Long life and prosperous distillation future to (not only) Chinese open weights models. Hopefully they will not stop releasing them ... anytime soon at least.

Having the code and the training data open would be much better but don't let the perfect be the enemy of the good.

Microsoft at least recognizes that memory safety features of Rust are a good thing and it is worth using (despite existence of unsafe keyword). If it would be according some slashdot users we would be still using assembler because real programmers do not make (stupid) mistakes.

Well, MTP will take a bit of memory lowering maximum inference context. But not much, probably only about 30k tokens. I'm not sure. I know that a friend runs Qwen3.8 27B with MTP and full context 262k on Radeon AI PRO R9700 32GB. He gets speed about 29 t/s. He used some of the Q6 quantizations. Not sure which one. He is likely using Q8 for KV chace, not sure. KV cache precision should be at least the same (preferably higher) than the weights precision. He also switched to Q5 model (at least for a while) because of some bug influencing inference speed (not because of memory size).

Many people consider even Q4 additional error rate acceptable. It is typically recommended as good compromise between model capability and memory required. Depends on the application. Higher compression seems to degrade reasoning from bigger contexts (e.g 100 kB of loaded source code) more than just answering a question directly. But Q8 seems to be as good as F16 almost always.

Q3_K_XL model with Q5_1 KV cache without MTP will give you about 110k token context on 16 GiB VGA., 21 t/s speed at low context (will drop to about 16 t/s when context is filled to about 60k tokens).

Yeah, 32 GiB is on the low side, but it is usable with small models. E.g. Qwen3.8-27B-Q6_K_M has KL divergence to F16 only 0.002 and will allow for full training context. It will do for some applications.

The subscription is a tax on stupidity :D
The more clever ones just install llama.cpp (or something similar).

You can get Intel or AMD VGA much cheaper (than Nvidia VGA with the same VRAM size) and run a local model. E.g. Radeon AI PRO R9700 32GB is about 1600; Intel Arc Pro B65 Creator 32GB is about 1200.
I'm sure the Radeon can run llama.cpp pretty well (a friend uses it (about 29 t/s, Qwen3.6 27B, vulkan backend)), the Intel one works too (though I do not know anybody personally who uses it).

Well chopping ones fingers is rather risky. But other methods you pointed out were good. I would also add:

Look at a second indicator on a clock. Check whether it progresses well.

Try to look closer at something to see more and more details (are the details expected, are they stable in time).

Try to influence the environment just by wanting it to change. If it is a dream you likely can do it. E.g. add or remove an object from a scene just by willing it.

Check whether physics laws hold in general. E.g. jump up and will yourself to fly, if you can fly then it is a dream.

The last two points are the biggest fun. Although my problem is that the more I bend a dream to my will the more quickly I wake up. The fun level seems to be inversely proportional to the amount of time the fun lasts :-(

If it is an LCD (i.e. not CRT, plasma or OLED) then no. You may get a temporary image retention but it will go away after a while of showing something else.

I clicked the story just to find out what does it mean. I hoped for a post explaining it but no luck.
Google indicates it likely stands for Good Old Games.

My experience with the posted hiking estimates is that they vary a lot. Sometimes they are shorter, sometimes longer than my times. ±30% time difference happens sometimes.
As for as cycling, 17 km/h in mostly flat terrain seems appropriate for a non-competitive cyclist. I'm doing about 23 km/h average (full suspension recumbent bicycle on flat asphalt) but I'm going for a leisure ride every day. I do not do any competitions nor do I care about them.

I just ordered 50 pairs of bicycle brake pads from them. Hopefully they will survive the fine :)

I have a good experience with AliExpress. Though I order only simple items which cannot be easily screwed up and which are rather easy to check for proper functionality. And only if the result price is at least 33% lower than what I can get locally. I spend a few thousands with them and only one item I received was defective: some bang-bang temperature regulator. I kind of expected that. I ordered 3 pieces and one was bad. I needed two working regulators so it turned out well.

I would never order from them anything which is hard to check for safety ... like e.g. cosmetics. Bicycle brake pads are my favorite item. They are not distinguishable from the ones I can get locally and AliExpress price is 1/3 of the local one.

... or makes the battery to catch fire. With a bit of bad luck also a nearby house will burn down.

Only if your ECU supports it. Manual transmission can be useful in snow though.

LLMs likely will not be the future AI but maybe JEPA will be. We will see.

It can make sense. If China's government believes that surveillance stabilizes government then they have an incentive to prevent surveillance in USA so that USA's government can be more easily destabilized. Less surveillance will also allow easier insertion of China's spies into USA.

But I think that the big data center investment is primarily driven by hopes of investors that the AI is going to have huge enough (non-surveillance) commercial impact to justify the investments. The surveillance part of it is negligible. But AI will be used for surveillance as well of course.