GuB-42

Distillation is not illegal by every definition of the word.

I am waiting for a precedent on this one. In general, training on copyrighted material is legal, there is a lot of precedent there. But every now and then there is a case where the owner of the training material wins.

I don't remember the details but I believe one of these instances was when one company trained its AI on the knowledge base of another company and turned it into a competing product. Fair use was denied because of that direct competition. Distilling a LLM to make a competing LLM looks kind of like this, or maybe not, I don't know.

It would make sense for distillation to be legal in every way, LLMs are built on a broad interpretation of fair use, but sometimes, law is weird.

But you still need backups right?

For most people "the cloud" is where you backup stuff. If you have a personal backup strategy that doesn't involve the cloud, you are not "most people".

Doing backups the right way take some skill and investment if you want to do it by yourself. It may involve setting up a NAS, and some discipline with physical media. You have to do your own security too. Most people don't want to do that, so, cloud backup it is.

It is not a perfect solution, even beyond the privacy considerations, like you can still lose your data by losing your account, but from my personal experience, people lose their data less often now that they have cloud backups.

Maybe it is out of habit, but I never managed to get into PowerShell, in fact, I am not at easy with the Microsoft way of doing things, with few exceptions. Too much UNIX I guess. Nushell seems to be based on the PowerShell philosophy of using structured data and not text, not my thing.

I really like the UNIX way of using text I/O, it has it flaws but it works for me. But that being said, I still hate bash and all its family. It has so many footguns it is an entire armory at this point, mostly related to spaces and escaping.

Something Perl-like could be a saner replacement. It is already a bit shell-like, it doesn't struggle with escaping the way bash does, and it has very powerful text processing abilities that go well with traditional UNIX tools.

It is quite representative of an embedded software project. TigerBeetle is not what we usually call embedded software, but it is built like it: static memory allocation, a self-contained executable, and a strong focus on determinism are typical in that field, especially for critical software.

And I think embedded software is a field where Zig will be at its best. The only thing it is missing is maturity. When project lifetimes are measured in decades and changing a single byte can cost millions, no one in his right mind will pick a language that is still in development. Things will become interesting when it reaches 1.0.

People will just buy diesel cars then. That happened in Europe, people bought diesel cars because it was cheaper, because of taxes. Not great for the environment.

Make $200 per gallon of diesel? This has the potential to collapse entire economies. Shipping will be more expensive, which will in turn make everything more expensive. Food will be more expensive too, tractors use diesel. Few things don't depend on diesel at some point or another.

Tax only the fuel used by cars? Many people still need their car to work, not every job can be remote, at $200 a gallon, you may as well quit, unless you get paid the difference, which again will increase the price of things.

Maybe a few decades from now, if could be $200 a gallon without breaking everything, but certainly not now.

It is easy to start working on it when you are rich, but people in the world aren't, and they aren't because we are and we exploit them. And these people don't care about how terrible the next centuries are going to be because they need to survive right now, so climate change is not a priority.

It is kind of depressing, really. We can't even agree on not killing each other. I mean, no one wants to get killed right, and the solution to this is the simplest there is, we just need to collectively decide not to kill. And if we did that, just consider how much resources it would free up: military, security, no need to rebuild what war has destroyed if there is no war.

So if we can't simply decide not to kill each other even though it would be a net, immediate benefit, how deciding to care for climate, which will involve at least a global reduction in comfort for many years before things get better is going to work?

More recent AI tends to be less agreeable, sycophancy is widely recognized as a problem and AI companies are working to fix it. And sure enough, some people don't like it, as shown when OpenAI shut down GPT-4o.

But for the people who do actual work with AI, who are the people who pay a lot of money for it, they prefer to be called wrong when they are, because it is better to be called wrong by an AI than being wrong in front of a customer.

It is not an easy problem to fix, because as much as we want an AI to give us the right answer rather than just being agreeable, we still need them to follow orders, and AI that doesn't would be quite useless. So there is a balance to be found. Smarter models tend to do better because they know what is right, compared to smaller models that don't and then assume you are right and hallucinate from there.

Why all the mess with Bun?

Couldn't they have rewritten Claude Code in Rust directly? No more need for a JS runtime, better performance, etc... If their agents can do Zig to Rust, why not JS to Rust?

The reason the adds don't count is not because they arbitrarily decided to ignore them.

Addition is O(n), multiplication is more than O(n), by the nature of the big-O notation, when you have to do a series of operation, you only have to count the ones with the highest complexity. So in the Karatsuba example where the formula involves both additions and (recursive) multiplications, the additions don't count only because the multiplications dominate.

