• 4 Posts
  • 2.85K Comments
Joined 3 年前
cake
Cake day: 2023年8月27日

help-circle
  • Software can be distributed almost for free. Once it’s written, sending another customer a copy costs practically nothing. Profit margins improve as software companies grow.

    AI doesn’t work that way. Every answer requires expensive chips, electricity, and data-center capacity. The next unit of software is nearly free; the next unit of intelligence is not.

    Software runs on fairy wishes, apparently. No chips or electricity involved.

    Local models run on your own CPU and GPU - like a video game. This fearmongering bullshit article keeps yapping China China China as if the American models and companies aren’t orders of magnitude worse. Worse about training, worse about inference, worse about money set on fire. Business fucking Insider is not taking a grand stand for software freedom. They’re just doing that one Tom Gauld comic for ‘our magnificent technical infrastructure’ versus ‘their hardware-devouring server farms.’





  • Does dumping even work, as an accusation? The shocking cheapness of Chinese models is in their training. Unless they’re just straight-up lying, and by orders of magnitude, that cost is not a factor in adoption.

    Tokens are not commodities. The cutting edge is all broadly similar, but people getting any real work out of these things have justified opinions about which ones are better and how. It is not an industry where you could half-ass your results and compete on subsidized price. Especially when half-assed results are possible with local models that run on your laptop.

    Nor do open models provide any serious retention. These companies emit an avalanche of whitepapers. If China tried to close the gate now, the horse is fuckin’ gone. Anyone can see how to produce a from-scratch competitor using institutional quantities of money instead of nation-state quantities of money. Plus you still have the old models. They’re behind, but they’ve narrowed the gap at an alarming clip. Increased pressure toward lower training cost and strictly local inference wouldn’t slow that down. Headline accomplishments from a few years back are already within reach of hobbyist budgets.


  • As the London subway illustrator said, “It’s not a map, it’s a diagram.” The technique conveys fine detail in a way that feels okay to your brain.

    Also, scrolling through this for the first time in a decade… all of these graphs show darkening toward the corner. Every single one of them is a shallow V. At best, one side / one face is uniformly flat. Ambient occlusion works because light really is less common in convex spaces. For any point near a wall… there’s a wall in the way. The closer you get, the more it’s in the way.

    Video games exaggerate the hell out of this effect (especially from Crysis through to this 2012 article), but heightened reality is kinda their schtick. Real people’s eyes aren’t that big, either. SSAO techniques advanced soon into wider low-frequency impact. I still have some dated PDFs with smooth-shaded church interiors.