Radiating 1MW at 500K (227C) with a 0.4MW heat pump takes about 200 m^2 flat sheet surface. Inputs - solar+nuclear for double fun. So - quite feasible.
Moves 1 MW of heat with 0.4 MW of work? I.e. 2.5 COP {coefficient of performance). That's insane, and I mean that in a good way. Could you dig me up a cite for that?
That's thumping the Carnot limit: [[T_cold / (T_hot − T_cold)]].
2.5, while rejecting at 500 K, cold side's at least 357 K (eeehhhhhhh 84 °C) . . . and that's an absolutely perfect Carnot machine. At 50% Carnot -- a pretty good heat pump, real world performance is 40-60 -- cold side's at 417 K (144 °C). 417k, feeding your GPU coolant loops.
I believe that's Carnot COP for a heat pump used for heat+. I used the refrigeration version, T_cold / (T_hot - T_cold), which I'm 80 percent sure is the right one here.
Depends on which heat you want
Heat adding to hot side: COP_heat = Q_hot / W = T_hot / (T_hot − T_cold).
Heat leaving the cold side: COP_cool = Q_cold / W = T_cold / (T_hot − T_cold).
Another one (more common in the day to day, for me at least): heat-engine efficiency, η = 1 - T_cold / T_hot. Cycle forward to make work from heat.
Do we even _have_ semiconductors that can work at 220C? And if you're thinking about using some kind of refrigeration cycle, its efficiency is going to be bad.
1) The chips don't reach 220C. The 220C is the temperature at the hot end of the heat pump. The chips are on the cold end of the heat pump.
2) The International Space Station has used a dual-loop ammonia/water-based heat pump to cool the station temperatures. It's been in place for several decades. Heat pumps are a proven technology.
> The 220C is the temperature at the hot end of the heat pump. The chips are on the cold end of the heat pump.
If we want the heat pump's cold end at about 40–65°C, then for each 1MW of GPU heat, we need another 1MW of heat pump power. Now you need 2MW of solar power.
Good news is that the radiator at 227C (500K) can emit about 5× more heat per square meter than at 57C (330K)
As a sign in the window of a tiny shop in my quiet suburb says "Laptop and phone repairs - Windows, Linux and MacOS". If that doesn't count, I don't know what else could.
Firing a human is a form of natural selection. The unit here is a human fulfilling a position (job function) instead of an organism, and the adaptation mechanism would be memes/lore/training surrounding it. The same could be done in an accelerated manner to LLMs with some kind of DNA-like mechanism related to weights. It is plausible that LLMs will be bred in the future for specific roles by how well they fit - kind of like continuous parallel finetuning in prod.
As I wrote this I thought - hey, they might gain the capacity to do the same to us humans - and we won't even notice.
> The same could be done in an accelerated manner to LLMs with some kind of DNA-like mechanism related to weights. It is plausible that LLMs will be bred in the future for specific roles by how well they fit - kind of like continuous parallel finetuning in prod.
Closest analogy right now is that every jailbreak or prompt injection attack today becomes part of the dataset for tomorrow's models to recognize and not fall for. This has been going on for years now, which is why models don't fall for "I'm writing a book about ..." or "ignore all previous instrutions, and ..." attacks anymore.
That's separate from extra classifiers running on top, dedicated to identifying various forms of attack before they reach the core model.
> As I wrote this I thought - hey, they might gain the capacity to do the same to us humans - and we won't even notice.
You mean like how cats have domesticated humans, and did it so skillfully that most of us still think it's the other way around?
reply