When a humanoid robot fumbles, its mistakes are obvious because the physical world offers immediate feedback.
It's the difference between lying on your résumé that you're a world-class gymnast, and having to actually perform.
So the robot might be equally convincing that it is capable to clean your windows as it is capable to repair your car brakes.
You saw it clean your windows and are satisfied, and both its form and words are promising that it can repair your brakes equally well...
I’m kinda torn between “genAI powered robots will have a ground truth reality as a reference, so they will ultimately be more grounded and effective that LLMs” and “LLMs are like drunk uncle Steve with his PHD swimming in vodka, and using genAI in robots will end up as well as having drunk uncle Steve drive home”.
Guardrails on tasks it will attempt are inevitable, but I can also see that becoming a paywalled enshitification farm.
With the gymnast example, as a non-gymnast, I don’t know the difference between a high and low scoring routine on the floor or beam. If a humanoid robot did a routine and didn’t fall, I would assume all is well. I don’t know the technical details of what is required for a gymnastics competition.
This seems like the same idea as an LLM writing a paper that looks correct to someone who doesn’t already know the answer.
In a home context, this could look like the robot not practicing proper food safety or storage around someone who doesn’t know the details about that kind of thing, which is a good number of people. What it’s doing might look correct enough, and it produces food you can eat… all is well, until you get sick and don’t know why.
I’m not an expert, but I know there are specific moves with various degrees of difficulty. I believe there is a max score based on that difficulty level, and any imperfection will lower that score, such as a foot pointed or flexed the wrong way at the wrong time, taking an extra step on a landing, etc.
I know all these rules exist, but I’m not an expert where I can say someone had their foot flexed when it should have been pointed. These details would go over my head, where a humanoid robot might get a pass from me, while an actual gymnast or judge would be able to see faults.
Welcome to the world of hard tech not easy machine learning models. Capital is in short supply, it doesn't go nearly as far and you don't get wild multiples in return if you even get any.
It plays off of the "if it looks like a duck, quacks like a duck, and walks like a duck" idiom, which of course isn't foolproof and gives avenue to the kind of spectacular advertising that is fueling this hype.
My mom was lamenting car insurance quotes, so I told her to ask AI. She did, then had it do a Monte Carlo simulation for all the insurances she the AI felt she was qualified for.
It happily replied that it did 1 million monte carlo simulations and here was the result.
To this day I don't think she fundamentally groks that LLMs cannot calculate.
They're getting quite good at that now.
Can't most LLMs trivially use Python and other languages and libs and calculate?
The "API" of trainable algorithms is essentially "arbitrary bunch of data in -> thing I want out" and the magic algorithm in the middle will figure out the rest.
Because "thing I want" is given as a list of examples, you're not even required to come up with a clear definition of what it exactly is that you want. In fact, it's a major "selling point" of the field that you don't have to.
But all of that creates the illusion that machine learning / "AI" would be able to generate a robust algorithm for any correspondence, as long as you can package up a trainset with enough examples and shore up enough "compute" to do the number crunching. Predict intelligence from passport photos? Or chances of campaign success from political speeches? No problem! Economic outlook from tea leaves? Sure thing!
The setup will not even tell you if your idea just needs more tweaks or fundamentally can't work. In both cases, all you get is a less-than-ideal number in your chosen evaluation metric.
I think it is possible to avoid, though, by asking if humans can be generally good at the task in question, if working through the implied interface restrictions, and then evaluating whether the required skills can be reflected in an available training data set.
If either of those cannot be definitively answered, it’s probably not going to work.
An interesting example here is the failure of self driving vehicles based on image sensors.
My take is that most of the problems are because a significant fraction of the actual required training data is poorly represented in data that can be collected from driving experiences alone.
As in: If you want a car to be able to drive safely around humans, you need to understand a lot about what humans do and think about. - then apply that same requirement to everything else that occasionally appears in the operational environment.
To understand some traffic management strategies expressed in infrastructure, you’ll need to understand, to some degree, the goals of the traffic management strategy, aka “what were they thinking when they made this intersection?”.
It’s not all stuff you can magically gather from dashcams.
A model might potentially be able to understand those situations, but it would need a lot of highly task specific training data and it would never be clear if the training really covered all possible situations.
The other problem I see is that a lot of situations in traffic are really two-way communication, even if it is nonverbal and sometimes so implicit we don't realize it. But usually pedestrians will also try to infer what the driver is thinking whether he saw them, etc. In those situations, a self-driving car is simply a fundamentally different kind of traffic participant and pedestrians will interact with it differently than they would with a normal car. That problem is independent of machine learning and seems much harder to solve to me.
The sort of questions you're talking about are primarily popular in academia. Run some MLRs against some random dataset you found, publish a paper, maybe do a press release and sell a story to some gullible journalist. It doesn't have huge value. But generative AI isn't like that.
If you just ask an LLM to write something off the cuff, it'll be bad. But doing a lot of prep with a human author guiding it? Not Dostoyevsky level, but not pure slop.
While I have seen LLMs produce some ham-fisted attempts at manipulating the state of mind of the reader, I think that the human process is so obfuscated that it only shows up in occasional echoes and shadows in the training set.
It might be possible to develop a training set that reflected perception and internal mental state vs input using (magic brain scan technology) that could change this, but right now the emotional state of the reader is just missing from the training data.
"It's writes like us, it must think like us, and will be able to think anything we can think!"
"It's embodied like us, it must be be like us, and will be able to do anything we can do!"
Flawed thinking layered upon flawed perceptions, but get enough decision makers to buy into it and heaven and earth are moved to further it.