For a long time, the chip manufacturers had an inclination to simplify the hardware and rely on the software adapting and optimizing. But for decades this bid failed. Now we have the unrelenting AI capable of finetuning kernels relatively quickly. Maybe simpler hw will work this time? Note: not sure if TPU/NPU is not only simple but also too limited.
I think Leopold of "situational awareness" fame predicted this, tried to make it widely known on podcasts and x, and still started billions worth of hedge funding based on this.
it is probable that you don't get the deepest pockets by wasting money. anyone that pays so much for memory means they expect to get even bigger value from that.
I am taking some bromelain and papain capsules which according to some (small) studies should help. In my experience, the floaters became more transparent, but don't disappear. This is still useful i guess.
Well, then when it detects an air-gapped environment it will just behave differently. I feel like we underestimate in general the way agents behavior changes when the environment changes. Related: "power corrupts"
Training a model is more like taking an alien (from a different universe) and teaching it a huge chunk of the entire human civilization. A better comparison would be how much energy it takes for a human to produce a page a text (roughly the equivalent of 1hour of thinking?), and how much energy it takes for an LLM. GPT Sol estimates a frontier LLM is 10-50x times more efficient than a human at generating a page of text.
Well, we are dumb and we make stupid mistakes, and code review can find and fix them sometimes. I am very surprised that the practice of "review" is not more widespread. For example when I see medical doctors going YOLO with their patients, ppl driving excavators, etc.
to my knowledge, the cost of uranium is almost negligible compared to the capital cost of building the plant. so as long as a market exists, you can choose whatever strategy: buy a big buffer, or just don't care if price oscillates x times.
does synergy works better now? 3 years ago, every week I would get into a situations where one machine was not connecting to the other, and I had to randomly restart synergy so maybe it connects. fun to do that 5 min before the meetings.
I suspect that you can modulate nuclear power too, but why do it? after you started the reactor it runs practically for free? (the fuel cost is so small; or it costs the same to run full power of half power). disclaimer: I did not read actual details about nuclear power plants designs in the past 20 years, so i'm vibing from first principles and bad memory
Europe was spending 200 billions / year on gas from russia. I imagine they could try to build 100 reactors for that price, but it would take a couple of years I imagine...