The Year Energy Woke Up To AI - Audioblog 14: Generative AI - The Power and the Glory

[@CleaningUpPod]

Thank you for joining us for this episode of Cleaning Up. Can you tell which Michael is the real Michail? For a full reading list to accompany this audioblog, visit and subscribe to our substack: open.substack.com/pub/cleaninguppod/p/a-super-secret-ai-reading-list

[@korona3103]

My friend works for a consultancy firm advising companies about integrating AI into their businesses. The last couple of projects concluded that the only thing AI could replace was their own consultancy work 😅

[@PazLeBon]

Pmsl, sounds about right too

[@Ben-gm9lo]

The AI you used to produce your artificial bits has some unusual tempo - you speak in a consistent way, the AI doesn't mimic that well enough to escape detection! Thank you for another rivetting blog.

[@MLiebreich]

Agreed, it's not perfect. You have to tweak each paragraph, and then generate it a number of times, keeping the best version. It's painstaking stuff. Every so often it does a paragraph that's absolutely spot on, which is very uncanny, but yes, most of the time it's close but no cigAIr! Give it two more years...

[@PazLeBon]

@@MLiebreich everyone getting lazier

[@ronaldgarrison8478]

There are enough trends and counter-trends going into this to give someone a splitting headache. And it creates enormous room for hype.

[@itsureishotout-itshotterin3985]

Excellent summary of the AI-Electrical supply issue.

[@sleep_gaminglife]

Excellent episode!

[@blandineetarthurdelassus8852]

Brillant, thanks!

[@PazLeBon]

They need Ai to help them get off our dying planet that they caused and are panicking

[@BobQuigley]

The 7 pound brain inside my 180 pound body uses 20% of the energy i consume.

[@ronaldgarrison8478]

If there's a concern about power demand for data centers, maybe it would be wise not to go down rabbit holes like bitcoin mining.

[@adus123]

Analog Could be the Future of AI. Analog GPUs hold significant promise for the future of AI due to their potential for increased energy efficiency and processing speed. Unlike digital GPUs that represent data as discrete 0s and 1s, analog GPUs leverage continuous physical phenomena, such as voltage or current, to perform computations. This approach can lead to substantial energy savings, as it avoids the overhead associated with digital encoding and decoding. Additionally, analog computations can be inherently faster for certain types of AI workloads, such as those involving matrix multiplications, which are fundamental to many AI algorithms. Energy Efficiency Energy consumption is a major concern for AI, especially as models grow larger and more complex. Analog GPUs offer a potential solution by significantly reducing power requirements. This is crucial for both reducing operational costs and minimizing the environmental impact of AI.

[@beatreuteler]

Dear Michael and Dear AI bot: For any future version: Is there a way to remove the sort of sharpiness in the voice that your natural voice doesn't seem to have? It would be more agreeable to listen to. Regarding the matter: Listening to the episode and thinking about the implications to me it seems that in certain cases it may make a lot of sense to delay some of these projects by just a few years, for example like the Irish do it due to power considerations, resulting in the harvesting of the opportunities to then use even more efficient xPU's so the same computing power will be able to run 24/7 without a new power plant need being built next to it. On top of that it may make most sense to tear down some 6 of the running data centers in a broader region and replacing them by one that beats them all in processing and uses 1/10-th of the power of just 1 out of the 6 did before? Hence reducing power demand for the same output by /60.

[@MLiebreich]

I learned a lot creating this episode. To train the voice, I used all of the previous audioblog episodes of Cleaning Up. Once I was using it to make the recording, I realised that I might have been better off using less training material, but of a more consistent quality. The voice, intimation and pace were quite variable - for some paragraphs it was absolutely uncanny, but others were pretty poor. I had to regenerate some of the paragraphs dozens of times to get one acceptable version, and then Oscar (our brilliant producer) had to adjust volume levels and do lots of other tweaks to make it sound like one piece rather than a collection of paragraphs. Regarding rebuilding data centres, I'm guessing some of that might go on. But if you have space, power and cooling for 20MW, but what you need is 200MW, there's not much you can do. And if you need 2GW, you're in a whole different world!

[@mr.e7379]

GOOD WORK!!!

[@eclecticcyclist]

Building AI data centres require electricity now, they can't wait fouteen years or more for a new nuclear power station, or rely on gas generators for fouteen years, besides they are more likely to go for the cheapest energy which is currently from renewables + storage. What happens when they have to shut the power station down for refueling anyway? AI data centres have a verry erratic demand for electricity so every on is likely to have battery storage to smooth out demand, hence Tesla have an advantage as they also produce battery grid storage and they already have a couple of hundred transformers in stock.

