Ultimately the dotcom fantasy kind of panned out. A few tech companies have massive control over society now, with what is essentially cloud/internet business. They have a moat.
…But with the AI bubble, I think folks are underestimating how fast and low the “race to the bottom” is.
As random examples:
Look at something like Nemotron 4B, which makes a lot of mundane ‘AI’ data processing people assumed to be big and power hungry (with these data centers) basically free: https://huggingface.co/jet-ai/Jet-Nemotron-4B
And all this is accelerating. Alternate attention is catching on (see: Qwen 80B, Deepseek experimental, IBM Granite, probably Gemini). bitnet is already proven and probably next, and reduces the cost of matrix multiplication by an order of magnitude or two.
In other words, AI as a “dumb tool” is rapidly approaching “so cheap, it’s basically free to run locally on your phone,” and you don’t need all these megacorp data centers for that. There’s no profit in it. It’s all fake planning, and the ML research crowd knows it. That’s much more extreme than the dotcom hype, where cloud hosting/dev cost is kind of a fundamental thing.
Allegedly Google, early in this craze:
https://semianalysis.com/2023/05/04/google-we-have-no-moat-and-neither/
Ultimately the dotcom fantasy kind of panned out. A few tech companies have massive control over society now, with what is essentially cloud/internet business. They have a moat.
…But with the AI bubble, I think folks are underestimating how fast and low the “race to the bottom” is.
As random examples:
Look at something like Nemotron 4B, which makes a lot of mundane ‘AI’ data processing people assumed to be big and power hungry (with these data centers) basically free: https://huggingface.co/jet-ai/Jet-Nemotron-4B
Look at GLM 4.6. I can run it on my Ryzen/3090 desktop, for free, and for the first time, I feel like it’s beating Claude and Gemini in some stuff, at 7 tokens/sec: https://huggingface.co/Downtown-Case/GLM-4.6-128GB-RAM-IK-GGUF
These are both literally from the past day.
And all this is accelerating. Alternate attention is catching on (see: Qwen 80B, Deepseek experimental, IBM Granite, probably Gemini). bitnet is already proven and probably next, and reduces the cost of matrix multiplication by an order of magnitude or two.
In other words, AI as a “dumb tool” is rapidly approaching “so cheap, it’s basically free to run locally on your phone,” and you don’t need all these megacorp data centers for that. There’s no profit in it. It’s all fake planning, and the ML research crowd knows it. That’s much more extreme than the dotcom hype, where cloud hosting/dev cost is kind of a fundamental thing.