It’s safe to assume that any metric they don’t disclose is quite damning to them. Plus, these guys don’t really care about the environmental impact, or what us tree-hugging environmentalists think. I’m assuming the only group they are scared of upsetting right now is investors. The thing is, even if you don’t care about the environment, the problem with LLMs is how poorly they scale.

An important concept when evaluating how something scales is are marginal values, chiefly marginal utility and marginal expenses. Marginal utility is how much utility do you get if you get one more unit of whatever. Marginal expenses is how much it costs to get one more unit. And what the LLMs produce is the probably that a token, T, follows on prefix Q. So P(T|Q) (read: Probably of T, given Q). This is done for all known tokens, and then based on these probabilities, one token is chosen at random. This token is then appended to the prefix, and the process repeats, until the LLM produces a sequence which indicates that it’s done talking.

If we now imagine the best possible LLM, then the calculated value for P(T|Q) would be the actual value. However, it’s worth noting that this already displays a limitation of LLMs. Namely even if we use this ideal LLM, we’re just a few bad dice rolls away from saying something dumb, which then pollutes the context. And the larger we make the LLM, the closer its results get to the actual value. A potential way to measure this precision would be by subtracting P(T|Q) from P_calc(T|Q), and counting the leading zeroes, essentially counting the number of digits we got right. Now, the thing is that each additional digit only provides a tenth of the utility to than the digit before it. While the cost for additional digits goes up exponentially.

So, exponentially decaying marginal utility meets exponentially growing marginal expenses. Which is really bad for companies that try to market LLMs.