Posts

Showing posts with the label energy consumption

AI Subscription Price Hikes Imminent

Image
The human mind is far more energy efficient than large language models. Even a genius like Einstein or Poincaré used 30 watts max to power their brain. 🙂 In contrast, estimation of META's Lama 2.0 front tier model required 700 MWh, and later versions much more. AI service providers are using brute force in terms of lots of resources such as power and GPUs as opposed to design improvements in their struggling attempt to achieve human intelligence and creativity. Throwing lots of power and GPUs at the challenge will probably not lead to the Singularity as larger models yield less and less incremental return. AI advocates calls it 'scaling', but I call it 'diminishing returns'. 😀 As a result of the brute force approach, AI service providers are losing money on their $20 a month subscription plans. VC money is currently paying not just for the service provider's infrastructure, but also subsidizes the end user pricing. For example, in New York video rendering comp...

Interest Rates Hit AI Service Firms

Image
Companies like SpaceX or Anthropic are issuing junk-rated debt at 1% to 2% interest rate premiums to government bonds or around 7% to 8% in total. This creates severe financial risk. Smart investors are no longer willing to buy equity in private or publicly traded AI companies because they know they are wildly overvalued. Funding will try up as these AI services simply don't generate sufficient revenues to justify trillions and trillions in capex. They are also not cost effective due to their high energy costs. Compare the power required to estimate an LLM relative to the human brain's modest 12 to 20 watts. It is not a contest.