Nobody Cares: PRS

[Cindy Claveau]

 

[Innula Zenovka]

My sweet summer child....

On the contrary, I DO think you've been spoiled by working with an unusually sensible organization. I would love to believe that is the norm, but I very much doubt it.

My cynicism is fueled by my own work experience in a corporate business consulting firm that is on the Fortune 500 list. I watched executives enthuse about the future of block-chain technology, metaverses, cybercurrency, and NFTs, all without any acknowledgement of the serious concerns that were, at the time, already quite well known. In the last few months before my retirement, AI was all the rage and AI-generated content was already being used by various departments, with only the most formulaic cautions about reviewing the result. There was no comprehensive review or governance, much less training, people were just running with it.

The projects I have in mind are a bit different, in that we're mostly considering ways to use AI chatbots for various customer-facing purposes, so everyone's very nervous about the damage an hallucinating chatbot might do. For internal purposes, people use AI at their own discretion and at their own risk. Certainly I know the designers use it to brainstorm, seeing if whatever Dall-e comes up with when given the appropriate prompts helps them come up with ideas, and I've certainly found ChatGPT at least as helpful as, and far faster than, Googling technical questions. I don't trust it implicitly, of course, so I check references and sources, just as I would for any other form of online research, but it's certainly a great time-saver.

I don't actually see the harm in using AI as a timesaver to generate simple content about a field you know well, so long, of course, as you check the results.

 

[Veritable Quandry]

The projects I have in mind are a bit different, in that we're mostly considering ways to use AI chatbots for various customer-facing purposes, so everyone's very nervous about the damage an hallucinating chatbot might do. For internal purposes, people use AI at their own discretion and at their own risk. Certainly I know the designers use it to brainstorm, seeing if whatever Dall-e comes up with when given the appropriate prompts helps them come up with ideas, and I've certainly found ChatGPT at least as helpful as, and far faster than, Googling technical questions. I don't trust it implicitly, of course, so I check references and sources, just as I would for any other form of online research, but it's certainly a great time-saver.

I don't actually see the harm in using AI as a timesaver to generate simple content about a field you know well, so long, of course, as you check the results.

A number of my students have started doing their research projects on using AI in medical fields, either to help doctors with diagnosis or to answer patient questions. The general finding is that customer facing uses are not there but where they aid a professional who interprets the results they are getting very good. These are LLMs trained on medical research and charts mostly, not ChatGPT.

 

[Innula Zenovka]

A number of my students have started doing their research projects on using AI in medical fields, either to help doctors with diagnosis or to answer patient questions. The general finding us that customer facing uses are not their but where they aid a professional who interprets the results they are getting very good. These are LLMs trained on medical research and charts mostly, not ChatGPT.

In England and Wales and also, I think, New Zealand, the judiciary have issued guidelines on the use of AI in court (I think I posted about it earlier, but I'll dig out the article if anyone's interested). In both jurisdictions, judges may use ChatGPT or similar to save time when drafting their explanations of the law to juries, so long, of course, as it's a simple area of the law with which they're familiar and that they take responsibiity for the accuracy of the summary.

I've just asked ChatGPT to provide simple summaries of a few common topics in English criminal law (e.g. the distinction between causing Grievous Bodily Harm and causing Grievous Bodily Harm with intent) and it's been perfectly accurate.

 

[Veritable Quandry]

I have an assigned reading about the New York lawyers who were sanctioned because they submitted a ChatGPT text as a filing that included three citations to cases that don't exist. I think the judge did a good response about how to responsibility use LLMs but that it is the responsibility of the attorneys to fact check anything with their name on it filed with the court, no matter if the draft is from an AI or from an intern.

 

[Innula Zenovka]

I have an assigned reading about the New York lawyers who were sanctioned because they submitted a ChatGPT text as a filing that included three citations to cases that don't exist. I think the judge did a good response about how to responsibility use LLMs but that it is the responsibility of the attorneys to fact check anything with their name on it filed with the court, no matter if the draft is from an AI or from an intern.

You and your students might be interested in this

The perils of courtroom AI

Look out for hallucinating claimants, judges are told

rozenberg.substack.com

Evernote link

Joshua Rozenberg is one of our foremost legal journalists, who has long been regarded as one of the leaders in explaining developments in English law to the general public.

 

[Argent Stonecutter]

What worries me most is that people think that AI is far more capable than it is

What worries me is that people think spicy autocomplete is AI, and when software with actual reasoning ability comes along people won't have a fucking clue that it isn't just a better parody generator until it's too late.

 

[Kamilah Hauptmann]

Since it seems more likely that 2/3 Americans base their opinion on what they've heard, or think they've heard, about AI taking everyone's jobs than that they've reached an informed conclusion based on a detailed analysis of AI's capabilities and how those might apply to their jobs, I'm not sure how much this tells us, other than that, no matter whether they're real or groundless, fears about AI taking people's jobs are likely to become a major concern in US politics.

This reads remarkably well in the voice of Sir Humphrey Appleby.

 

[Bartholomew Gallacher]

A whistleblower at Boeing explained, how quality control is actively circumvented to get faulty airplanes out, namely the 737 MAX 9. It is a nightmare to read. As consequence you probably should not book flights with Boeings.

Internal error statistics at the door in the mid section of the airplaine look like this: "this check job that should find minimal defects has in the past 365 calendar days recorded 392 nonconforming findings on 737 mid fuselage door installations."

Since Boeing has not enough work labour force, the outsorce the fixing to a contractor; of course they didn't fix the problem, but oh boy that paint job looks nice: "The Boeing QA writes another record in CMES (again, the correct venue) stating (with pictures) that Spirit has not actually reworked the discrepant rivets, they *just painted over the defects*."