[@beatreuteler]

Well, if they would be allowed to wait 14 years they won't need the nps any longer as within 14 years power demand of this AI application is shrinking a few orders of magnitude.

[@laughinggas5281]

Does anyone wonder why the Chinese are rolling out renewable energy at a frantic pace? It's because they realize that the country with the lowest power cost is the country that will win in New Era of Bitcoin and Ai and of course with industry in general. So when these clowns in the administration think that they're going to solve this AI issue by building gas-fired power plants they are absolutely wrong.

[@ronaldgarrison8478]

20:20 Watch your language.

[@MLiebreich]

I didn't name it FERC!

[@PazLeBon]

Lol

[@Nikoo033]

😂

[@macmcleod1188]

The human brain runs on 20 watts at rest and roughly 100 Watts at peak performance. Over the last year techniques have been developed to reduce the electrical requirements of AI significantly. There's no reason that AI will not become more efficient over time. I agree that the cost of electricity for AI may make some our options prohibitively expensive in the short-term.

[@PazLeBon]

Only cos we don't understand it, we may be utlising power 10,000 times more in our conscience

[@macmcleod1188]

@@PazLeBon no, energy consumed by the brain is easy to measure,the work was done decades ago, and has been verified independently. Today, it's at the point of "common knowledge," Our current implementation of AI is energy inefficient but it improved a lot over the last 18 months.

[@PazLeBon]

@@macmcleod1188 lol aren't measuring consciousness , they don't know what it is ;)

[@macmcleod1188]

@@PazLeBon Wheaton may never get a robot with consciousness or it may be an emergent behavior which is unavoidable. Doesn't matter. We're going to have a eyes smarter and the more capable than most humans even if they like consciousness. That's going to attack the foundation of the capitalist economy unless it turns out these hey eyes and embodied eyes cost hundreds of dollars to operate. Currently they seem to cost less than $10 per hour to operate.

[@laughinggas5281]

Why can't we use AI to solve the power consumption of AI

[@jimgraham6722]

Ultimatly there will be AIs running on a few hundred Watts if not less. At this point all these power stations will be toast.

[@mmimpact-2024]

Fantastic wrap up about this hit topic. As the publier is endorsing the content délivred, Who gîves a f**k about the part that was generted with IAs. (For those who cares, the answer is given at the end of this crucial episode)

[@PazLeBon]

Lots of folk

[@DirkLondon]

People will soon wake up to the realization that not predictive & generative Alghos are the main constraints in systems optimization & resilience But the sensors, actors, valid, stable context data & ‚market‘ rule designs. As well of human oversight to identify ‚overfitting‘. As the inventor of UML said: A fool with a tool …

[@PazLeBon]

Don't require comma after valid

[@beatreuteler]

@@PazLeBon Cool. This said, Dirk London is not an AI bot.

[@PazLeBon]

@ talks as much sense as one tho;)

[@glyngreen538]

I don’t ever see me using AI software personally. Definitely not paying for it. I hope AI will overall prove positive for humanity but I’m cautious and sceptical. Things could go badly wrong and I don’t trust the billionaires in control.

[@beatreuteler]

You are already using it unknowingly. Just listen to a wheather forecast.

[@adus123]

Don't knock it until you've tried it. Lol Its just a tool just like your Smartphone or Desktop PC/ laptop. Anything can be bad in the wrong hands, it's just a fact of life.

[@ryuuguu01]

I think this podcast will age in a matter of weeks. DeepSeek has released its model as open source. It only needs 132Mb and uses much less energy than ChatGPT. It is not the same as ChatGPT but does do things that ChatGPT can not do. Also, China being 5 years behind the US is wishful thinking by American analysts, as shown by its release of DeepSeek. On the 10-year scale, China also seems to be ahead in optical computing although it is still early.

[@PazLeBon]

It's Chinese

[@ryuuguu01]

@@PazLeBon Yes, I said it was Chinese. What is the point of your comment?

[@beatreuteler]

@@ryuuguu01 I don't speak Chinese, maybe?

[@ryuuguu01]

@@beatreuteler It runs in English. Also, it is completely open-source so people in the field can use it as a starting point to build English versions. The only thing I have seen people speculate about is what type of answer it would give if you asked it about Tiananmen Square.

[@PazLeBon]

@@ryuuguu01 who'd use a communist system, kinda obvious question