Eh?

There's a supply capacity crunch and obviously NVidia are going to heavily prioritize manufacture of the more profitable cards. That's basic capitalism for you.

But they don't half-arse CUDA on consumer cards.You get a different mix of functional units, but the code still runs fine. If you think a consumer card cuts it, they'll take tour money for one.

AMD is showing definite signs of competence with AI, and with its product portfolio, robotics.

Yeah but not very strong signs. They are still desparate to segment their nonexistent market share so you need to buy their punishingly expensive compute cards instead of the relatively cheap gamer ones just to see if it works.

NVidia don't give a shit as long as you are giving them money. If you want to run a 4090 in a datacentre, then go nuts it's on you. Your dime.

Also for some reason they are obsessed with NVidia compatibility rather than simply making the ML stuff run properly. Given the difference in architectures stuff optimized for NVidia is never going to be optimal for AMD. The problem with AMD is not that you can't run CUDA stuff out of the box it's that you can't do fucking anything out of the box.

If Europe and the US were doing that too we would be in a much better position regarding climate change. It would doubtless be a huge boon for those economies as well.

This goes to show the utter fecklessness of the right wing positions. If you ignore the environmental side (which is also just abject stupidity), it's obviously good economically, and for national security. These are things conservatives claim to be in favour of, though I've never seen any evidence of it.

Being insulated from externally imposed energy price shocks is good. Circulating money in the local economy is good. Heck if you're a true blue Tory, there's even British companies run by cronies to support.

I could go on...

I'll do it. Ones I've read below. Some of them are cracking. Some get annoying, but have really good bits (cough cough Neal Stephenson) for miss out the entire ending (Philip K. Dick y u). Some of them are completely bananas (City& the City, the Ninth series) but compelling. Some are just trippy (inverted world, anything by Dick), Some are a book I didn't even know I wanted until I read it (Spinning Silver). Some are great series (Vorkosigan series), there's good space opera, good mil-sci fi in there and on and on.

I won't say every one of these is the best book I've read, but some are really great,many are good and none I found bad (depending on your tolerance for particular author's style). Many of the older one are still really good.

A canticle for Leibowitz
The man in the high Castle
Cat's cradle
The inverted world
The Dying of the Light
Ringworld Engineers
Ender's Game
Speaker for the Dead
Falling Free
The Vor Game
Barrayar
Xenocide
Mirror dance
Memory
A Storm of Swords
A feast for Crows
Old Man's War
Rainbow's End
Halting state
Anathem
The City and the City
Cryoburn
A Dance with Dragons
3 Body Problem
Goblin Emperor
Ancillary Sword
The Fifth Season / Obelisk Gate / Stone Sky
Uprooted
Ancillary Mercy
Seveneves
The Collapsing Empire
Spinning Silver
Gideon the Ninth / Harrow the Ninth / Nona the Ninth
Project Hail Mary

He's made a few very good calls, and also has benefited greatly from his competitors being quite mind bogglingly incompetent.

Deep learning went big in 2012. NVidia called it right with cuda and then with CUDNN in 2014. But it's 20-fucking-26 and we still have this shitshow:

https://interconnectd.com/foru...

AMD and Intel are still tooling around dicks in hand trying to segment their nonexistent market share. Random people have got it working on a much wider range of intel and AMD GPUs using things like OpenCL and Vulkan shaders. And yet neither of these two multi-hundred-billion dollar companies has the wherewithal to make their shit just work.

The equivalent for NVidia? pip install torch
Compatibility matrix: anything released in the last decade.

Huang made a good call pushing GPGPU stuff and machine learning. He was given a rocket engine by intel and AMD for resolutely fucking it up.

Yes, that is what I feel,

Clearly.

What I described above is my lived experience.

Lies or delusion? Who can tell?

I'm the one who has been writing everything in rust for the last 6 years. Not you.

Yeah? I'm not pretending to only be able to code in one language. Only knowing one language isn't a brag. It's even worse when you obviously don't actually understand it well enough for this to be true.

and they pay me a hell of a lot more than you get paid

Also it get you lot of hot chicks who let you do the sex with them. Real ones, not anime.

Goodbye Aston Martin. Goodbye Jaguar. Goodbye Land over. Goodbye BMW. Goodbye Rolls Royce. Goodbye Bently.

BYD is going to come for your car industry,

Lol what car industry?

Jaguar: American since 1990, Indian since 2008
Land Rover: Joined Jaguar in 2000
BMW: that's not our car industry.
Rolls Royce: makes jet engines and nuclear reactors. divested motors in 1971, bought by BMW in 2003.

Apparently people in Luton are really unhappy with Labour because 120 years of car manufacturing finally ended in Luton last year under a Labour government. Naturally they voted Brexit in order to cut themselves off from the European market, but the fallout must be someone else's fault.

No you didn't forget, you just make shit up that your feelings tell you is correct.

It doesn't matter how hard you shit on C++, you still won't be able to understand Rust.

I've found that if you don't use weakly typed, anti-pattern riddled languages like C++ or python,

OK, troll.

Given the right prompts and direction, AI engineered software is faster and more bug free than what a human can provide at this point.

That's a marvellous little concoction: AI is better by construction. Any time it's not, it's the human's fault for not providing the right prompts and direction.

We have moved beyond hallucinating LLM's generating slop code. Linux kernel just fixed over 2000 verified bugs just in the past few months directly related to AI.

One of those doesn't follow from the other.

I've been setting up agents for scanning for security hole at work. It's pretty good, and as someone pointed out recently AI is making us pay back 40 years of technical debt. It's also quite good at finding security holes in its own code. But also it still has a tendency to produce slop and vastly overengineer solutions to simple problems.

OpenAI on november 4th wrote a story

OpenAI say lots of things. Jut because they say it doesn't mean it's definitely not true, but also doesn't mean it's true either. They will definitely round up "done by a team of security experts with million of token credits and AI tooling with unlimited compute" to "AI did it".

When used and validated correctly, pipelines like this can be used to add features, fix vulnerabilities, find new optimizations,

I didn't find it very good at optimizations, so far. It's... OK. Well, probably better than most programmers because most programmers are allergic to profilers and will hyperfocus on std::unordered_map because they know it's slow, rewrite it and assume their work is done. AI agents will use the tools and find the hotspots and will make something a bit faster. But they don't do a great job on the whole I've found. At some point it's better to stop the agent churning for hours and repeated re-prompts and simply write the faster version by hand.

I am going to blow your mind.

Unlikey.

Manmade global warming enthusiasts

You misspelled "people with a concrete connection with reality".

say that the solution to all our problems is emissions.

They do not.

Natural cycle enthusiasts

You misspelled "deniers of objective reality".

Does it do youtube?

https://www.youtube.com/watch?...

Though it's only really watchable during daylight hours in Florida. It's also not an aquarium.

Aw cute! The sheep thinks it's a wolf!

The actual wolves love having people like you around. You make their lives so much easier and they don't need to give you a penny for your efforts.

Even without the silly "$minimum wage to $million dollar" game, you can still advertise a very-low-end-of-industry-standard actual day-1 salary range with a clear statement along the lines of "opportunity for very large increase after 30 days" or something like that, and get the word out that you really mean it.

You'll never get the word out well enough and this way you'll basically never entice someone from a job that pays better than your very low end salary because they won't know.

Some friends of mine were doing this sort of thing (split development) years ago, pre AI. They had a desktop client in C++ and JS client for the web, which was slow. They compiled the core of the C++ one into ASM.js (this dates it) and it ran a lot faster. After a while they scrapped the pure JS one.