Do the LLMs actually spit out entire sentences from their training data? If so, then yes - I agree that is a problem.
I was at the impression that LLMs basically only generate probabilistic next words.
Iām not debating that computer-generated content canāt have valueāit can. I get value out of AI grammar checking and use it for ideation. But what Iām primarily addressing is what I stated above: āPassing off machine-generated content as oneās own lacks integrity and obscures the writerās voice.ā
As I said, I suspect all of us will be using AI. That seems inevitable and, in many cases, desirable. But I donāt want us degenerating into impostures.
I cannot care the less if a support call is attended by an AI ā¦
You must have had better experiences than I have. Mine have been enough to āmake a preacher cuss.ā 
Well, there was the āas long as I get a prompt & satisfactory answer.ā part 
But to be fair, I usually donāt get satisfactory answers from human operators either: companies cannot automate their way out of a broken business process by using AI.
Iāll give you that. 
And, thatās assuming I can even understand the person on the line.
+1
I have no idea how many hundreds of hours I spent over the years trying to get problems solved.
Sometimes I couldnāt find the right person to answer my questions. Sometimes the problem was that neither I, nor the person with the answers, could communicate with the other.
If an AI could have provided the answer, there were times I would have paid for it myself.
Yes they can and do. Try taking some of the answers in ChatGPT and doing a google search on them in all quotes. You can usually find the original item that way. Itās especially noticeable in coding answers where often times the variable names and even typos are repeated verbatim.
2 Likes
Here is an example, and the response.
Yes, entire sentences. Sometimes from public domain works, but often from Websites. This has been fairly well covered in consumer-facing work, but itās a known issue, in part because of the training methods and corpora. Hereās a decent discussion from 2023. https://pike.psu.edu/publications/www23.pdf
1 Like
I was teaching myself Python using the book Python Crash Course, using VS Code with an extension Continue and models via Ollama. And it would output word for word examples and solutions from the text.
2 Likes
Youāre either going to work for an AI or have an AI work for you. Which would you prefer?
Thatās a nice little quip. To paraphrase H. L. Mencken, āneat, plausible, and wrong.ā
While I fully expect that non-biological intelligence is possible, LLMs on their own are not the technology to achieve so called āAGIā.
LLMs may be the front end module, doing the syntax processing, but there will need to be a semantic processor that provides actual understanding. Something LLMs are not capable of.
LLMs are tools. I, along with many others, use these tools. But they are only tools. I donāt work for any of the tools I use, nor, except in a sense different from the quip, do these tools āwork for meā. LLMs are no different.
4 Likes
Without being pendantic, or maybe is it Iām being pendantic? I think the thought here, which I totally agree with, is that AI tools (not just LLMās, but the entire category of ML/AI) will be able to replace some jobs - especially entry-level or bureacratic tasks.
Those most threatened to be replaced can protect their jobs, and continue to grow in their career instead of stagnating, if they adopt AI tools.
Whether they work for AI or AI works for them, IMHO, is only a pendantic argument about semantics, not substance.
That is definitely not the sense the quip was meant in the post I responded to. And in truth it doesnāt really fit the rest of Sethās piece which, is more in line with what we are both saying. The quip is at best misleading. It would be better phrased, āYou will either use AI or be made redundant by someone who does.ā Much more realistic, and certainly less threatening, in my opinion.
And I expect that one day there will be non-biological intelligences, some on par or even surpassing human intelligence. But it will take more than ML + LLMs to get there. And I fear the current infatuation with LLMs will only delay things.
And all the while snake oil salesmen are out in full force, sadly even in the US government, replacing people with LLMs. Mistaking mimicry for understanding.
This is of course trivially true, just as it is true of any other technology. And drives home my point that LLMs are just tools. And tools have been making jobs redundant since at least the printing press.
ā[I]tās so critical that you are clear to your customer that there will be always a human if you want,ā he said.
While chatbots are cheaper than humans, it seems theyāre just not as good, according to Siemiatkowski.
āAs cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality,ā he told Bloomberg. āReally investing in the quality of the human support is the way of the future for us.ā
It turns out Klarna is far from alone.
