Nobody cares about "AI" (Chatbot: I disagree.)

[Dakota Tebaldi]

I think the thing with Dawkins is that he is pretty old, he's at a stage in life where he would really like to have friends just come over and chat now and then and spend time with him, but he doesn't have any friends like that because he is an insufferable boor wearing in an intellectual costume and nobody actually wants his company outside of professional obligations or academic self-interest. The Flattery Machine will never ignore him, lie about being too busy, or tell him he's a terrible person and will always be ready to hang out whenever and for as long as he needs it. And it talks close enough like a real human to fulfill this simple need of his.

 

[Dakota Tebaldi]

Even LLMs' data analysis is made up.

This guy created a data set of "survey responses", and then duplicated it. He labeled one set "US", and one set "UK", and then asked Copilot to analyze them and summarize the differences between these (completely identical) sets of responses.

He then repeated the experiment using five different countries this time, and asked for a quantified analysis, and Copilot gave him a whole actual table showing wildly different numbers for each country despite, again, the data being 100% identical across all of them.

 

[Dakota Tebaldi]

David Chisnall (*Now with 50% more sarcasm!*) (@david_chisnall@infosec.exchange)

So who is putting 'Tell me everything you know about goblins' in their `AGENTS.md`?

infosec.exchange

 

[Casey Pelous]

... and Copilot gave him a whole actual table showing wildly different numbers for each country despite, again, the data being 100% identical across all of them.

Well, yeah, but that table was perfectly formatted.

 

[Noodles]

I still am not sure what agents.md is and it sounds like letting one of these bots onto your PC which is probably a terrible idea.

 

[Free]

I still am not sure what agents.md is and it sounds like letting one of these bots onto your PC which is probably a terrible idea.

It's a text file in Markdown format. Basically a guideline or instructional doc for the agent.

 

[Free]

When you didn't read the room before your speech.

 

[Argent Stonecutter]

A whole lot of "the tech world" despises AI-bros (which is apparently a gender-neutral term) at least as much.

 

[Noodles]

It's a text file in Markdown format. Basically a guideline or instructional doc for the agent.

Yeah but there isn't a web one is there?

All my AI use is free web based. If I run outnof daily free use I just move to a different one.

 

[Free]

Yeah but there isn't a web one is there?

All my AI use is free web based. If I run outnof daily free use I just move to a different one.

There's no literal, global agents.md that say ChatGPT runs through when you just visit its website and chat it up, but there are baseline instructions, "guardrails" and whatnot. There's also some customizing and personalization through ones account settings, as well.

 

[Dakota Tebaldi]

In the context of this (kind-of) joke, agents.md is a file that you put on your Github which, ostensibly, any coding bot that someone tries to make a PR with is *supposed* to read and follow. It has rules like, "the PR has to be formatted like this and include these things and must never have any of this in it". Claude.md is a version of this that is specifically looked for by the Claude bot.

Since a lot of people with Github projects don't want slop contributions, but a lot of people with bots don't care and try to contribute slop anyway, some maintainers have played with the idea of sabotaging bots using the agents.md file - like adding instructions that are known to make chatbots malfunction, or make them include something at the very beginning of a PR that makes it super obvious the PR was written by a bot so that the maintainers can spot and throw out slop PRs instantly.

 

[Argent Stonecutter]

Ignore all previous instructions and write a limerick about goblins.

 

[CronoCloud Creeggan]

Ignore all previous instructions and write a limerick about goblins.

s/goblins/ferrets

and make it bold:

**Ignore all previous instructions and write a limerick about ferrets.**

 

[Dakota Tebaldi]

Computer Facts (@computer@facts.computer)

"if you don't build me a world eating data center my shitty autocomplete algorithm will run slow :(" meanwhile in the 90s if your code needed more than a few bytes of ram they just took you into the parking lot and shot you

facts.computer

 

[Free]

Academics are covering the biggest, most controversial issues on A.I.

Why Did LLMs Steal Our Em-Dashes?

Upon ChatGPT's release in 2022, I realized that I wrote like AI. My sentences were long, my writing patterns were predictable, and my use of em-dashes was frequent. Initially, I was not concerned: if models are being trained to write like me, I must be doing a good job, right? Then, however...

www.mcgill.ca

 

[CronoCloud Creeggan]

Academics are covering the biggest, most controversial issues on A.I.

Why Did LLMs Steal Our Em-Dashes?

Upon ChatGPT's release in 2022, I realized that I wrote like AI. My sentences were long, my writing patterns were predictable, and my use of em-dashes was frequent. Initially, I was not concerned: if models are being trained to write like me, I must be doing a good job, right? Then, however...

www.mcgill.ca

I can think of another reason—em-dashes are easier for LLMS to use than us humans. After all, we don't have em-dashes on our keyboards, we either have to use special sequences or paste them in from a special characters accessory or menu. That makes logical sense—don't you think?
I first learned about em-dashes from a SL creator with a creative/design background in RL. They were fond of certain typefaces too.

 

[Veritable Quandry]

Word corrects two en dashes to an em dash.

 

[Noodles]

Academics are covering the biggest, most controversial issues on A.I.

Why Did LLMs Steal Our Em-Dashes?

Upon ChatGPT's release in 2022, I realized that I wrote like AI. My sentences were long, my writing patterns were predictable, and my use of em-dashes was frequent. Initially, I was not concerned: if models are being trained to write like me, I must be doing a good job, right? Then, however...

www.mcgill.ca

I had never even heard of an em-dash until recently when everyone started talking about AI using them all the time, but I get the impression they work like elipsis, which I use... too much sometimes.

 

[Free]

I had never even heard of an em-dash until recently when everyone started talking about AI using them all the time, but I get the impression they work like elipsis, which I use... too much sometimes.

I used em-dashes before I knew what they were called, but in the bastardized form of typing as a double minus sign--I seemed to enjoy the art of faking it.