News for hackers. Stuff that matters.
Re:AI's energy footprint by dr_dshiv on Wednesday May 21, 2025 @07:57AM (#44050497)
Distillation seems like a pretty robust approach. There is no good reason to think that GPT5 or 6 won’t have massive rapid efficiency gains.

He’s one of the most knowledgeable persons in the world on the topic. It’s like saying that the ceo of BMW isn’t citable on conversations about expected cost decreases in electric cars.

Any environmental topic is susceptible to a huge amount of groupthink. The world isn’t so binary as people make it out to be. It is far from truth that LLMs=bad for environment, any more than computers=bad for the environment.

Re:AI's energy footprint by wyre on Wednesday May 21, 2025 @09:47AM (#44051419)
I think its more like saying that the ceo of Tesla isn't citable on conversations about the expected advancement in electric car technology.

What I mean is that I have a healthy level of skepticism with Altman. He has to constantly battle for funding. Surely, he must be knowledgeable about LLMs, but he's the CEO of the largest AI company in the world, his PR needs to give "most knowledgeable person in the world on the topic" but I think that title should go to all of the engineers and developers working on these technologies and not a capital founder.

All that said, I agree that LLM's being bad for the environment is a complex topic. I think it would be more accepted if people had safety nets and could be excited for AI to take their job instead of having to be terrified, or if AI isn't just used as another tool for increasing wealth inequality.

Re:AI's energy footprint by dr_dshiv on Wednesday May 21, 2025 @10:22AM (#44051746)
One reason to take these expected cost reductions more seriously is because it DOESN'T fit the desired narrative for OpenAI. Ideally, from a front-runner competition perspective, it would NOT be possible to create the same level of quality at < 1/10 the cost/energy within a year. But, distillation and other approaches seems to make this possible. There is a LOT of room for optimization with LLMs.
Re:AI's energy footprint by Fraterkes on Wednesday May 21, 2025 @11:45AM (#44052723)
Alright, so can you actually name some numbers so that the stuff you are claiming is at least fasifiable?

Altman mentions a 150x increase in efficiency, you claim that trend will continue through to gpt6. At that point these models would be 22500 as efficient as they currently are, which would mean generating a 10 hour long video would cost around the same amount of electricity as running your microwave for 15 minutes. Will you have some introspection if that doesn't come to pass?

Re:AI's energy footprint by dr_dshiv on Thursday May 22, 2025 @10:09AM (#44062223)
Sure, I'm game. What's your prediction?
Re:AI's energy footprint by dr_dshiv on Thursday May 22, 2025 @12:10PM (#44063458)
Of course I predict this will turn into a logarithmic curve not exponential over time. But I’m always happy to bet on future outcomes with specifics. Like, that gpt5-level intelligence will be able run on contemporary MacBooks within 2-3 years of release.