I believe the only reason for Commercial Orgs are looking at ChatGPT is to eventually replace all support people with some kind of AI.
Think of the classic "is 0.002 cents the same as 0.002 dollars?" call - https://verizonmath.blogspot.com/ - I just tested, and chatGPT seems to understand the difference - so maybe this would be a case where the AI is better.
edit: to be clear I'm fine with labeling if that's desired, it's the "I want to talk to a real person" that I think is silly (and oddly similar to the "I want to talk to an American" that people also say).
So anything that you don't value is silly?
I think there's a value/don't value axis but also a serious/silly axis. For food labelling as an example:
Contains gluten. Don't value personally, serious.
Contains GMO ingredients. Don't value personally, silly.
edit: more specific to the support desk example - we could label each tier by "has the authority to do $X", I think that would be far more useful.
Well, sure. Or prejudiced, anyway, because none of those things affect how they perform their jobs. But whether or not the entity on the phone is a human absolutely affects how they perform their job. I would deal with a machine very differently than a human as a result.
> we could label each tier by "has the authority to do $X", I think that would be far more useful.
That would be useful as well, but the difference between human and machine isn't really about what authorities they have.
Except now you're back to where I was, talking about outcomes.
I had a very minor issue with my ISP today - they sent me an email to return some equipment (that I never had). I used the chat widget on their site to contact them. I got (presumably) a real person and it went perfectly well; nice and pleasant, sorted out quickly.
If instead it was an AI, what's the difference? If I didn't get to the same result, or it took 10x as long, then sure I'd be ticked off. But for the same outcome? Why waste a real person's time?
I don't think we're at the point where an AI can reliably perform tier-1 support at the same level as a human, but I don't think that's too far off either. When we get there, I'll happily deal with the AI. If it doesn't work, then I expect an escalation path - the same way I do with the human support agent.
I don't see that the current state of AI is anywhere near being able to do that.
> When we get there, I'll happily deal with the AI. If it doesn't work, then I expect an escalation path - the same way I do with the human support agent.
I understand. If we ever get there, I suspect that I'll have no choice but to deal with an AI. Hopefully there will be some method by which I can just immediately escalate further and talk to a person.
Part of what I want from support is to know that my problem has been heard and understood by a person. If it's an AI, that is just more separation between me and the company I'm having a problem with. It may not be rational, but I want to at least be able to think that some person at the company actually understands my issue, the importance of it, and cares.
No matter how capable and "intelligent", no AI can give me that.
This is really interesting. I think this might be another "oops, I might be neurodivergent" moment for me, because I realize you're probably in the majority here. Can I give you a hypothetical? I'm not trying to trick you or anything, just calibrating my sense of how folks view this.
You have a minor issue with your ISP. They're overcharging you for some service, $60 when it should be $50. You've tried to resolve it through the site and can't do it self-serve. It's very clear, there's no ambiguity, some mistake on their end. Would you rather:
A) Call in, be on hold 10 minutes, deal with a pleasant/empathetic/helpful person, get it resolved after 10 more minutes.
B) Call in, no hold, deal with a robot/AI (still voice) who is all business, does not attempt to be human-like, get it resolved after 5 minutes.
But if I'm going with something like B, I'd much prefer it not be done with voice. Doing it on a website would be much preferable.
Now, "I'd like to speak to someone who I can understand" is a reasonable ask.
There are many dystopian movies that showcase this and the resulting oppression.
Seriously. A couple of years ago I hit a truck tire tread and roached my radiator 30 miles east of Denver on I25. I have State Farm auto insurance, I probably pay more than I should. Their after hours support was terrible, as if they'd never heard of someone needing a tow on an interstate highway outside of city limits and street/number/zip code addressing. "Westbound I25, east of Denver at mile marker 316" is a good location, but I had to talk myself hoarse because the idea of "Interstate Highway 25" was incomprehensible. Several reps wanted a street address, which just wasn't possible.
The towing company also basically held my car for ransom, I guess that's an experience an AI can't provide. Yet
The location "30 miles east of Denver on I-25" literally does not exist.
Perhaps you meant I-70?
And perhaps neither support humans far from Colorado nor an AI with a map would know to make that correct.
I have never seen nor heard of any "ai" system for customer service that remotely approaches a barely competent human, and those are frustrating and annoying enough.
ChatGPT4-level systems may be able to get much closer if they are provided the training set, and if they can, then great.
