Artificial Intelligence thread

[henrik]

OpenAI's capabilities are in fact a contributing factor for US National Strength, the same way Huawei's capabilities are for China.

Open models might be where China is dominant now, but in enterprise coding, where the real revenue is, Claude and OpenAI are lapping every one up. And that matters.

Enterprise coding is just taking coding jobs from Indian contractors, and many other top US software companies. They are just destroying the stock valuation of those US companies, creating unemployment and lowering spending power for those affected.

 

[bsdnf]

This has had a significant impact on the Chinese academic community.

It's not just this particular incident; AI has had a major impact on some disciplines (such as mathematics), with many doctoral students abandoning their desire to continue academic research. For them, employment, especially entering the LLM industry, is almost the only high-income path.

Xi Yin simply further confirmed their judgment: top scientists are leaving academia, so I should leave too.

 

[bsdnf]

Minimax M3 Released

Highlights: Reproduces the ICLR 2025 award-winning paper; model runs autonomously for 12 hours.

Optimized CUDA operators; model runs autonomously for 24 hours, achieve 9.4x speedup

What did I say?

 

[PopularScience]

This has had a significant impact on the Chinese academic community.

It's not just this particular incident; AI has had a major impact on some disciplines (such as mathematics), with many doctoral students abandoning their desire to continue academic research. For them, employment, especially entering the LLM industry, is almost the only high-income path.

Xi Yin simply further confirmed their judgment: top scientists are leaving academia, so I should leave too.

Xi Yin is well-known dalai lama supporter. He won't go back China anyway.

 

[horse]

On coding, it’s clear Western frontier labs are still significantly ahead in enterprise engineering work where you’re dealing with huge code bases & simple iterative prompts (since people rarely write sophisticated technical specs to communicate with LLMs, much more often you’ll just see “fix this code for me it’s not working.”) This is exactly the setting this benchmark tries to capture.

But strong as they are, the costs of hyper scaling are also starting to pile up. True, they are “cost effective” relative to open models in situations where the user puts in minimal effort (since the open models will just repeatedly fail in those situations), but they still increase the cost of building software dramatically. The extra productivity they provide isn’t justified by the product value due to there being a finite market & only so much money to go around. Nobody is going to spend more on software services than rent, food, gas, etc.

So companies are ironically caught in a catch 22 with AI that is only really solvable via lay offs and scaling down the number of engineers. This is however a vicious cycle for capitalist economies where supply & demand are intertwined. The higher productivity from AI is not being absorbed by the market because consumers are being eliminated by the lay offs. Longer term there will be a ceiling at which the costs of using AI will exceed the market value & once we get there (and it’s probably soon), there will be opportunities for Chinese labs to catch up & overtake via value since it’ll become a race to the bottom on price.

I think we are already here on that part.

That balance point on that scale, where AI usage will met its equilibrium.

This AI equilibrium, that point where the value of using AI is maximized, it already here for some use cases I believe. However, this has not spread everywhere yet, into all corners of all industries.

What I firmly believe with AI, it will be different that other technological revolutions, in that it will speed things up.

So in some ways, we already at the end.

 

[horse]

Nope, I work with huge code bases and the idea of going back to using Claude is inconceivable at this point, it has no measurable capability difference from Kimi but you can use far less of it with the same budget. I have not had a single instance of Kimi failing to do something, nevermind "repeatdely fail".

And nobody actually work with the huge code base in entirety because that's just not how you do thing. It's also extremely expensive to load a huge codebase into context so you'd especially not want to do that if you're paying an arm and leg.

IMO western frontier labs are significantly behind in term of enterprise use, they just don't know it because the nature of enterprise is decisions aren't made by people doing the real work and not based on actual comparisons.

The catch 22 is only really for western model users, which is why Anthropic will be the death of western industry, as we're starting to see with companies either overspending or have to cut back usage.

Case in point, debugging at system level is one thing that you might want to load a lot of your codebase into context, but debugging also tend to be open ended and you can't predict how long you'll take. With Kimi I can just let it do its thing and not have to think about it, and I'll usually have multiple instances going at it while I do something else. With Claude you either risk using up your quota or you risk blowing your budget, which actually significant limit the amount of utility you can get out of it.

Reading your post, I thought of something.

I am not a coder, nor do I use AI. I used it a couple of times, to answer a question, that I did not want to research.

Here is my question.

Can we give AI a prompt, that tells it to code in assembler language?

Is that possible?

 

[horse]

I mean why do you think US companies are all starting to limit Claude use and removing token use targets? Because in real enterprise enviroments one-shot benchmarks means absolutely nothing while the productivity of using AI to do real work is entirely a function of how much you have to spend per token. Enterprise management are notorious for having no clue what engineers actually do, that's why they set those targets, now they know, but unless they discover Chinese models, they're going to be lapped by companies that uses Chinese models.

Okay, now I know what you mean by one-shot. Never heard of that before, but I am not a programmer.

But it confirms what I always thought of those coding benchmarks, that they were bullshit.

The code, is like an Excel spreadsheet. It changes. Over time, things do change. In fact, sometimes we find a bug we never found before!

Once DeepSeek came out, it was game over. Now we can see AI be deployed everywhere, for really low costs, and no one has to pay those few AI companies in the United States if we do not want to.

In the long, the US government will be the most important buyer of American AI. That means they are going to lose, big time!

 

[Halcyon66]

Reading your post, I thought of something.

I am not a coder, nor do I use AI. I used it a couple of times, to answer a question, that I did not want to research.

Here is my question.

Can we give AI a prompt, that tells it to code in assembler language?

Is that possible?

Yep, you can get an AI to write assembly code, but you’ve got to spell out the details — the CPU architecture, the calling convention, the syntax, all of it. If you don’t, it’ll just guess and the output ends up all over the place.

AI loves making stuff up

Regards,

 

[iewgnem]

Reading your post, I thought of something.

I am not a coder, nor do I use AI. I used it a couple of times, to answer a question, that I did not want to research.

Here is my question.

Can we give AI a prompt, that tells it to code in assembler language?

Is that possible?

Of course, theres no point since compiler already does that and AI can just use a compiler, but I did have one instance where it found optimation to an assembly snippet in our code. At snippet scale, which is basically the limit of modern assembly usage, its actually not that complicated, and I'd imagine the training dataset is huge since you have infinite amount of compiled and pre compiled code to compare. That and assembly is a pretty niche skillset so its not hard to do better than people.

 

[iewgnem]

https://twitter.com/i/web/status/2061453137500643384

An example of what I mean by, if you're using western models, you're just choosing if you want to die by going bankrupt paying for tokens, or die from getting lapped by people with access to Chinese AI.