tlb

When paper data is breached, the crooks don't steal the paper and put in their own locked file cabinet. They take pictures of the documents and sell the data on the dark web. So the end result is the same.

I expect companies busy self-flagellating with prophecies of doom to fail, and companies selling the best models for the lowest price to win. But individual companies booming and busting won't disrupt the overall progress of the technology very much.

"<Q>What model are you?<A>Qwen." is surely in Qwen's training data. It's quite standard to include such meta knowledge during instruction tuning.

I learned to machine on their machines, before graduating to full-sized units. They were great for learning, because they had all the essential features of professional manual machines, but couldn't cause as much damage when something went wrong. And they were delightfully made and easy to take apart and modify. Sad.

That's true for production models, but a lot of research involves working with fine-tuned models made for one experiment. RL involves constantly updating weights. Those may well run in the same cluster as the eval.

It's shocking that anyone can spoof the number of, say, my bank. At least large institutions should have authentication.

Please don't submit pages with multiple unrelated stories, as it's not suitable for discussion.

You can engage with an AI-written document! Simply paste the document into the AI (same or different as the AI that wrote it) and ask questions. You can get an answer in seconds compared to an hour (if it's a blog post by an HN reader), days (if it's a professional blogger), or never (if it's by a journalist).

Perhaps AI-written articles should include the prompt as hidden text on the page. Or include a link with the prompt embedded, that causes the AI to spit out the same article that you could then ask questions about.

Librarians have systems that can classify every published book. And scientific publishers can classify every paper. Those don't add up to a complete classification of all human knowledge, but it's a substantial fraction of it.

If you skim the high-level book categories at https://en.wikipedia.org/wiki/List_of_Dewey_Decimal_classes, you may conclude this is is beyond a lifetime's work.

Gmail is at least as large a target, and they don’t keep having breaches.

I just ask LLMs questions about things I'm curious about. I don't really want to follow someone else's predefined curriculum. A curriculum is necessary when the information has to be delivered in a lecture, but not in a 1-1 interactive setting.

If you type "I want to learn CUDA programming. Where should I start?" into Claude, it gives a pretty plausible outline. It's not clear what an app could add to it.

Some of mine do. They use WASM64 with threads to run robotics simulations.

This is out of date. There are now a number of fusion energy startups, including Helion which has raised $1B from investors, mostly Silicon Valley ones you've heard of. It's not yet proven that the tech will work, but it is proven that investors are willing to fund it.

They do as soon as that company makes a profit, or anyone sells shares in that company.

Yes, if you run `git blame -C`.

15 years ago you could argue that venture capital wasn't funding enough advanced tech, so ideas were failing to cross the gap from pure research to commercial development. But lately there's capital available for quantum computers, fusion, synthetic bio, space exploration, asteroid mining, and lots more. The government is going to suck at funding the right things. They should leave tech transfer to private investors, and focus on funding pure science.

git actually does this. `git diff --find-copies`

0 for 10 on some startups (large and small, YC and not) that came to mind.

It's easy to scrape YC startups from https://www.ycombinator.com/companies. Scrape that and a dozen other investors' portfolio pages and you'll have a useful fraction of startups.

In terms of political power, programmers are more like yeomen than peasants. Yeomen have power due to their essential skills because if a good fraction of them stopped working the system would collapse. (Whereas if peasants stop working, they just starve.) If a sufficient number of programmers said "We're not going to implement or support privacy-invading systems" the government would have to back down.

England also has a lot of disused airfields, often with huge hangers and stupendous concrete runways built during WWII. A few are open as museums. They can be worth a quick visit.

As a kid I found saltpeter at an old-fashioned pharmacy and made gunpowder, and it also barely fizzled. I think you have to grind the ingredients much finer than a kid has patience for.

Generator excitation is not the hard part of a black start. You have to run coal feeders, blowers, and water pumps for an hour before you can spin the generator. Then you get power instantly upon applying power to the field windings.

Yes, that is basically the plan. It's based on the belief that unfettered AI would let anyone be a supervillain and destroy the world. There are enough would-be supervillains out there, but they rarely get far because they can't get teams of smart people to build doomsday machines for them. So the AI has to not let anyone do evil with it.

Unfortunately, that won't feel very much like freedom.

It's ridiculous to consider MITM attacks out of scope for taking over your computer. Also, there are probably ways to exploit this without a true MITM like DNS cache poisoning. But it's best to just assume the whole internet is MITMed.

They should compile it to WASM and host it.

Building a swap device at user level used to be one of those classic unsolvable problems, because what if your daemon needs to swap in a page in order to swap in a page? Or at least it was discussed at a reason why microkernels will never work. I’m not sure what the solution is here.

There's a package to do this in Julia: https://juliaarrays.github.io/StructArrays.jl/stable/

There are still things you can't do with an open-weight model without the training data, like modifying the architecture and training from scratch. That's different from true open-source code, where you can do anything the authors could do.

Linux on Arm works great. I barely notice the difference except everything is a bit faster. Most SaaS companies can and should switch.

Yes, for example taking off or landing a rocket on the surface blasts particles of sand out sideways at 1000s of m/s. The particles can fly in the thin atmosphere for kilometers and sandblast everything. Our intuitions about how far and fast tiny things can fly are only true in an atmosphere of similar density.