I've been using ChatGPT and Copilot to help with some things related to standing up my new business and putting the website together, and overall it's been extremely helpful. This is not to say that it doesn't go off the rails from time to time.

I was trying to determine how many rounds I should list as a minimum with each course I teach and so I asked it (Copilot, in this case, as I didn't want to pollute the HTML/CSS/JS discussion I was having with ChatGPT), figuring that it could give me a response that reflected what other providers of the course say, and it did that quite well, but then it said:



I guess these LLMs are far more versatile that I gave them credit for!

So, being naturally curious, I responded, "So, you teach this course? How many times have you taught it?"

To which it replied that, no, it doesn't teach this course or any other course and that it's role is to provide "rigorous, contradiction‑free, audit‑ready information."

This is something the Copilot is particularly bad about. It frequently starts it's responses with things like, "Here’s the clean, authoritative answer," or "Here’s the authoritative, practical breakdown." Then, when you point out something that it got glaringly wrong (and here I'm talking about substantive, factual errors, not the kind of slop demonstrated by the above) it will correct them (sometimes wrongly) and tell you why what it said before wasn't really wrong (although this behavior does seem to have largely gone away and been replaced by more mea culpa type replies). But the "Here's the authoritative, rigorous, accurate answer" is clearly boilerplate prattle that has zero value and, in my mind, is detrimental because it preys on human nature to accept such claims at face value provided the following nonsense is confidently and plausibly worded.

I think that LLMs sometimes "absorb" the context of the source material they are referencing. In your case, it probably came across a teacher's personal anecdote relevant to the query. The LLM likely absorbed and exposed the first-person context into its reply to you.

Just a guess.

Humans also sometimes speak badly (or use inappropriate context) as well. We as listeners usually recognize the real underlying meaning and successfully continue the conversation with that understanding, without finding it necessary to correct our interlocutor.

 

I think that LLMs sometimes "absorb" the context of the source material they are referencing. In your case, it probably came across a teacher's personal anecdote relevant to the query. The LLM likely absorbed and exposed the first-person context into its reply to you.

Just a guess.

Probably absorbed from more than just one teacher's anecdote in order to rise to the level of statistical relevance in token selection.

Humans also sometimes speak badly (or use inappropriate context) as well. We as listeners usually recognize the real underlying meaning and successfully continue the conversation with that understanding, without finding it necessary to correct our interlocutor.

It tried to blow it of as "I was just using a figure of speech." Not a figure of speech I've ever heard. Now, if it had said, "If I were teaching this course....", that would be reasonably in line with the persona it tries to but forth pretending that you are conversing with a human being.

 

Probably absorbed from more than just one teacher's anecdote in order to rise to the level of statistical relevance in token selection.



It tried to blow it of as "I was just using a figure of speech." Not a figure of speech I've ever heard. Now, if it had said, "If I were teaching this course....", that would be reasonably in line with the persona it tries to but forth pretending that you are conversing with a human being.

Here's something you can try:

Ask the agent a moderately complex question to which you already know has a definitive answer.

When it replies with the (presumably correct) answer, tell it is wrong, give it a reasonable explanation (that supports the given correct answer), and then give it the same answer rephrased.

A normal human would get exasperated ("That's what I said!").

I predict the agent will:

1) Apologize for the mistake;
2) Agree with your reasoning;
3) Rephrase (or parrot your version) of the same correct answer.

I have noticed this behavior.

If it gets exasperated instead, I will be both pleasantly surprised and encouraged that the models are making progress.

 

AI is harming us in many ways, for example it is now apparently considered sensitive, controversial to ask about election results from completed elections, I think we all know why that is:

 

I think we all know why that is...

Do we all?

You should tell us so we can confirm your assertion.

 

1,000 AI Agents Started Agreeing Without Anyone Telling Them To

​

"Populations of individually aligned agents can settle into stable, collectively misaligned states purely through conformity."

– computational social scientist Giordano De Marzo

 

1,000 AI Agents Started Agreeing Without Anyone Telling Them To
View attachment 370453​

 

https://www.bbc.com/news/articles/cvgx4yd1gl2o
Tech leaders say AI means less work - their staff say they work up to 90 hours a week

For years now, executives at companies that are pouring hundreds of billions of dollars a year into developing various artificial intelligence tools have insisted that the technology will ultimately mean people will spend less of their time working.

An engineering director at Google said four years ago that AI would deliver a four-day work week by 2025.

Earlier this year, and just one year after that engineering director's prediction, OpenAI took up the challenge, in a manner of speaking. It formally urged companies to start testing out a four-day work week (with no change in pay), claiming that AI will soon be able to speed up so much human labour that the
corporate world should prepare itself.

However, a former OpenAI technical employee who left the company last year told the BBC the firm never actually trialled the four-day work week it suggested others should try while they were there.

Instead, the person described what was often a gruelling work culture marked by frequent "crisis meetings", working on weekends, and "super cut-throat" performance reviews that would see colleagues suddenly let go.