Or, as a formula, O(n log n) + O(n) = O(n log n), and btw O(n/2*10) = O(n), in big-O, constant factors don't count

Reflecting and radiating heat are mutually exclusive, and if some material is good at radiating heat, it is also good at absorbing heat, otherwise it would violate the laws of thermodynamics.

It is a bit more complicated than that because light come in different wavelength, but it is the general idea.

Railroad rails tend to be dark except for the polished top surface, so they absorb/radiate the heat more than they reflect it. Painting them white changes that, it makes them less effective at radiating their own heat, but because they don't produce much heat by themselves, it is not a big effect, but it also makes them less effective at absorbing the heat from the sun, a much greater effect, so, the rails get cooler overall.

TigerBeetle is interesting because even though something like TigerStyle is unusual for a database, it is very much at home for embedded systems.

In fact we could summarize TigerBeetle as "the financial database developed like embedded software". Determinism is a cornerstone of critical software like in avionics, and naturally, dynamic allocation is forbidden, as in TigerStyle.

So even though it is not obvious, TigerBeetle represents one of Zig biggest strengths: embedded systems.

This. I talked about Zig to my company CTO for embedded systems, he told me he was happy that someone mentioned it as he finds it promising and not much talked about, especially compared to Rust.

He then told me that it is, as you said, unfinished, and that he can't reasonably recommend it for embedded projects that are often long term and demand stability. C, a 1970 language is still king for a reason! So he is keeping Zig on his watchlist for when it becomes more mature.

I don't know of any useful model that would match the usual definition of "open source". That is, where everything needed to build the thing is provided, such as the training data and code. Not that it would be tremendously useful anyways considering how expensive training a model is.

I much prefer the "open weights" term. It is not open source in the sense that you only get the finished product, not the actual source, but it is still open in the sense that it is not only accessible as a service.

For an analogy, take Quake for instance. When it was launched, its game server was available as an executable, so you could run it your machine, but that didn't make it open source. Only much later it was released as true open source software.

I like Centre Pompidou, and if it wasn't technically problematic, it could have aged really well.

For those who don't know, Centre Pompidou is built "inside out", with all the technical parts like ducts, framing, elevators, etc... fully visible from the outside. There is some beauty from function here, like looking at an old locomotive. For people in the future, it would be an insight into what a 20th century building is made of.

Problem is: for a form-from-function design it is not very functional. Because the technical parts are also part of the aesthetics, it makes maintenance problematic. Every single pipe you change has to be of the same model or you would ruin the façade, which is crazy. Unfortunately (for me), it is one of the least likely building to stand the test of time, from a technical perspective.

Other than that, the Louvres pyramid is starting to gain some acceptance, even though it was almost universally hated when it was built, who knows how it will be seen a few hundred years from now, if it still stands.

LLMs don't have feelings, but they imitate humans who do. If your prompt sounds rude to a human, LLMs might respond like a human who is talked to rudely, if you don't want that, prompt politely.

Chatbots go through a reinforcement learning phase, which should be enough to make the LLM helpful regardless of the tone of the prompt, but you may still get better result if your prompt is made in a way that would get the most helpful response from a human.

It is not anthropomorphizing, it is helping the LLM recognize a pattern.

It is not.

You can make sand by crushing rock, we don't because it is cheaper to extract natural sand. Should we run out of suitable natural sand, we will just have to go with the more expensive option, not a big deal.

All structures are worth preserving. The article is about a latrine. Do you think ancient Romans thought a literal shithole was worth preserving? And yet, it is now an important piece of history.

As for beauty, you never know, take the Eiffel tower for instance, when it was built, people found it ugly, and it was to be a temporary structure. Now it is the symbol of Paris, more popular than the Paris Pantheon, which much better fits the classical standards of beauty.

The Eiffel tower is also more historically significant, more representative of its time. The Paris Pantheon, is an imitation of the Roman style, beautiful, but probably not as interesting for future historians. If both building survive that is, which won't happen without continuous maintenance, at least for the Eiffel tower.

Maybe we can make an analogy with taking the plane vs driving to your destination.

A plane literally goes 10x faster than a car, so a 5 hour drive becomes a 30 minutes flight. But if you have to drive to the airport, arrive early, pass check-in, security, boarding, then pick up your luggage on arrival, rent a car and drive to your destination, you may realize that your 30 minute flight took you more than 5 hours in total.

You also have to consider the Amdahl's law: a 10x speedup on 10% of the project is just a 9% speedup overall.

It sounds like a pretty good deal, so I assume you got it before prices went crazy, like mid-2025.

