- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
Silicon Valley has bet big on generative AI but it’s not totally clear whether that bet will pay off. A new report from the Wall Street Journal claims that, despite the endless hype around large language models and the automated platforms they power, tech companies are struggling to turn a profit when it comes to AI.
Microsoft, which has bet big on the generative AI boom with billions invested in its partner OpenAI, has been losing money on one of its major AI platforms. Github Copilot, which launched in 2021, was designed to automate some parts of a coder’s workflow and, while immensely popular with its user base, has been a huge “money loser,” the Journal reports. The problem is that users pay $10 a month subscription fee for Copilot but, according to a source interviewed by the Journal, Microsoft lost an average of $20 per user during the first few months of this year. Some users cost the company an average loss of over $80 per month, the source told the paper.
OpenAI’s ChatGPT, for instance, has seen an ever declining user base while its operating costs remain incredibly high. A report from the Washington Post in June claimed that chatbots like ChatGPT lose money pretty much every time a customer uses them.
AI platforms are notoriously expensive to operate. Platforms like ChatGPT and DALL-E burn through an enormous amount of computing power and companies are struggling to figure out how to reduce that footprint. At the same time, the infrastructure to run AI systems—like powerful, high-priced AI computer chips—can be quite expensive. The cloud capacity necessary to train algorithms and run AI systems, meanwhile, is also expanding at a frightening rate. All of this energy consumption also means that AI is about as environmentally unfriendly as you can get.
The comparison of GPT parameters to neurons really is kinda sloppy, since they’re not at all comparable. To start with, “parameters” encompasses both weights (ie. the “importance” of a connection between any two neurons) and biases (sort of the starting value of an individual neuron, which then biases the activation function) so it doesn’t tell you anything about the number of neurons, and secondly biological neurons have way more dynamic behavior than what current “static” NNs like GPT use so it wouldn’t really be surprising if you needed much more of them to mimic the behavior of meatbag neurons. Also, LLM architecture is incredibly weird so the whole concept of neurons isn’t as relevant as it is in more traditional networks (although they do have neurons in their layers)
Another sloppiness that I didn’t mention is that a lot of human neurons are there for things that have nothing to do with either reasoning or language; making your heart beat, transmitting pain, so goes on. However I think that the comparison is still useful in this context - it shows how big those LLMs are, even in comparison with a system created out of messy natural selection. The process behind the LLMs seems inefficient.
I wouldn’t discount natural selection as messy. The reason why LLMs are as inefficient as they are in comparison to their complexity is exactly because they were designed by us meatbags; evolutionary processes can result in some astonishingly efficient solutions, although by no means “perfect”. I’ve done research in evolutionary computation and while it does have its problems – results can be unpredictable, it’s ridiculously hard to design a good fitness function, designing a “digital DNA” that mimics the best parts of actual DNA is nontrivial to say the least etc etc – I think it might be at least part of the solution to building, or rather growing, better neural networks / AI architectures.
It’s less about “discounting” it and more about acknowledging that the human brain is not so efficient as people might think. As such, LLMs using an order of magnitude more parameters than the number of cells in a brain hints that LLMs are far less efficient than language models could be.
I’m aware that evolutionary algorithms can yield useful results.