That wasn’t my takeaway from it at all. I think he’s quite precise in his definitions and the point he’s making about Rust, even if I found the prose a bit fatiguing due to the LLMisms.
Personally I think that the world needs more of your optimism. But what does your crystal ball say about energy costs and political stability?
It’s one thing to say that AI will help everyone create immersive games, but skyscrapers won’t be free unless energy is free. Do you also assume that AI will solve fusion?
What happens to the individuals who are not “strong” and how do they hold onto their remote work jobs? Why will the biomedical research technology you’re positing not be used to create biological weapons, or do you assume that AI also creates universal peace and harmony? If so, how does it do that while also preserving our ability to have our own ideas and disagree with each other?
If we want the great future you’re imagining, I think history teaches that we need to give at least as much attention to these questions as we do to making the technology work.
I think their work earns "theory" because it makes specific predictions both about how to make more effective prompt injection attacks and what activations you'd observe in the LLM during those attacks, and can also be plausibly extrapolated to suggest useful future research directions.
They were adopted by professionals long ago, and those highly tuned and validated proprietary models are going to kick the butt of the models that you have access to every day of the week.
Claude is already shockingly good at reverse engineering. Try it – it's really a step change. It has infinite patience which was always the limited resource in decompiling/deobfuscating most software.
I think the origin is in the phrase, "Will the dogs eat the dog food?" which was common VC-speak in the 90's and 00's, referencing dog food commercials that once ran on TV, and meaning something like "this has been made to sound great in an internal powerpoint presentation, but will customers actually like it?"
In 2015, Marc Andreessen memorably said of Mixpanel's success at product-led growth: "The dogs are fucking jumping through the screen door to eat the dog food. And he hasn’t done any marketing yet."
https://www.newyorker.com/magazine/2015/05/18/tomorrows-adva...
That then led to the idea of "eating your own dog food", because if even you won't eat it, what credibility do you have saying that the other dogs will?
It's not "the largest representable number" because you're not representing numbers in any rigorous sense. If I give you 64 bits, you can't tell me what number those bits represent (first, because the rules of the game are ambiguous - what if I give you 8 bytes that are a valid program in two different languages; and second, because even if you made the rules precise, you don't know which bitstrings correspond to programs that halt). And if I give you a number, you can't tell me which 64 bits represent that number or even if the number is representable, and that's true even for small numbers and even if I give you unbounded time.
It seems far more natural to say that you're representing programs rather than numbers. And you're asking, what is the largest finite output you can get from a program in today's programming languages that is 8 bytes or less. Which is also fun and interesting!
Heat is not by itself waste. It's what electricity turns into after it's done doing computer things. Efficiency is a separate question - how many computer things you got done per unit electricity turned into heat.