Tade0

Imagine my confusion when people started throwing around the officially accepted, in my native language, abbreviation for artificial intelligence.

Depends how it's done I guess. France has a decent population tower, for lack of a better expression:

https://www.populationpyramid.net/france/2026/

If they keep it up, their population will remain stable up to the 22nd century. There will be no further noticeable ageing and thus no threat to anyone's retirement.

Reminds me of Stanisław Lem's Terminus:

https://en.wikipedia.org/wiki/Terminus_(short_story)

Who wrote all this? Not humans, that's for sure. But the style is that of human writing.

I'm getting:

  uBlock Origin has prevented the following page from loading:

  https://tech.criteo.com/blog/human-skills-ai-cant-develop-junior-engineers/

  This happened because of the following filter:

  ||criteo.com^

Umm, okay? The entire domain is blacklisted?

LLMs "think" with tokens - it's the only unit of state they know, so it's really more of an accident of having larger models with more context.

This problem is older and existed already before the pandemic, but instead of Polish plates it was German/French, as those countries didn't have such expensive insurance and it was easy to have a foreign branch of a company register one's company car.

The first crackdown was also before the pandemic and being Polish and in-and-out of Italy I fully expected to get caught in the crossfire, but that never happened.

Apparently I was driving too much like a tourist - indicators and all.

That being said a few people from the local Polish community got into trouble, but it always first due to driving into a Zona a Traffico Limitato with a car that shouldn't be there - plates notwithstanding.

For me fun left the room around 2016, but I can relate.

Now it's just work. There's not a ton of pride in it.

Yeah, sadly true. But it did stop it in the Netherlands.

Well, yeah. Earthquakes bad, don't cause them.

> Does a system need the knowledge of the full works of Shakespeare to be able to output Javascript?

Turns out that actually - no. Researchers have managed to prune half the Experts in a MoE model that had a low probability of getting activated during coding tasks, resulting in a more focused model:

https://arxiv.org/abs/2607.16721

Main benefit is that it greatly reduces the amount of RAM required to run these models. Of course you could just cache those unused experts on disk instead, but the main point here is that you know which ones matter.

But aside from that recent models, like Qwen3.8-27b are reportedly more durable under heavy quantisation, e.g. 3bits or even ternary. With additional techniques like TurboQuant, you can feasibly run these models on consumer hardware - even if at 1/4th the speed you'd get from rented infrastructure.

VS Code has extensions such as Kilo Code or llama-vscode which let you work with local models much like you would with cloud based solutions.

I dare say review is often way more important than the act of laying down code.

Everyone is stealing from everyone else.

Bielik largely exists because, and this is also my own experience, even frontier models still make grammar/spelling mistakes in Polish - even more so now than in the past.

> we can't compete on energy and labour costs for training,

Specialists are expensive everywhere. China is functionally 80s Japan surrounded by several Brasils and all the frontier AI work is being done in that first part, where labour is expensive - if only due to competition for top talent.

It's no secret that everyone is dis-stealing from everyone else.

Nah, just overprovision like crazy and store what you can.

Germany has currently around 1.5kW of solar installed per person - even in October this yields double-digit GWs at 9am. They can't put all that power to use in summer, but it sure puts a damper on prices off season.

Sure, storage is still being built as there's a market for it, but for now it's still cheaper to just build more solar capacity.

Difference is these speakers never output the 80dB+ those modified exhausts do. Also, they don't produce infrasounds, which to me personally are actually the worst aspect of all of this.

There was an interesting study showing that while social mobility is generally greater in Europe than in the US, the European tendency to underestimate vs American to overestimate it, results in broad agreement that social mobility is in fact greater in the US.

Poking holes in every new idea is a pan-European pastime - it's how people in this region of the world managed to survive despite exhausting its resources.

To add to the other comment, there's also Ridge quantisation - the majority of weights are indeed Q3_x, but the most sensitive layers are FP8.

Yeah, but some prefer this if it means they own the infrastructure.

I spent some years in the pharma industry, where AI came really late, as there was (justified) concern that sensitive data might leak - even by accident. Eventually a solution was implemented - it was some kind of open model (they didn't say which) running on-premises.

You couldn't tell these people that this and that powerful model is free or inexpensive, because it's useless to them if it's running on someone else's computer i.e. the cloud.

In my area half the issue is cars being parked on the way, which is why the truck has to approach the bins in reverse, which is particularly loud.

The other half is buildings being spaced 7m apart, because that's the minimum allowable at this height. The echo is enormous.

That being said it's still more tolerable than the people absolutely punching it through our 30km/h street with their modified exhausts and windows rolled down - as if the point for them was to lose hearing as fast as possible.

Sounds like Switzerland - different bin for each colour of glass - at least in Zürich.

Define "small".

The other day I managed to get a context of 195k for Qwen3.5-9b Q4_K_M using a llama.cpp fork that supports TurboQuant:

https://github.com/TheTom/llama-cpp-turboquant

I think you could replicate this with a larger model on your device.

Overall with the right quantisations for both the model and KV cache you can get a lot of mileage out of this old hardware. Speed remains the main limitation, as IIRC I was getting ~26-30tps on a 7700S.

My concern is that there might come a time when this cost is passed onto employees.

Suppose everyone starts moving faster thanks to LLMs and it becomes an expectation to use them. Budgets aren't infinite, so one of the two has to happen:

1. People get laid off.

2. Costs are shifted onto employees - either through lower salaries or having them bring their own subscriptions. I don't even make $500 a day!

I dare not ask about the cost, having burned $60 on a task running for 1h 16min once.

That is a useful hint, thanks!

I make Abathur rephrase multi-line comments, because every LLM out there enjoys peppering code with them.

A successful recconect resets the counter. Efficient.

Doesn't always catch that distinctive style, but at least all the fluff is gone.

This Zerg is a software engineer at heart and has no tolerance for nonsense.

> Strangely, there were many times where ChatGPT decided it would answer our medical question, but not in the form of a TV medical script. The tides had completely turned!

I've noticed this happens a lot with characters that don't have any long lines and/or have limited vocabulary.

Furby trails off pretty quickly, Abathur from StarCraft 2 (great for shortening code comments BTW) starts talking like a typical LLM after a few sentences.

Common issue with this game. The red bomb in mine had been stepped on, at which point it became too conspicuous.

I'm afraid internally they would still look like chromium/blink anyway.

I was wondering what kind of client name they were talking about.

It's really silly that this was the only means of blocking access. I expected them to have, I don't know, some sort of cryptographic signature or something.

An additional JWT with a long expiry time would even work here, anything.