I believe it is a trueism that, for most of us, there has always been someone who is bigger or stronger or smarter or better looking than we are. And yet most of us have found our way in the world. Pundits predicting a dire future have missed many of the positive twists and turns of history. People are resilient and full of surprises. It is too early to give in to a future that we donāt want. It is too early to fear the apocalypse.
1 Like
Sometimes yes, sometimes no.
When I began my professional career at a Fortune 50 company, it was, if not quite dawn, the very early morning of the desktop PC era. (How early was it? My desktop PC had a 10 megabyte hard drive.) Every middle manager above a certain level had a secretary whose job it was to keep his calendar (it was always a āheā), answer his phone, take messages, word-process hand-written memos, word-process dictaphone content, open mail, etc etc etc. My department of 14 people counted four secretaries in that total.
Less than 10 years on, the department had 11 people in it, and all the secretaries had been replaced by one administrative assistant, who rarely, if ever, answered a phone or typed a memo. Why? We could all type our own memos, did everything by email anyway, and let voice mail handle our phones.
The women (and they were all women) who were secretaries worked hard to keep their skill levels with tech up-to-date, but there was no way for them to āprotect their jobs.ā With every budget cycle, their jobs would be eliminated and replaced with nothing. Most were without college degrees and had spent their careers honing skills that were no longer needed.
It was people like me who were able to protect our jobs by adopting new technology, because our jobs were of the kind for which technology was rocket fuel. Some of the more senior members of my department dismissed tools like Lotus 1-2-3 and Excel and clung to their calculators and 13 column analysis pads on the theory that using tools like Excel eroded oneās reasoning skills. (Sound familiar?) They too lost their jobsāor, more likely, got passed over for promotion or shuttled off to one of the dreaded āSpecial Projectsā positions, where the old and in the way were sent to die a quiet corporate death. (We used to call it the Special Projects Ice Floe.)
AI will open a wonderland of opportunity for some jobsāmuch as Excel and other important tech-based tools in the toolbox did for me. But there are other jobs that AI will simply eliminate and the people who currently hold them will need to transition to other work. As a society, we should do our dead level best to make that possibleāwhich includes allocating actual blood and treasure to the problem.
3 Likes
This. When I was a kid, we had a Uniroyal plant in our city. My understanding is that the factory made bias ply tires, and the cost of refitting the plant to make radials was deemed to be too expensive - so the plant shut down.
That put 1000-ish people out of work, and it was a massive hit to the local economy. Some of the people were managment and such, where skills might transfer more easily to a management position elsewhere. But most of the people were factory workers who knew how to make bias ply tires, which basically donāt exist anymore.
The local government actually started working with the local tech college to help the workers figure out how to transition into other jobs.
Thatās one factory, comprising maybe 2% of the population of a small city. AI is going to be crazier. Itās not that people arenāt resilient; itās that itās going to be a significant disruption that is (a) inevitable for the vast majority of workers that will be replaced, and (b) will likely take quite a bit of time to resolve.
2 Likes
Currently the US economy needs to create around 180,000 jobs each month for new people entering the workforce. If a large percentage of companies created in the future will use AI from the start, displaced workers may end up competing with new graduates, etc .
Iām of similar age, but worked in tech all my career (first as engineering, then as mgmt).
When I started, we didnāt have PCās, and the terminals connected to mainframes or minicomputers we used were in hot, cramped āterminal roomsā where there was always just not enough terminals for the people waiting.
As time progressed and we got PCās, those of us that self-taught ourselves word processors and spreadsheets, in addition to just code editors and other tools, wrote our own memos and documents.
This was to the consternation of department secretaries (both m and f) that were always begging to do it for us. (Now I realize they were trying to stay relevant; at the time, it was simply much more efficient to do it myself than to iterate back and forth reviewing drafts and waiting for them to be re-typed and re-printed.)
The best admins didnāt solicit typing assignments, they shadowed managers or other ops people and were promoted to asistants, project managers, and other jobs that stood the test of time.
I think thatās the key though. āThe bestā¦ā Not everybody is the best. And there need to be jobs for people in society that arenāt, or else Bad Things happen.
2 Likes