaring bitmaps?')

Been using Starlink and honestly it's been great for fast, reliable internet at home and on the go. If you sign up through my link, you get $100 in cash or service credit: starlink.com/?referral=RC-DF…

Competition ! Meta just came out with a competitive coding agent, Muse Wonder Full. You can subscribe for either 50$/month or 200$/month. For coding, 'Muse Wonder Full' seems to be a match for the best models from Anthropic and others.

Thanks to everyone who gave me ideas as what to talk about regarding AI and programming in my upcoming lecture. I will try to give credit.

With Python 3.14, you can use actual threads in a single Python process... which makes it possible to parse JSON (for example) at insane speeds. See article for details.

I am giving a talk on AI and programming in a couple of weeks. I have an abstract but the talk is not ready yet. What should I talk about? The audience will be made of young people.