You think?

AI tools like this should really be viewed as a calculator. Helpful for speeding up analysis, but you still require an expert to sign off.

Very much so. As a nurse the AI components I like are things that bring my attention to critical results (and combinations of results) faster. So if my tech gets vitals and the blood pressure is low and the heart rate is high and they're running a temperature, I want it to call both me and the rapid response nurse right away and we can all sort out whether it's sepsis or not when we get to the room together. I DON'T want it to be making decisions for me. I just want some extra heads up here and there.

Ideally, yeah - people would review and decide first, then check if the AI opinion confers.

We all know that's just not how things go in a professional setting.

Anyone, including me, is just going to skip to the end and see what the AI says, and consider whether it's reasonable. Then spend the alotted time goofing off.

Obviously this is not how things ought to be, but it's how things have been every time some new tech improves productivity.

I mean... duh? The purpose of an LLM is to map words to meanings... to derive what a human intends from what they say. That's it. That's all.

It's not a logic tool or a fact regurgitator. It's a context interpretation engine.

The real flaw is that people expect that because it can sometimes (more than past attempts) understand what you mean, it is capable of reasoning.

plenty of people can't reason either. the current state of AI is closer to us than we'd like to admit.

And if you think your little 401k makes you a meaningful shareholder

"In this company we're all like family, you don't have to worry about anything."

As far as I know, the Deepmind paper was actually a challenge of the OpenAI paper, suggesting that models are undertrained and underperform while using too much compute due to this. They tested a model with 70B params and were able to outperform much larger models while using less compute by introducing more training. I don't think there can be any general conclusion about some hard ceiling for LLM performance drawn from this.

However, this does not change the fact that there are areas (ones that rely on correctness) that simply cannot be replaced by this kind of model, and it is a foolish pursuit.

That's probably it's primary function. That and maximizing profits through charging flex pricing based on who's the biggest sucker.

We also have hoverboards. Well, "hoverboards", because that's the branding. They have wheels, and don't hover.

Yep, a great summary.

I keep telling people that what they call AI (e.g. LLMs) are fancy autocomplete. Little more.

Well it seems like a pretty natural fallacy to think that if something talks to us, in a language that we understand, that it must be intelligent. But it also doesn't help that LLMs, aka. fancy text generators built with machine learning algorithms, are marketed as artificial intelligence.

The LLMs can also be EXTREMELY useful, if used correctly.

Instead of replacing customer service workers, use the speech processing to highlight keywords on the service workers PC, so they can quickly find the right internal wiki page? Atlassian Intelligence works pretty neat in that way, a Help desk ticket already has some keywords highlighted and when you click on it, it shows an AI summary of what this means from resources in the Atlassian account. Helps inexperienced people to quickly get up to speed and it's only helping, not replacing.

This AI just needs to pull itself up by its bootstraps so it can move up from working fast food

I mean I don't know how it is where you live but here taking the orders has been 99% supplanted by touch screens (without AI) So yeah, a machine can do that job.

Yeah, all the control systems are in-band, making them impossible to control. Users can just modify them as part of the normal conversation. It's like they didn't learn anything from phone phreaking.

I have to had so many conversations with people still thinking fast food is only for high school kids. It's odd. If I say how will they be open during school hours, they make up some bullshit 'get a better job.' It doesn't make snese. Most of these people don't have good jobs and are lucky to be supported in their current lifestyle. They don't see that though.

I try to push the point of 'they are paying for your time and for you to be on standby.' you don't need to be actively moving all 8 hours. Your bosses don't. I've seen so many waste of time meetings to justify their welfare jobs. It's comical. They don't produce value. They are leeches. Not all, but too many.

Can we also talk about how much everyone, everywhere relies on service industry workers and how much everyone would absolutely lose their goddamn minds if they had to make their own burgers and fries twice a week, AND how these staple institutions, jobs we deemed so important that we made people work at them during a pandemic, how much the prices of these sandwiches and snacks has gone up in the last few years, how even bringing up the possibility of increasing minimum wage for these difficult and demanding jobs leads to an entire social "discourse" and fierce debates about if people should be able to afford things.

I don’t think it’s shitting on fast food jobs at all. The point of this is that taking orders at a fast food is, in the micro, an extremely easy task. What makes the job as a whole exhausting is the fact that you have to do that for a full shift and the human brain gets stressed from doing that. But AI doesn’t, and yet it’s messing up the simplest part of the job.

