The AI dooooooom thread

[crazy canuck]

Not to go into an old person thread - as I think on changes in the last 3-5 years, this isn't to do with ChatGPT but I have several friends who are teachers who have all said there's a huge shift in behaviour since covid. They all say that basically all their colleagues with pre-covid experience agree that something happened in behaviourally.

Sadly one is even quitting - which is a job he has loved, he's had students go on to study his subject at university and return to the school as teacher - but his experience has become that it's really grim right now.

I think it is a mix, the person who wanted to give me a written summary was going to us ChatGPT.  I agree that the reduction in attention span has additional causes, mainly related to social media use.

[Hamilcar]

I am working on replacing you all. 

[Jacob]

I am working on replacing you all. 

Presumably you're already posting on a languish populated by AI an its much better, which is why you're not posting here.

[Savonarola]

It's Goals and Objective time of year at Knorr Bremse.  Last year we put our goals in Alstomspeak (I'll deliver my KPIs with 30% greater Agility in years past by March!), there isn't really a Knorr-Bremsespeak yet, so after I submitted my goals by boss, we discussed some changes and he asked me to run them through Microsoft CoPilot.  I did so, sent the results to him, and he eliminated all the CoPilot "Improvements."  I thought that was amusing.

Some of my coworkers are using CoPilot to write their goals; and have been saying that their Goals this year sound a lot more sophisticated than last; but also their goals don't sound like something they would ever say.

[mongers]

It's Goals and Objective time of year at Knorr Bremse.  Last year we put our goals in Alstomspeak (I'll deliver my KPIs with 30% greater Agility in years past by March!), there isn't really a Knorr-Bremsespeak yet, so after I submitted my goals by boss, we discussed some changes and he asked me to run them through Microsoft CoPilot.  I did so, sent the results to him, and he eliminated all the CoPilot "Improvements."  I thought that was amusing.

Some of my coworkers are using CoPilot to write their goals; and have been saying that their Goals this year sound a lot more sophisticated than last; but also their goals don't sound like something they would ever say.

 

Live by the sword*, die by the sword*.



*the sword in question is two sided and dripping in bullshit AI.

[Josquius]

I spent some time playing with a tool called cursor. It's a "vibe coding" environment.

I've had a little project in mind for a while but never really pushed myself to do it so gave it a go with AI, avoiding any manual editing.....

And it gets to basically the right place very well and very quickly. I have something fully functional that does all I expected and looks mostly right.

It's the mostly where the issues lie. It did look better. But along the way of fixing other bugs it lost that.
Getting the small details done right is a painful process of back and forth and flipping between different wrong versions.

Still.... The stuff that needs tweaking to get it how I want it is a lot smaller and easier than the major problem solving that it solved no problem.

The AI just bloody can't handle "make the margin between those two things a bit smaller" but changing the number oneself doesn't need anything but the most basic of knowledge.

The grand scope stuff where I would have had to push myself... It did well. It did two versions even. The first used a python plug in that loaded terribly and then it suggested redoing it in another when I complained.

I'll see about sharing what was made when I get time to do the last tweaks and figure out somewhere to upload.

[Baron von Schtinkenbutt]

It's Goals and Objective time of year at Knorr Bremse.  Last year we put our goals in Alstomspeak (I'll deliver my KPIs with 30% greater Agility in years past by March!), there isn't really a Knorr-Bremsespeak yet, so after I submitted my goals by boss, we discussed some changes and he asked me to run them through Microsoft CoPilot.  I did so, sent the results to him, and he eliminated all the CoPilot "Improvements."  I thought that was amusing.

Some of my coworkers are using CoPilot to write their goals; and have been saying that their Goals this year sound a lot more sophisticated than last; but also their goals don't sound like something they would ever say.

Gods I hated that shit.   So glad I don't have to deal with it any more.  If I did, I would definitely be using an LLM to write that shit because it was painful.

I spent some time playing with a tool called cursor. It's a "vibe coding" environment.

I've had a little project in mind for a while but never really pushed myself to do it so gave it a go with AI, avoiding any manual editing.....

And it gets to basically the right place very well and very quickly. I have something fully functional that does all I expected and looks mostly right.

It's the mostly where the issues lie. It did look better. But along the way of fixing other bugs it lost that.
Getting the small details done right is a painful process of back and forth and flipping between different wrong versions.

Still.... The stuff that needs tweaking to get it how I want it is a lot smaller and easier than the major problem solving that it solved no problem.

