setsubyou

Size is not the only limitation imposed by physics. And the biggest reason we want to make chips small in the first place is electrical signal travel time, one of the limitations that has been around forever. Another way we’be been approaching it is parallelization, and there is no limit imposed by physics in sight there. Another limit we’re not even close to is the optoelectronic switching limit, which is 1 petahertz.

And average users having, at most, 8GB of RAM

Average users when I started using computers had 64k and were happy. They will buy average computers but average computers become more powerful.

mid-spec computers from 2013 are almost perfectly serviceable today

Yeah, why not. I have an old (not quite 2013 though) Atom netbook that’s serviceable. As long as nobody installs Windows on it at least.

Because current fabrication process of chips seems to be very close to limits imposed by the laws of physics.

So? That’s pretty much always been the case. And then someone figures something out. That something doesn’t have to be related to fabrication at all. Just a few years ago people were in awe at what Apple managed to do with simply a better design.

All of this talk is under the premise that AI works and will keep getting better. Surely if that’s the case, at least the AI will come up with something. You say it won’t, but that really just means it’s not the super intelligence the AI companies are chasing and their whole plan failed.

Why let the user have the model when you can rent it off to them?

Because most companies are not AI companies. If the AI is just a part of your product that makes users more likely to but it, you don’t need to rent it out.

It’s not like there’s much of a choice anyway as long as there are labs that are just publishing their weights. There’s already a trend now that AI useing companies are cutting costs by dropping online models in favor of self hosted open weights ones.

the average person isn’t going to use a locally hosted model, ever.

True.

I don’t think that’s true at all.

Platform vendors have already started shipping small LLMs for translation and for AI assistance. If you use Apple Intelligence you’re using local models to some extent for the smaller tasks. Consumer hardware is perfectly capable of running much larger models too if you give it enough RAM, and is only held back by RAM being expensive. My five year old laptop can run 30B-ish LLMs, and if RAM prices hadn’t exploded, then by now surely those 64 GB wouldn’t be an outlier and newer devices would be running more powerful models for “average people” too.

But prices for RAM will eventually come down again and hardware in general will keep becoming more capable. It’s weird to me to suggest anything else, considering that we went from 8 bit PCs with 64k of RAM to modern smartphones in my lifetime. Why would hardware stop improving exactly now?

The same applies to LLMs themselves. Qwen 3.8-27B beats expensive online models from two years ago in its stronger areas (e.g. coding). Why would that development suddenly stop now? There’s a hard limit here somewhere regarding the compression of knowledge (just based on information theory, not specific to AI) - which is why smaller models hallucinate more - but that doesn’t rule out e.g. improvements to how external knowledge can be integrated at runtime.

Accordingly, the AI models shipped with consumer devices will keep becoming more powerful and online models will be needed less and less for “average person” tasks.

One part of the AI bubble is the assumption that AI will remain compute intensive, which is what drives the data center buildout. But historically, new technologies have often started out as expensive and barely obtainable for average people and then became cheap and ubiquitous within fairly short time frames. It’s weird to me to suggest that LLMs will be an exception especially while also suggesting, at the same time, that the technology will accelerate research and development across the board.

I guess it depends on what you think prompts look like. If it’s “Claude, write GTA, make no mistakes” then yeah that’s probably never going to be programming. It’s also not what prompting looks like that produces these results that make people like the author see a gap. Although personally I think a big part of the gap is the automatic verification coding agents can do and that’s impossible in many other areas.

As someone who barely (i.e. not professionally yet) wrote assembly almost 40 years ago, structured languages were well accepted by that time, but compilers were legitimately not that good at optimization, so much like today, we did have to do things by hand instead of using a compiler if we wanted good results. That compilers got better at this than humans in most cases is a somewhat recent thing. Personally I’m too young to have ever thought of structured languages as “not real programming”, but I know people who definitely did think that. The other thing that wasn’t considered real programming is scripting languages. The perception of those changed dramatically in my lifetime.

One argument I see often is that even though a modern structured language’s compiler does generate a ton of code from a small amount of input, at least you know exactly what that input does. Which is funny to me because for the last 20 years I’ve been complaining that people don’t understand what the machine does when it runs their code does anymore, because how are they supposed to learn that when they started on Java or similar.

It’s a weight class in China, anyone can tell they’re 49 tons because that’s what these trucks are marketed as.

In the US the heaviest class is ~36 tons, in much of Europe it’s 40 tons, and 49 is not that much more. But in Scandinavia you can find much heavier ones, up to 76 tons in Finland.

They have different fixed lengths, no prefix needed to tell them apart.

If you can afford taking a few years off in the first place, you can also afford starting a company and doing your own projects for a while. Don’t leave the field behind, put some software on app stores and pick your topics by what skills you want to learn or maintain. Maybe it even makes money.

I don’t think it’s really less elegant in Python or Swift. Especially not in Swift where you can write it very concisely in some cases, like numbers.reduce(0, +). Admittedly Python is a bit more verbose.

But maybe Python and Swift programmers are not used to it in the same way as Clojure programmers are.

It’s a target for capacity added by this particular initiative (REPowerEU), but member states have barely allocated any of the funding available under it. So REPowerEU is not effective, even though capacity has been added in other ways (e.g. just the market doing its thing).

One thing I feel like I’m noticing is Claude being super eager to read my own changes between turns even when they are in a completely different area. It’s like they trained it to suck in all confirmed human generated content before doing actual work. It’ll be like “The user’s changes are in some text files and don’t touch anything I’m working on. Let me read all of it anyway before I continue just to make sure”

There are some unanswered legal questions around this. In the US there is fair use, so US companies argue that training falls under that. If they’re right it’s difficult to prohibit because fair use literally means they don’t need a license in the first place. If they’re wrong and they’re found to be making derivative works, then there is likely no need as the terms of the GPL would then apply to their models too and they would not survive that. Other countries have different situations.

Any license prohibiting use for AI training would not be an OSI approvable open source license as OSI prohibits discrimination by field of endeavor.

The comments on HN mention a watch made by waveshare, I didn’t know they made that. It’s basically an esp32 dev board in watch shape. Similar price point too.

I don’t know about dumb. First off, she lives in Switzerland with her wife and children, so she can mess with German politics as much as she wants and it won’t affect her private life that much. Might have to finally start paying taxes there though if she has to flee Germany.

Secondly, she lived in China for years, speaks Chinese fluently and probably is more popular on Chinese social media than in Germany. So there is another option why she might be interested in breaking Germany other than just stupidity.