• 51 Posts
  • 2.29K Comments
Joined 1 year ago
cake
Cake day: October 23rd, 2024

help-circle



  • For my language, J, I can’t get autocomplete.

    Even though J is a functional language (on extreme end), it also supports fortran/verbose python style, which LLMs will write. I don’t have the problem of understanding the code it generates, and it provides useful boilerplate, with perhaps too many intermediate variables, but with the advantage that it tends to be more readable.

    Instead of code complete, I get to use the generation to copy and paste into shorter performant tacit code. What is bad, is that the models lose all understanding of the code transformation, and don’t understand J’s threading model. The changes I make means it loses all reasoning ability about the code, and to refactor anything later. Excessive comments helps, including using comments as places to fix/generate code sections.

    So, I get the warning about “code you don’t understand” (but that can still happen later with code you write), and comment system helps. The other thing he got wrong is “prompt complexity/coaxing”. It is actually never necessary to add “You are a senior software…”. Doing so only changes the explanation level for any modern model, and opencode type tools don’t or separate off the explanation section.

    LLM’s still have extreme flaws, but article didn’t resonate on the big ones, for me.











  • Wanting to sell to China just means that demand isn’t exceeding supply, or maybe even that they have access to more supply that they’d use if they could sell to China, which is a massive market. Or even if they don’t have any excess supply, higher demand means they can set higher prices and still expect to sell all inventory.

    Nvidia has to sell to a Chinese buyer for 25% more than a US buyer would pay to have equivalent profit. It’s certainly possible that China is willing to pay more than that difference, but US private sector is supposed to be in desperation mode for skynet, in addition to having direct white house access of lobbying against China for mere trinkets in tribute. MSFT and others have the power to tell whitehouse/other republicans that they want to buy the H200s instead, and amplify warmongering BS as the reason. They just don’t want to buy them.

    Like the US car companies wanting to sell cars in China doesn’t imply that they are unable to sell cars in the US, it just means they want to sell cars to China and the US.

    US car companies are not supply constrained, including some of them with factories in China, and aren’t prohibited from selling all of their cars there if they were competitive. Nvidia has not been making H200s recently. It has astronomical record inventory levels (likely H200s based on lobbying win). Thier H20 cards that they sold to China the last 2 years, are the best value inference cards on ebay from China, but Americans were not allowed to buy them directly. Since about half of Nvidia GPUs are assembled in China, they have 0 problem with black market access to them, and massive secret Singapore customers of Nvidia are likely them directly profiting from Chinese black market with payment to Nvidia instead of pilferage of GPUs. I get that B200s B300s are better value/FLOP than H200s, but H200s could be priced to Americans/colonies on the same/similar $/flop, and if US/MSFT was really supply constrained, they’d buy or lobby government to force Nvidia to sell them at good $/flop. The Nvidia corruption is also likely to create new H200 production making newer GPUs “scarcer”


  • Microsoft’s finance chief, Amy Hood, argued that the cloud result could have been higher if it had allocated more data center infrastructure to customers rather than prioritizing its in-house needs.

    This is 98% chance a lie. Refusing azure clients wasn’t happening. They are saying the dedicated GPUs to copilot 365/windows/bing, but they would just slow tokens/second delivery or raise prices if they were constrained. Open AI/copilot service is flattening out is the far more likely explanation, and China/Anthropic/Google gaining share is apparent with frontend and LLM innovation.

    That said, windows 11 copilot is going at about 7tps on simple queries about its QOS, and slow service of paid models could impact azure. In Nov 25, they did drop big customer volume discounts. There were big price increases earlier in the year, so growth was in part pricing growth, and likely a drop in usage volume from previous quarter, or at least very stable. The AI frenzy, mostly openAI/msft/oracle/coreweave block of absurdly impossible capacity growth depends on keeping up with supposedly massive (token) demand growth. There are still a lot of free alternatives in the space, and app download figures usually accompany free promotional usage of latest breakthrough model (sora2 was free use on release. kilo code this week has free Kimi K2.5. Other coding tools have fully free or generous free tiers)

    Overall, this, and highly promotional industry, means its very hard for datacenter/LLMs to meet the hype. Deepseek 4 is hyped as a big leap forward, to be released in a couple of weeks. Everything AI boom is likely a lie, and Nvidia bribing Trump to sell H200s to China, at 25% export tariff, is proof of incapacity or unwillingness of US industry to deploy them.






  • this week‑long autonomous browser experiment consumed in the order of 10-20 trillion tokens

    at $60/m, that is $600M to $1.2B in full price cost, but 1/4 this is current standard pricing. Still, even if a buggy piece of shit, a 1-3m line code project in a week is impressive. OTOH, netscape 1.0 cost $4M to develop, with the advantage of working (though other advantage that it was your web page’s fault for not working).

    They set a very challenging experiment. There is a reason for chromium being a popular base for a browser. The more interesting experiment result is if it is ever usable. Are the bugs solvable by AI.