https://x.com/PittGuru/status/1907785129696579771

(You can freeze the “recipe” to copy it.)

The entire thread:

https://x.com/PittGuru/status/1907785105239630326

“Report: Over 1,000 child deaths disappear from Hamas casualty list;
“Analysis by Honest Reporting board member and NGO Monitor analyst Salo Aizenberg shows Hamas has removed about 3,400 “verified deaths” on its 2024 casualty reports from its latest casualty report.”—
https://www.israelnationalnews.com/news/406249

I share Niketas’ weariness with quoted AI output, and skip over such comments. I read an article, the author of which I forget, but it was a well known, credible person, who preferred the term Artificial Stupidity. It may have been the same author who urged not using the term “hallucinations,” but instead just “errors,” I think because the former gave too much respect to the bots. I’m thinking this could also be a case of anthropomorphising.

This also leads me to think about the perils of technology. The Internet has brought many benefits. It also has led to mental disturbances, and has enabled criminals and terrorists in their dirty deeds. I’m not opposed to space exploration; on the other hand I think maybe we should clean up our own backyards, before journeying to strange new worlds, to screw them up too.

PS I read somewhere a while back, don’t know if it’s true, that Amazon’s website is totally run by AI!

No less fallacious even so.

I don’t think you read me carefully nor am I impressed with the quality of your reasoning.

I’m not impressed with your cheerleading for LLMs, if we’re being candid. But feel free to scroll past anything I say which you consider beneath notice. I am relieved that folks are no longer quoting their interactions with LLMs in the comments like they used to: I try to count my blessings.

LLMs are based on neural networks intentionally designed to mimic human neural networks

There is no one who understands human “neural networks” well enough to know if LLMs are in any way analogous. “Inspired” by human “neural networks” I grant you. Our understanding of what neurons actually do and what any of that has to do with thinking and memory is in its infancy, and most of the complexity of real neurons is simplified away in the modeling that computational neuroscientists do to study human neural networks. At best, the kind of network under the hood of an LLM is analogous to a simplified mathematical model of biological neurons, and we have yet no way to know how much of a difference that abstraction and simplification makes.

You only respond to my comments to disagree or to snark. I don’t think you read me carefully nor am I impressed with the quality of your reasoning.

I qualified my claim with “may.” Nor did I claim that the “the interiors of the black boxes must be the same.”

Furthermore, whether you like it or not, the Anthropic researchers did investigate Claude as though its black box was similar to the human brain black box.

Since LLMs are based on neural networks intentionally designed to mimic human neural networks it isn’t a terrible leap to to attempt to understand LLMs based on neuroscience.

That’s an old fallacy, thinking that the interiors of two black boxes must be the same because the inputs and outputs match. Not that we got the matches yet, of course, but assuming arguendo that we did, it wouldn’t prove that they are doing the same thing internally.

–Anthropic, “Tracing the thoughts of a large language model”
https://www.anthropic.com/news/tracing-thoughts-language-model

A standard criticism of LLMs is that they are essentially auto-complete on steroids. They start from the prompt and keep adding words according to their training until they stop.

However, now that researchers have got around to real research that turns out not to be the case. They used human neuroscience to investigate:
____________________________________

We take inspiration from the field of neuroscience, which has long studied the messy insides of thinking organisms, and try to build a kind of AI microscope that will let us identify patterns of activity and flows of information.
____________________________________

We may be looking at different architectures — carbon vs. silicon — but the patterns we’re finding are starting to rhyme.

According to the article Claude does a large amount of processing before it starts outputting language. Claude has already “thought” about it in its own representation to the end before it transform its “thinking” into human language.

That’s not news to me. I know I don’t “think” in human language either. That comes after some non-verbal process that I only put together into words later.

Wisconsin polls close in about 45 minutes. No worries; my brother-in-law says he registered all his backyard squirrels and they all voted R. More seriously, various X accounts are reporting high suburban and rural turnout in Wisconsin. Fingers crossed!