...and such an airplane then goes out of the door as properly made.

Final quote: "In Boeing production speak, this is a “process failure”. For an A&P mechanic at an airline, this would be called “federal crime”."

How did they get the airplane marked as good? They tricked the production system by entering a part of the door as "replaced", for which you have to open the door while they falsely marked the door as "unopened", so there was no need in the system to inspect the door.

So in other words Boeing actively mislabels faulty airplanes, which can and will harm its passengers, as flawless.

Boeing Whistleblower: Production Line Has "Enormous Volume Of Defects" Bolts On MAX 9 Weren't Installed

A reader at respected airline industry site Leeham News offered a comment that suggests they have access to Boeing’s internal quality control systems, and shares details of what they saw regarding the Boeing 737 MAX 9 flown by Alaska Airlines that had a door plug detach inflight, causing rapid...

viewfromthewing.com

 

[Soen Eber]

This reads remarkably well in the voice of Sir Humphrey Appleby.

Bernhard's too, but curiously not Hacker's voice. Not self-serving enough, I suppose.

 

[Khamon]

What worries me most is that people think that AI is far more capable than it is, and that they will use it with decreasing amounts of supervision.

A number of my students have started doing their research projects on using AI in medical fields, either to help doctors with diagnosis or to answer patient questions. The general finding is that customer facing uses are not there but where they aid a professional who interprets the results they are getting very good. These are LLMs trained on medical research and charts mostly, not ChatGPT.

We constructed "medical expert systems" in school using Prolog on an IBM 370. We were so 7eet. The professor required backward chaining and that we include cautionary statements detailing the inherent limitations. Stats and Logic were prerequisites but the other languages (Assembly and COBOL) were not; that always seemed odd to me. The point is that people now don't bother to ask these chatbots to site data or reasoning to support their "conclusions." Businesses might though where profitability is concerned.

 

[Innula Zenovka]

Fascinating article, with links to some of his columns

Erwin James, the writer of an influential Guardian column about life in prison who would go on to be a leading voice on criminal justice, has died.

James, real name Erwin James Monahan, was convicted of murder in 1984 and served 20 years in prison. He wrote the Guardian column, A Life Inside, from 2000 until his release in 2004, and continued to write extensively for the newspaper afterwards, offering a glimpse of the reality of life behind bars and the challenge faced by those trying to build a rehabilitated life after it.

 

[Argent Stonecutter]

The point is that people now don't bother to ask these chatbots to site data or reasoning to support their "conclusions." Businesses might though where profitability is concerned.

If you ask the chatbots to cite sources they treat it like any other input and create a plausible continuation of the prompt based on the connections between tokens in the data set. They will therefore make up citations, because that's what a plausible continuation of the prompt looks like.

That's all these programs do. It's all hallucinations. If it's actually correct it's a coincidence.

 

[Innula Zenovka]

If you ask the chatbots to cite sources they treat it like any other input and create a plausible continuation of the prompt based on the connections between tokens in the data set. They will therefore make up citations, because that's what a plausible continuation of the prompt looks like.

That's all these programs do. It's all hallucinations. If it's actually correct it's a coincidence.

ChatGPT 4 and Perplexity AI both provide hyperlinked references if asked. Whenever I've bothered to check them, they've been sound. It's a matter of knowing the limitations of the particular application, I think, and always checking sources -- which people should do anyway.

They're a great time-saver if the person using them knows the field and knows how to check and assess the sources on which they rely.

 

[Veritable Quandry]

If you ask the chatbots to cite sources they treat it like any other input and create a plausible continuation of the prompt based on the connections between tokens in the data set. They will therefore make up citations, because that's what a plausible continuation of the prompt looks like.

That's all these programs do. It's all hallucinations. If it's actually correct it's a coincidence.

My main response to concerns about students using LLMs is to have them turn in a bibliography with hyperlinks and annotations before they write the assignment.

 

[Argent Stonecutter]

So basically they finally gave up using the generative system itself to answer questions about references.

 

[Bartholomew Gallacher]

Some "there are still decent people around" news: 15 citizens of Houston, Texas, should serve their jury duty in a court room. The offense was that members of the group "Food not bombs", which do feed homeless people for free, did that in Houston. Feeding the homeless in Houston is an offense and can be fined with 500$ penalty.

Too many of the potential jurors said that even if the defendant would be guilty, they are unwilling to the 500$ the city attorney was seeking. So now a new jury will be built.

https://www.houstonchronicle.com/news/houston-texas/housing/article/food-not-bombs-trial-busted-jury-panel-homeless-18617041.php

 

[detrius]

Trump’s Not The Only Republican To Face Disqualification Over Jan. 6

 

[Innula Zenovka]

This reads remarkably well in the voice of Sir Humphrey Appleby.

Bernhard's too, but curiously not Hacker's voice. Not self-serving enough, I suppose.

My father was a senior civil servant and my uncle was Professor of Politics at a major university, so I have a certain view of how things work.

 

[Free]

Holy crap, their planes are coming apart while just rolling down the taxiway.

Nose wheel comes off Delta Boeing 757 shortly before takeoff

A Boeing 757 plane lost its nose wheel shortly before taking off at Hartsfield-Jackson Atlanta International Airport, according to US aviation officials.

The Delta Air Lines plane had been preparing to fly to Bogota, Colombia, on January 20 when the “nose wheel came off and rolled down the hill,” according to a preliminary report from the Federal Aviation Administration (FAA).

Passengers were able to disembark the plane after the incident, which occurred at around 11:15 a.m., and were bused back to the terminal, the FAA has confirmed. No injuries were reported.