orbifold

they have exactly two customers, both of whom are also investors.

More precisely they can't cache it.

Same can also be said of IBMs computers at the time and the US efforts to build computers.

Step 1: steal all of the worlds IP and make a judge rule it is fair use

Step 2: declare end of the Anthropocene (defined by man made influence on earth), while simultaneously accelerating mass extinction

Step 3: ???

Step 4: Profit?

They are increasingly being trained on generated tasks and even (parts) of the pre-training data is 'distilled' (e.g. Clibmix as an open-source example), so there are many ways in which the vocabulary can seep into the model.

I mean EU can cope all it wants, the most successful startup in Germany being a weapons company is all that you really need to know. It is sad how far these countries have fallen and I am German myself.

If SAP appears in a list of successes then clearly EU is fucked.

Maintining this type of long time fork is the kind of mind numbing, tedious thing LLMs are actually exceptionally good at. I did some backend work on Lean4 compiler and then they switched out parts of the codegen, it took Codex less than an hour to recover the feature set I had implemented but on the new backend.

All this AI doomerism is hiding the obvious and immediate harm that will cause a mass extinction event. Funnily enough people at Anthropic are predicting the end of the Anthropocene, completely absolving themselves from any responsibility for the actual reality that the Anthropocene is defined by human made geologically noticeable changes. Like causing a global climate catastrophe.

Article reads AI generated, or maybe those are the kind of publications is where they got this from?

"What it means and what it does not": article is clearly written by claude, which for me at this point means that I pay as much attention to it as I do to Claudes output: I skim it and then move on. It actually requires far more effort of me to read it than it was to generate the content. Before it would have always been the other way around. Whoever wrote the article would have had to spend way more time writing it than me reading it.

can you give examples of colo facilities in SF / New York that are shutting down?

I was thinking that would be great, too. What would be the equivalent for the property developer: one gpu server is 450k.

The hardware division has 80%+ margin and still makes the systems that process 75% of all financial transactions. Their processors for those systems are on par or better than any other, I don’t think that is a business at all. This cash cow is not going away any time soon and gives them the profits to make bets on the future of computing.

I wish that people would not use AI to generate copy or websites, at this point it induces an almost instinctual negative reaction: 'What Modelloop does -- in three movements' is a classical example of AI lingo. I don't have proof but I feel that the most recent models have gotten worse. The rhetorical devices they use are tiring and off-putting at this point.

I learned programming in QBasic in MS-DOS, you could just start the IDE and the documentation was filled with cool examples. Super easy to run any program. I made music / weird drawings etc.

Here is how I think about it: Learning to program is learning a new way of thinking. When you learned to do mental arithmetic the point was not that you would necessarily do mental arithmetic at all times in the future. Programming is the last step when solving a problem with a computer, learning to program teaches you how to solve problems more generally.

I recommend reading a book like https://mitp-content-server.mit.edu/books/content/sectbyfn/b..., going through it will hopefully as enjoyable as it was for me when I read it in high school. There are many kinds of programming which are not super enjoyable (to me), so I gladly leave those to AI, but based on personal observation, my experience programming lets me be much more effective at using AI to solve problems than a fresh MIT / Oxford grad with less programming experience.

Finally it depends on your interests: If your interests are computers and X, than combining both to solve problems you find interesting can make using AI worthwhile, because then programming isn't the main point.

Writeup by MIT news: https://news.mit.edu/2026/new-study-bridges-classical-and-qu....

is there a jj hosting service?

A lot of that is compute commitments and NVIDIA has a 80% margin

they were definitely totalitarian, slightly different mix of ideology. Fascist is a fairly good description here, it describes close collaboration of government with corporations to advance national goals. US had somewhat fascist tendencies for a long time now.

I don’t get that, the use of these books was instrumental and necessary for the success of the training run. The expected value of these training runs is high as the build out of 100 billion+ infrastructure demonstrates, so the book publishers should at a minimum be paid a licensing fee, a small fraction of every inference run revenue or whatever they decide. The fact that authors and publishers didn’t get any say under what conditions their intellectual property can be used is pretty outrageous.

In fact 5g and all previous standards have a provision for lawful intercept. So your domestic intelligence service and police can always turn it into a listening device.

I guess we will find out in 20+ years once the next electron positron collider at CERN has been build

Curious how this relates to what lean4 is doing, I guess in lean's case some of the data structures are special cased (Array) and there is no easy way to implement such data structures yourself

At some point we will be so tired of distinguishing between AI generated content and human content that we will stop using the Internet and it will be left to bots.

It took cerebras less than a billion to get to where they are now, CPUs are not that hard. You would probably be able to reverse engineer them for ~100 million

I mean in a normal math curriculum you would define only the multiplicative inverse and then there is a separate way to define fraction, if you start out with certain rings. It is kind of surprising to me that they did a lazy definition of division.

One other thing I've observed is that Claude fares much better in a well engineered pre-existing codebase. It adopts to most of the style and has plenty of "positive" examples to follow. It also benefits from the existing test infrastructure. It will still tend to go in infinite loops or introduce bugs and then oscillate between them, but I've found it to be scarily efficient at implement medium sized features in complicated codebases.

200 is a lot of monkeys