mickey_mouse_demonstrates_a_thing_that_will_definitely_end_well.gif

@judgedread

> In every case where coders have told me that 'AI doing X is impossible' they have been wrong...

I don't know what claim I've made about which thing is impossible; I didn't even say they're not worth research. I just said "this is how you do context-sensitive memory with LLMs, you do a big-ass vector database of semantic knowledge and feed that into the LLM to bias generation" and "we're getting towards the end of what you can do with LLMs". LLMs aren't "AI", they're just token predictors. This is why they have the problems they have. Mickey Mouse, army of mops. You also get a lot of terrible feedback loops if you try to put them in charge of *which* knowledge they should treat as true and which they should treat as specious. Bigger context windows (i.e., an order of magnitude more RAM) solve some of it, better training solves some of it; Kimi-K2 is so far the only one that has correctly ascertained the purpose of a tiny chunk of awk code I wrote and that thing is so huge (1e12 parameters, so close to a terabyte at 8 bits per parameter) that it requires a GPU farm. (Feel free to try: `echo '++++[->++++<]>[-<+++++>>+++++++>++<<]<------.>>+++++.--.+.>.<[-<+>>>+<<]>[--<++>>-<]>---.+++++++++++++.+.<<<.>>>-------.---.<<<--.>.>>+++.-------.++.++[->+<<+>]>++++++.<<.<<.>[-<<->>]<<++++.[>>>--<<<-]>>>+.' | sed -E 's/(.)/\1\n/g' | awk 'BEGIN{print "BEGIN{p=0;"}END{print "}"}/\./{print "printf \"%c\",a[p]"}/\+/{print "a[p]++"}/-/{print "a[p]--"}/</{print "p--"}/>/{print "p++"}/\[/{print "while(a[p]){"}/\]/{print "}"}' | awk -f /dev/fd/0`. Most of them think it prints "Hello, World!" and even then, only the bigger ones can reliably guess "brainfuck to awk transpiler".) You want to see what's going to be workable in the next few years, assume RAM gets bigger, so you look at something too big to plausibly run in your house for less than five figures because that'll probably be four figures in a few years, and then reason "If we solved this specific mechanic behind this problem, how would it perform?" Absent an architectural shift (which cheaper GPUs with more RAM might get us), we're looking at something that is like Kimi-K2 with a handful of problems solved. Even that article I sent, right, the guy's like "This has some use but we're not there yet."

Previously, I did say (and maintain) two things: we can't simulate a guy and that we'd have to hollow out the moon (which I explained), and also that this is the panopticon. There are no "can't use amassed LLM sockpuppets to manufacture consent or to conduct massive real-time surveillance campaigns" restrictions against the government (and even when there are restrictions like that, the 2007 warrantless wiretapping and PRISM and the Biden Twitter/Facebook thing demonstrate that the government doesn't really care when there are restrictions) and you can see that the PRC and US are aggressively pouring money into this while governments that don't have much fear of uprisings are making minimal investments. (The room full of Russkies posting at boomers on Facebook is probably going to be replaced with LLMs from Yandex, but Russia's been half-assing their influence campaigns last 20-ish years at least.) That's going to be a qualitative shift in government and that is bigger than clickbait journos getting the ax. I showed you, right, on consumer hardware, it'll do a passable job at summarizing my notifications. You look at the NSA's exabytes under the desert in Utah, you look at the https://www.top500.org/lists/top500/2025/11/ and the US, which had been slipping in the rankings, suddenly ate the top 10 again, China went from owning the top slot to #24. And where you saw only governments before, you see Microsoft and NVidia. (OpenAI has, at current revenue levels, an order of magnitude gap between their $250B commitment to minimum buy for Azure and their revenue of about $3-4B/month. Microsoft already owns 27% and will acquire them unless they get a big contract. They are talking Q4 IPO and they're discussing advertising; advertising will not make up the shortfall and is a dry-hump for businesses that have run out of ideas. But "persuasive reasoning based on natural language analysis" is, in the case of LLMs, dual-purpose: they are eyeing government contracts.)

Whatever claims "all coders" make, I ain't those guys; I know what claims the median techno-mysticism enjoyers make, though. You're talking to me, not coders in general.
mickey_mouse_demonstrates_a_thing_that_will_definitely_end_well.gif

Public

mickey_mouse_demonstrates_a_thing_that_will_definitely_end_well.gif

Notices where this attachment appears