McDonald’s is removing artificial intelligence (AI) powered ordering technology from its drive-through restaurants in the US, after customers shared its comical mishaps online.
A trial of the system, which was developed by IBM and uses voice recognition software to process orders, was announced in 2019.
It has not proved entirely reliable, however, resulting in viral videos of bizarre misinterpreted orders ranging from bacon-topped ice cream to hundreds of dollars’ worth of chicken nuggets.
McDonald’s told franchisees it would remove the tech from the more than 100 restaurants it has been testing it in by the end of July, as first reported by trade publication Restaurant Business.
You do realize LLMs couldn’t even reason in 2024 right? Using sometimg that happened multiple eras ago (at the speed this is going) to push a narrative is a bit disingenuous.
They still can’t reason, because that’s not how they work.
It’s glorified auto-complete that cannot tell the difference between training data and input commands.
Albuquerque New Mexico.
Stop drinking the kool-aid my man, just because the AI companies label things as “thinking” and “reasoning” doesn’t it’s the same thing as those terms when applied to humans. There is an argument to be made the analogy works well enough with some terms, but especially with reasoning it’s pure marketing BS.
Internally reasoning is called “internal recursion” and it’s more of a scratch pad. It allows the model to split up tasks into smaller bits, where it would choke if it tried to tackle the whole thing at once. It goes through the question, tries to come up with results and writes those down in its scratch pad. That whole scratch pad including everything that came before (system prompt, prompt, context etc) gets put into the model again to produce the final result.
Not to be confused with Chain-of-Thought, which is similar but different. That’s where multiple passes get made and the tasks are actively broken up with different system prompts. This especially helps in tasks where the context would run out, but doing the same thing in smaller steps does work.
It’s very annoying how AI companies heavily anthropomorphize everything related to LLMs, right down to calling it AI instead of just an LLM and leave the intelligence out of it. This easily dupes and confuses people into thinking the tool is something it’s not. People think because they use terms like reasoning and thinking that these models have motivations, internal dialog, actual thoughts and even do things like planning, anticipating and predicting. This simply isn’t the case, it’s a mathematical model, you input numbers and it spits out numbers. When you input the same numbers, the same numbers come out (although parts can apply randomization, usually controlled through a parameter called temperature). It isn’t alive, it isn’t thinking or reasoning or smart. It doesn’t have motivations or evil intentions. It’s just numbers.
Well I’ll be damned, someone who actually understand how LLMs work, on this platform? That’s a first.
I wasn’t talking about the marketing, I read the papers.
The thing is, we’re pretty much doing the same when we think. There are plenty of convincing papers and theories that tie advanced intelligence to language. I don’t understand why everyone assumes that “predicts next tokens means it can’t be intelligent” when in reality humans do just that when they talk and think.
If you’re ever learned a second language through immersion you’ll know the feeling of stringing a sentence together just from having seen enough similar ones, even if your sentence is technically brand new. That word right there feels just right, even if you don’t know the logic why. How is that any different from an LLM?
Feelings are not facts, and LLMs are not thinking.
otherwise this wouldn’t work.
https://www.youtube.com/watch?v=lk3jCuITwcE
or this.
https://www.youtube.com/watch?v=fhi6JU5pOJk
I’m not sure how that’s related to what I said. Please bring an argument or something.
As for your videos, this is obviously using the dumb models, they wouldn’t have the budget to use decent ones, so not sure what’s your point here either.
The thing is, we have literally no idea what a thought even is nor how they work. You cannot replicate a concept you have no working understanding of. Your brain does not work the same way a computer does, point blank.
Well, we know two things for certain:
LLMs and neural networks as a primitive are trained like an optimised version of a human. Because we do know how we learn: positive and negative reinforcement. That’s seemingly how all animals learn, though there is supposedly some genetics at play too.
Verbalizing our thoughts (eg. rubberduck technique) improves reasoning
This is what LLMs are built upon. Specifically because language makes reasoning much easier, LLMs are the best current form of AI we have at it.
Besides, saying “we have literally no idea” is just straight up not true, just because you don’t know doesn’t mean nobody does. There have been a lot of papers on the subject lately, though it’s true we don’t fully understand it.
What we can easily figure out from black box analysis of our brain is this: when making a decision or when thinking, there is the initial input that causes us to have that thought, context, memory and previous thought progress (or CoT to keep things simple) bring used.
LLMs replicates all this, and given the same input, produces the same output as a human at an ever-increasing rate, these days often better in specific and limited fields. What does it matter if the internal algorithm isn’t the exact same?
Besides, you do know you can run an AI on human brain cells, right? Saying “your brain doesn’t run the same way a computer does” when we can literally put a computer in a brain…
“So, you see, our brains take information, send it through the black box, and then things come out. LLMs also take information, send it through the black box, and the. Things come out. We can say for certain that the processes are nearly identical.” - you
That black box is the entire process. We have no idea what happens inside the black box for either LLMs nor the human brain, nor do we have the tools or understanding to even know how to find those answers. We have no idea what the working mechanics of thought are let alone what it would look like to replicate the process. LLMs are just a .EXE. They have no intelligence nor thought. You are a fantasist.
Also, intergrading a board to interpret electical signals from cells as input for a computer is not turning a brain into a computer, at best it’s a novelty, at worst (if it works according to your interpretation) it’s slavery.
I don’t know, this entire comment is a rant on something I did not say.
The “board to interpret the signals” is nothing, it’s the fact that the neuronea themselves are trained the same way LLMs are trained that’s interesting. I think you’re under the delusion that the board does much here. It’s literally the GUI to the linux kernel.
Anyway, seeing as you’re not willing to have a good faith argument, I’ll just stop here.
Exactly, there are a limited number of electrical signals coming from the cells, they program the GUI to interpret each signal for a specific purpose and introduce a specific signal to get it to interact with the GUI in a preprogrammed way. There is no sentience just because it’s neurological tissue, you can make it do the same things with citrus fruits.
Do you think that using a nintendo emulator means that your computer is literally a nintendo?
But see that’s what you fail to understand. You can’t do that with a citrus because a citrus has no neurons. They don’t just interpret signals, they “code” them through RL training like they do with LLMs. In other words, it requires and is based on the brain’s neuroplasticity, aka its ability to respond to RL training.
To your argument “there is no sentience just because it’s neurological tissue”, I agree. However this is the very root of the idea that if the physical medium doesn’t matter, than why can’t AI running on silicon be sentient? The current era probably isn’t, but how will we know when it will be?
I’m sure you’re not as arrogant as to believe that out there in the universe the human brain is the only possible way to advanced intelligence. Why then is it so hard to believe that you don’t have to copy the brain 1:1 to create an artificial one? We’ve already copied such a big part anyway.