BUT, the absolutely need to be able to figure out when they are at the limit of their knowledge/ability, and then hand-off to a human. ChatGPT4 has been spectacularly unable to do anything even close to this; and it actually just starts fabricating bullcrap in a highly confident voice.
Corporate managers are already no good, horrible, awful, terrible at actually providing service, as it seems their only goal is to provide a cursory appearance of service and reduce their human costs. (Incidentally, this also massively misses the opportunity to gather excellent data on where their company could improve it's product/service and gain market share.) "AI" will only exacerbate the trend until several generations in the future when it is actually good enough, and it may become a competitive advantage to provide better service.
My problem hasn't been getting a barely competent human, it's been getting a barely competent human (or better) that also has the correct authority for my situation. Unfortunately, in the case of both AI or humans, unlocking the person that has both the correct competence and the correct level of authority to correct a problem is the hard part.
If it was resolvable with the automated system, I already tried those steps. If it's not, there's often a reason (for the company, and not necessarily a good one) that it's not. And that's where the authority part comes in; if the AI (or the call person) is doing the "figuring out" of the technical part of a problem but passes the authority on to someone without that knowledge, then how can the person with authority possibly provide authorization without first lending their authority to the end-user support provider?
Ultimately, the reason you can't do the things you need to do is because someone in management removed the authority of that end-user support provide to give it; that will still be true in the case of competent AI, and you'll probably have even less recourse since there is no person to advocate for you to the authority on your behalf.
We really need to stop personifying every project that has the categorical goal of AI, but not the result.
I don't see a problem with terminology getting more refined as a field matures. We used to have "cars" and "trucks." These days we have SUVs and sports cars and crossovers and who knows what else. Yes, they used to be "cars". People still find it useful to have more granular categories.
Today AI remains a catch-all, but having more narrowly scoped terms like ML, AGI, and ASI lets people talk without tripping over whether or not the rice cooker is going to turn us all into paperclips.
As soon as someone says the words, "an AI", they have entered a narrative about the thing personified. The original subject is lost to an anthropomorphized facade.
The adjective, "General" was originally meant as a classifier for quality, not category. That distinction is still present on the face of the word itself. An AGI is an AI that can compete with human intelligence. Adding "General" to the phrase, "Artificial General Intelligence" does nothing to define "Artificial" or "Intelligence".
The distinction I just made is entirely based on symbolic meaning. Words are defined explicitly, and then put together. In this approach, the resulting meaning is independent of context: it's literal.
Your distinction takes the opposite approach: an entire phrase - regardless of any symbolic definitions each word is known to have - is defined by the narrative it is surrounded with. In this approach, the resulting meaning is dependent on the context: it's literary.
Large Language Models are famous for demonstrating the second approach: they implicitly model whatever patterns exist in a prompt's text, so that they can generate a "continuation" that follows those patterns, without ever needing a literal symbolic definition.
This is exciting, because natural language - as the result of humans taking both literal and literary approaches - is ambiguous. The literal approach that traditional programming language parsers use is fundamentally limited to "context-free grammars", because ambiguity can only be resolved with literary context.
The problem is that LLMs cannot be literal. They are only capable of inference, so they lose all of the functionally that symbolic definition would provide.
The distinction I made earlier has the same limitation that parsers have: it ignores the surrounding narrative context that has evolved away from the original meaning: it's short-sighted.
The distinction you made has the same limitation that LLMs have: it ignores the original symbolic meaning that is still present in each word: it's confidently imprecise.
What matters is the result:
By ignoring the literary context, the definition of "AI" is written in stone, and we lose the ability to talk around it.
By ignoring the literal meaning, we imply "AI" tech is (by definition) capable of both literal and literary transformations.
We can resolve both limitations by simply using new words. There is no need to say "AI" when "NLP" is more precise.
I don’t see a problem with the term AI, it’s both artificial and intelligent
Historically, as soon as the system actually works for some task, it gets a new name. Thus expert systems, neural networks, logistic regression were all supposedly AI at one time and now are distinct. Already, we see a bit of a divergence in nomenclature where people distinguish large language models from other classes of techniques.
Of course, up to now, it has mostly been techies who have used, abused and redefined the term. More recently, marketing types have adopted the work AI so it may lose all meaning soon.
Support work is far from the only kind of human labor commercial orgs would like to automate if possible, and far from the only kind that AI could automate.
you want to make the call even longer than necessary to provide me with information i can do nothing about. excellent.