There needs to be a Hanlon’s razor but for suspecting something is LLM generated. The typo and grammar mistakes in the response lead me to think it’s a human, but no way to really know.

I GUARANTEE that was written by AI. This type of slop is unmistakable

> The typo and grammar mistakes in the response lead me to think it’s a human

People using LLMs to "spam" slop already caught up on this sentiment and are purposefully introducing grammar and spelling mistakes in the outputs now. I'm not saying yay/nay in this particular case, just sharing what I've seen in the wild. If a company get complaints that customer support is too robotic after starting to use LLMs for it, making it more concise and introducing subtle mistakes are two surefire approaches they'd get recommended.

I refuse to believe there are actual human beings so blatantly splitting a single concept into three artificial points.

Been happening for years. You are not talking to a Developer but someone employed to do "support" who 1) might not know much or 2) think they know a lot but they don't 3) can't tell you the actual reason so try and BS an answer that will make you go away.

There needs to be a Grey's law but for suspecting something is LLM generated.

This doesn't make sense, even if the LLM had access to the source code

But yeah I kinda understand why would they want to block stuff with 'web'

Strongly doubt that a LLM would say "domains that has .web". Maybe it's a human paraphrasing a LLM.

For what it's worth and I know it can't be trusted, one of the AI detectors reported that sentence as "Your most AI sentences"

In this case, the combination of "web" and ".one" triggers these security measures.

It also flagged the whole email as 100% AI.

Plus this idiot bot has confabulated a subdomain.

Has anyone ever tried giving LLMs a backspace? I have long wondered if their responses could be better if they had the ability to correct earlier parts of their text. Imagine how much harder it would be for us to write coherently with no backspace!

Thinking mode is exactly this. The LLM is allowed to dump thoughts, correcting itself along the way, and only surfaces the cleaned-up final answer to the user

It's not exactly this. It's something like it. But it's not a token that erases a previous token. Not that I know how that would work or how you'd get training data (edit histories of internet comments seem too few).

(You could have a token that hides previous tokens, but that'd be rather closer to CoT.)

I'm not sure how that would work. You'd need to be able to reverse the operations in the KV cache, and I'm not sure if it's doable or not. And then you'd need to pick another path, or "rebalance" the probabilities or you'd go the same token path. CoT is probably easier to implement (but not necessarily better).

You could set up checkpoints where you store off the current state, and roll back to those checkpoints?

Definitely, this is the easy "bruteforce" way, but it would require more memory, and would not resolve the second part of the problem (i.e "how to pick a better path").

I'm not an LLM engineer, but I just got an idea of how it could work, combining CoT + checkpoints :

- use checkpoints to save KV cache before trigger CoT

- trigger CoT, save result as a summary

- go back to previous checkpoint

- instead of generating tokens, add result of CoT summary as input tokens

- continue normally

For the price of twice the KV cache memory, the context stays perpetually small, allowing smarter sessions. You can even apply that continually by summarizing tool calls, etc.

This idea is free.

I'm not sure it's that advantageous though: it consumes more memory, and the sessions are already quite long at 1M+ tokens. One would need to run the economics down, and just test if the shorter sessions are actually smarter with the continuous summarization.

You could burn a few tokens by prompting the LLM to summarize the omit all of the "wrong" answers (where it overrode itself), after the user accepts the response. It'd probably reduce token use in the end, especially for long sessions.

Backtracking, of sorts. The LLM descends a seemingly fruitful path, and all of a sudden that path seems less so. So you backtrack a few tokens ("erasing"/"undoing") and resample the probability distribution a bit back in the stream.

Would it retain some memory of the fact that it backtracked? That path becoming less likely somehow?

Yeah that's what I was imagining. No idea how though, so this is just rambling on my part :)

ChatGPT does this. It will present an initial response, then change its mind partway through the thinking process and present a different response.

Oh wow, I'd like to see this. Any prompt which reliably triggers it? (Seahorse emoji comes to mind!)

It does it quite consistently on higher reasoning.

I believe they have made display answer a tool call so that it can keep updating its answer as it thinks, instead of doing all the reasoning up front, essentially inverting the usual flow. You can use this pattern with any model that supports tool calling.

Copilot does this too, kinda. It will give an initial response, realize that it might tangentially criticize Microsoft, then change that response to a refusal.

That's not a backspace, that's just a nanny filter. I think all bigco LLMs have something similar. e.g. Claude obviously uses this to prevent Fable from doing security work.

Chatgpt will erase text it wants to censor if you get near sensitive topics.

What sort of prompt would exercise this?

I inquired into a court case against a police officer who committed a sex crime. I only wanted to know the disposition of the case with no details. It briefly spit out a response involving the victim and wiped it away with a refusal to answer more questions. This was a year ago and the behavior may have changed by now.

That's just a standard nanny filter. The filter realized it was in the process of saying "sexual assault" or the like, and just shut the conversation down.

Plus the idiot bot has confabulated a subdomain.

Why did you post the exact same comment three hours before this one?

To do this you need to run

    hallucinated command

but wait, this command only existe for aws cli, not for oracle cli, so you need to run

    actual command

Regularly happens to me in chatgpt.