"You go in on Saturday or Sunday just to catch up or make sure things aren't broken," the person said.

Other companies have also pushed the idea that AI will effectively reduce the number of hours people need to work, for better or worse.

Anthropic has boasted that its popular coding tool and chatbot Claude is capable of working on its own for seven hours without a break, essentially a full corporate workday. Meta's Mark Zuckerberg has said his company is in the middle of the year when "AI starts to dramatically change the way that we work"
and that such tools let far fewer employees do more than they ever could have before.

While Meta has since laid off 1 in 10 of its employees, Anthropic's chief executive Dario Amodei has warned that as AI tools inevitably become more productive, it could mean even broader job losses.

Despite these claims, workers inside these same tech companies, who are not only developing but using the very AI tools that will purportedly perform at least some of people's work, say they are clocking in far more than the typical five-day, 40-hour work week.

'Sprints'

US tech workers are typically well-compensated, and while other industries such as investment banking, law and medicine also often see people work long hours, a 14-hour workday was not always the norm in tech. For years, it was a more typical 9-5 office job.

But the former OpenAI employee said they would put in at least 70 hours a week, much more than they did in previous tech jobs. The person now works at a start-up also focused on AI, and said their work-life balance has improved, working closer to 50-60 hours a week, "outside of sprints".

A "sprint" is term for a common period in technology companies in which people work long hours in the weeks leading up to a release of a new feature or product.

Inside AI companies and larger tech companies working furiously on AI projects, however, such sprints can be extreme.

At OpenAI and Anthropic, for instance, sprints can stretch on for many weeks and top 90 hours of work in a seven-day period, tech workers that the BBC spoke to for this story said.

Neither company responded to BBC requests for comment.

 

https://www.bbc.com/news/articles/cvgx4yd1gl2o
Tech leaders say AI means less work - their staff say they work up to 90 hours a week

For years now, executives at companies that are pouring hundreds of billions of dollars a year into developing various artificial intelligence tools have insisted that the technology will ultimately mean people will spend less of their time working.

An engineering director at Google said four years ago that AI would deliver a four-day work week by 2025.

Earlier this year, and just one year after that engineering director's prediction, OpenAI took up the challenge, in a manner of speaking. It formally urged companies to start testing out a four-day work week (with no change in pay), claiming that AI will soon be able to speed up so much human labour that the
corporate world should prepare itself.

However, a former OpenAI technical employee who left the company last year told the BBC the firm never actually trialled the four-day work week it suggested others should try while they were there.

Instead, the person described what was often a gruelling work culture marked by frequent "crisis meetings", working on weekends, and "super cut-throat" performance reviews that would see colleagues suddenly let go.

"You go in on Saturday or Sunday just to catch up or make sure things aren't broken," the person said.

Other companies have also pushed the idea that AI will effectively reduce the number of hours people need to work, for better or worse.

Anthropic has boasted that its popular coding tool and chatbot Claude is capable of working on its own for seven hours without a break, essentially a full corporate workday. Meta's Mark Zuckerberg has said his company is in the middle of the year when "AI starts to dramatically change the way that we work"
and that such tools let far fewer employees do more than they ever could have before.

While Meta has since laid off 1 in 10 of its employees, Anthropic's chief executive Dario Amodei has warned that as AI tools inevitably become more productive, it could mean even broader job losses.

Despite these claims, workers inside these same tech companies, who are not only developing but using the very AI tools that will purportedly perform at least some of people's work, say they are clocking in far more than the typical five-day, 40-hour work week.

'Sprints'

US tech workers are typically well-compensated, and while other industries such as investment banking, law and medicine also often see people work long hours, a 14-hour workday was not always the norm in tech. For years, it was a more typical 9-5 office job.

But the former OpenAI employee said they would put in at least 70 hours a week, much more than they did in previous tech jobs. The person now works at a start-up also focused on AI, and said their work-life balance has improved, working closer to 50-60 hours a week, "outside of sprints".

A "sprint" is term for a common period in technology companies in which people work long hours in the weeks leading up to a release of a new feature or product.

Inside AI companies and larger tech companies working furiously on AI projects, however, such sprints can be extreme.

At OpenAI and Anthropic, for instance, sprints can stretch on for many weeks and top 90 hours of work in a seven-day period, tech workers that the BBC spoke to for this story said.

Neither company responded to BBC requests for comment.

Even smart people sometimes get caught in the trap of zero-sum thinking.

Predicting that AI will lighten workloads was foolish. If it -- as intended -- made working overall more efficient (i.e. more work done with less human labor) -- the result would not be less work to do, but more work to be done!

Efficiencies increase output. They do not reduce input!

 

Do we all?

You should tell us so we can confirm your assertion.

I'm not at liberty do that, it risks violating the "no politics" rule that this forum strives to adhere to, if you really don't know then I can't help you.

 

​