The AI just bloody can't handle "make the margin between those two things a bit smaller" but changing the number oneself doesn't need anything but the most basic of knowledge.

The grand scope stuff where I would have had to push myself... It did well. It did two versions even. The first used a python plug in that loaded terribly and then it suggested redoing it in another when I complained.

I'll see about sharing what was made when I get time to do the last tweaks and figure out somewhere to upload.

I have been using ChatGPT in the past couple days to help me better translate raw SQL queries into Entity Framework Core (the standard C# Object Relational Manager), since Entity Framework can be quite verbose if you don't use the right patterns.  I also use it off and on for help with writing TypeScript code for the AWS CDK (Amazon's Infrastructure As Code tool for AWS).  In both cases, the official documentation is inconsistent on showing you how to properly use the tools, and in the case of AWS CDK even properly explaining what some modules of the library do.  While the model does get some details wrong, and in the case of CDK hallucinates methods that don't exist, it as least gets me pointed in the right direction.

I am not confident enough in its product to start having it write code for me.  I use JetBrains IDEs, which now have an AI assistant built in.  I have so far only used it for single-line intelligent code completion, but even in this it seems to be wrong about 40% of the time, though only slightly.

[DGuller]

I use AI to give me pointers as to how to write the code, and maybe for simple functions where I can easily inspect and understand them at a glance I'll just let it write them, but I'm hesitant to let it do the coding for me in general.  Sometimes it's going to be off in subtle ways, sometimes it's actually going to way overcomplicate it, and I'm also fearful of losing my own skills through attrition. 

The analogy I often make is that when I drive with a GPS, I drive with the map always pointing north, so that I don't get used to just being a monkey that obeys its orders to turn left or right.  Having it set up that way helps with training my ability to orient myself independently of GPS, and GPS is there just to lighten my mental load.

[DGuller]

One use that did make me very grateful is when I had to refactor the code and rename 30 data columns to match the other system.  You had to things like rename "wrkIndustry1" to "work_industry_1", and things like that.  It's not hard to do for a human by eye, but it is super tedious.  Turns out it's also not hard for AI to do.

Another use where it unintentionally impressed me was when I asked it to write a code to create a list of categorical and continuous predictors, based on whether it was a character or a numeric column.  It designated two of the numerical columns as categorical despite my instructions, and when I asked it why, it said that in real estate datasets the names of these two columns are usually categorical, with the numbers just being the indices of the categories.  It was right on both counts.  I would've figured it out eventually, but the fact that it has the ability to look past my literal instructions to understand what I was really trying to achieve was next level.

[Sheilbh]

Enjoyed reading about this Anthropic experiment
https://www.anthropic.com/research/project-vend-1

[crazy canuck]

One of my partners was in court yesterday making an application to strike a claim.  The Plaintiff was self represented and completely honest. He was not asserting any facts that were not entirely true.  The problem was the AI tools he used to research whether he had a claim, draft his claim, and draft his response to the application to strike, led him terribly astray.  Prior to the court hearing my partner tried to explain to the Plaintiff that he had no claim and that he should just walk away so that he did not have to pay costs.  But he had faith that AI had gotten it right.

During the hearing the judge gently asked him if he had used AI to draft his pleadings.  He said he had, and the judge was heard to say "well that explains it".

I think our courts are going to become flooded with unmeritorious claims as more people come to rely on AI.  I am also fearful of all the errors occurring in other fields where there is not a formal process for catching the errors AI makes.

[Razgovory]

Lawyers for the Mypillow guy, Mike Lindell, were fined yesterday for using AI to write court documents.

[crazy canuck]

Lawyers for the Mypillow guy, Mike Lindell, were fined yesterday for using AI to write court documents.

Yeah, the court did not hold back in its condemnation of the stupidity of using generative AI

In February, Judge Wang said, the lawyers filed a court brief in a defamation case brought against Mr. Lindell that contained "nearly 30 defective citations." It misquoted court cases, misrepresented principles of law and, "most egregiously," cited "cases that do not exist," she wrote.

Judge Wang said the lawyers, Christopher I. Kachouroff and Jennifer T. DeMaster, had not explained how such errors could have ended up in the filing "absent the use of generative artificial intelligence or gross carelessness by counsel."

[The Minsky Moment]

To be fair, AI hallucinated case law can be difficult to detect these days.  For example, a few days ago I was reading what I was sure was hallucinated case law.  The opinions contained long strings of nonsense and the citations were bizarre.

Upon careful examination, however, it turned out to be the actual case reports from the most recent Supreme Court term.

[crazy canuck]