“It’s gonna take everyone’s jobs!” though

Cheery woman’s voice- “Hi will you be using your mobile app to check in today”

Me- “no thank you”

Voice of chain smoking grizzled dude who is tired of this- “Go ahead and order”

McDonald's did not factor in the same thing you are apparently not factoring in- when humans at McDonald's fuck up your order, you can tell them about it.

Maybe if they hired enough people.

Or in this case, replacing humans being paid starvation wages

You asked a search engine for information from the internet in its early stages and got Frankensteined results from the Internet. That was its purpose was it not? Obviously the more info it scrapes the more info it will go off of.. but yeah it is still just scraped data, not even particularly from reliable sources. The language models job is to make sentences out of information it has. It doesn't do anything intelligent to disect the information.

A good example is many AI (programs) can draw you a leopard. If you ask the program to then draw an arrow to its tail... It doesn't know what a tail is so it will draw a random arrow that would point to where a tail would be on a stock image, not even the one it "drew"

What it knows "this entity is a leopard". We want it to know much more than that. We want the program to see "draw a leopard" and then it draw: 2 eyes: run the eye operation - which then says draw shapes, for items like retinas, colors, blood vessels etc and document all of that data while creating each hair and skin blotch. Then be able to comprehend what all of the items are and do, so when asked a question or given a task it can perform such actions..

But we haven't programmed it to do so, and thus it can't, because it doesn't think to learn, it just aggregates data and searches it. It can't compile the data and recycle itself back to the original operation, as that would be pretty intelligent.

Well, the LLMs got a lot better since the first release. I would guess that the main problem with this AI (probably not one of the bleeding edge LLMs, guessing from the timeline) is that they have piss poor micros - even humans have problems getting your order right.

Machine Learning is awesome for medicine, when they run your genetic sequence and then say "we should check for this weird genetic illness that very few people have because it's likely you'll have it" that comes from Machine Learning algorithms finding patterns in the old patient data we feed it.

Machine Learning is great for finding discrepancies in big data sets, like statistics of illnesses.

Machine Learning (AI) is incapable of making good decisions based on that statistical analysis though, which is why it's still a horrible idea to totally automate medicine.

It was an umbrella term for stuff we didn't have yet..then marketing teams remembered it existed. So now it's either a nonsense term (like a PID is AI now) or it means large language models. Since that is clearly both what mcd's used and what the follow up message is referring to, I don't think you need to gatekeep this message. Take that cape off, hero of AI-term-correctness. Flip the dictionary-signal off and turn in to bed.

I work adjacent to a group that does speech recognition. There's a massive amount of variation in regional dialects and that's before you get to non-native speakers. The you have people like my mother in law who doesn't have an accent, but her diction and grammar are... unique.

If someone is speaking in sentences you can use context clues to infer intent, but it's a lot more challenging when you're just getting spoken commands.

I suspect it's a training/sample gap, but it's likely going to be really hard to get to 100%.

If I'm understanding your question correctly, heres an example model.

Obligatory scene from Hot Fuzz

I was going to say, I went through a drive through, had the clearest exchange ever when placing my order, and it was only after pulling away and hearing the same voice at the same time that it was an LLM.

I'd personally take that over trying playing telephone on shitty mics and speakers any day. "Sorry, could you repeat that?" Etc etc.

Assuming you taught them how to enter orders into the till (the AI was "trained" on how to input orders, let's compare apples to apples here) no, they wouldn't fuck it up. They would be slower than a regular employee but they wouldn't fuck up what people wanted.

Oh, and if they weren't sure for some reason they would ask somebody for help instead of making shit up.

Who the fuck drinks orange juice with a Big Mac and fries!?

Or worse, its a burger. Who’s eating my cancer medicine?!??

They'd likely use different models, but they'd be based on the same fundamental technology of LLMs. The training data would be different, but they'd have similar issues, but the one that's supposed to be in a classroom would have significantly more signals to interpret and significantly higher consequences for being wrong.

They'd be different, but if it doesn't work for McDonald's, it can't work for the much more complex task.

Both would be built on the same architecture.

If anybody can build a teaching AI worth its salt then every other user facing AI service company would want to copy their architecture, and would be willing to pay for it

So if a megacorporation can't get the good stuff, and nobody's even seen the good stuff, it probably doesn't even exist

I was going to post a reply supportive of you but realized this is an anti-ai grandpa group lol. I can't stand hearing people talk about topics they don't know anything about. Yall need to give your order to ChatGPT, and then ask it to repeat your order for you. If you don't know why McDonalds fails but ChatGPT succeeds you need to shut the fuck up.

A stressed, underpaid human? Yeah, that’s not an appropriate tool for those things.