At that time, you could get a decent entry-level Windows laptop for around 700 euros that can do pretty much everything you want a laptop for if you are not a power user. More than that and for most people, it is "expensive". At 950 euros, you are paying for options, and it turns out that the package offered by that M4 Air is what you want, but it is likely that a gamer would want a different package for these 950 euros. Personally, I went for a 1000 euro ThinkPad because it was a better fit for my needs at that budget. In particular, I wanted more than the 256 GB of storage you most likely had (1 TB).

My parents (both teachers) always told me to respect books, and it means keeping them in good shape, among other things. It is almost a religious idea: books represent knowledge, something that is important and must be preserved.

To be honest, it is not like we were maniacs about that, books are meant to be read, and accidents happen, but damaging a book if you could avoid it just felt wrong.

And I am also thinking about libraries that have books that are hundreds of years old or more. The reason we still have them is that people cared.

It is not entirely rational, but it is one of these things, like playing with food, a moral consideration more than a practical one.

AI changed the perception of the "information wants to be free" idea.

In a sense, AI companies did a lot to "free" the information, they took everything they could, including pirated data, and put them all into a model, which you can then query to get something similar to "not free" information, but clean of copyright.

But now that information is actually free (or at least, freer than before) people realize that it didn't come out of nowhere. People worked to produce that content, and many of them are people like you and me, not billionaires and faceless corporations, and it is affecting them and their ability to produce more content.

That part didn't change, what changed was that before, piracy was a rebellious act, done by poor teenagers, something easy to sympathize with. Now, it is done by trillion (!) dollar companies on an industrial scale, not much sympathy to give here.

The first reason is price. Macs are expensive, you may argue that it is worth it but it is still a lot of money. The Macbook Neo change the deal somewhat, really good value, but even with that, it is not the cheapest laptop by far.

The second reason is compatibility. Apple does a pretty good job with x86 emulation, but it still a problem for many Windows games, Linux support is not the best either.

And then, there may be plenty of specific reasons, maybe you are missing some connectivity, maybe you don't like Apple, maybe you want a Nvidia GPU, etc...

These names are for humans to remember and to type-in, not just for computers. And whimsy names are more memorable, which is a good thing. Random strings are for computers, not humans.

I was also surprised, but it makes sense.

It is based on user-generated content, it has a comment system, followers and stars (similar to "likes"). There is even an attention economy, and some gamification.

Social media doesn't have to be bad, it just needs to be social.

I didn't notice that.

Some software got buggier. Other software improved. I didn't really notice a pattern in the recent months. Microsoft is doing quite bad, and they make a lot of software many people use on a day to day basis, so maybe there is that.

Often, I get this sentiment when my hardware gets old. Software progressively bloats, but with new, powerful hardware, there is enough margin not to notice it, but eventually, you will, and then they will stop active support. At that point, you may blame the latest release, but it is more of a continuous process. And if, by coincidence, some heavily used software is becoming troublesome for other reasons, it definitely feels like everything goes wrong.

And then we find something to blame. Today, it is AI slop, and it certainly the cause of some software becoming shitty, but I didn't notice a global trend yet. Maybe there is, but we would need quite a study to be sure, I don't even know where to start, so much potential for bias.

I trust the judges on that one, they know the law better than I do, obviously.

I am also not discussing about who is at fault, I agree that it is on OpenAI.

I just don't want, say, some company that is even shadier than OpenAI to launch an OpenAI branded protect with the intention to mislead people.

Maybe grant OpenAI the trademark, but do not allow them to use it on products that are not actually open, but I guess it is legally problematic.

I think you are confusing trademarks with patents.

Trademarks don't prevent you from copying anything, they only prevent you from being misleading regarding the origin.

As much as I hate OpenAI for hijacking the term "open", and I love the idea of OpenAI losing, I am not sure if I agree with it.

Trademarks are first intended to protect consumers, so that if it says Coca Cola, then the Coca Cola company made it, for the better of for the worse, but at least you know.

OpenAI is already a well known name in Europe, and when I see OpenAI on a product, I expect it to be a product of that company. It doesn't mean I will want to use it, I may even want to avoid it, but I don't want it to be from someone else. By denying that trademark, anyone could call their product OpenAI, and I don't think that situation would benefit the consumer.

When I had to fill up the ESTA form to enter the US, this is what was in the "social media" drop down list:

- ASKfm

- Facebook

- Flickr

- GitHub

- Google+

- Instagram

- JustPaste.it

- LinkedIn

- Tumblr

- Twitter

- Vine

- VKontakte (VK)

- YouTube

This is obviously outdated but it can give an idea.

I'd say farming. Not agriculture in general, Monsanto is one of the most hated companies, so is John Deere, and farmers hate them too, but the ones working on the field, preferably crops over livestock.

I think that farmers are somehow immune despite having their fair share of evil because we all understand that without them, we would all die.