Anyone who says the first is lying to you. LLMs are actually incredibly useful tool in the tasks they were initially designed for like machine translation, natural sounding text-to-speech and accurate speech-to-text.
In trying to generate hype (or more rather revenue), the companies responsible for these models have been throwing LLMs into all sorts of functions they just weren’t designed for - often to haphazard results.
It’s like asking a really well-trained parrot to fact-check for you, code for you, write stories to you. It knows what these things look like, so can make really convincing copies and mash-ups that look right on first read - but it can’t iterate and make new things because it doesn’t actually know what training data it has is fact/fiction, it doesn’t know what code actually does, and it has no idea what a cohesive story is.
The problem is that executives and shareholders are only aware of what’s being hyped up about LLMs, and not of the technical limitations underneath that make them rather unreliable compared to specialised neural networks or just plain trained professionals.
So it is simultaneously robbing people of their jobs because of hype, while doing an absolutely terrible job of it because it is fundamentally limited in what it can replace.
New being unfamiliar. LLMs can’t just abstractly create things they don’t have training data for the same way a human can. They’re parrots that rely on training data to “create” anything, and that’s why I said they’re good at creating copies and mash-ups.
A good example is DALI being famously unable to depict an empty glass of wine because its training data didn’t have one.
OpenAI had to feed it training data of empty wine glasses to undo that.
That need for base data to make literally anything is the whole reason why AI companies have been scraping the ever living shit out of the internet, to give as much training data to mash-up as is possible.
The more data it has, the more convincingly unique its output can be.
Iterate was a poor word to use, but you’ll have to chalk that up to me being a fallible human.
What I mean is that it can’t extrapolate from training data to make something unique. Everything it makes you will always be cobbled together from the data it has, because LLMs only know what things look like, not what they are as concepts.
Hell you want to see AI not understanding what good code actually is - look at MicroSlop’s Windows 11, where damn near every update has a crippling bug in it that could’ve been avoided
Ask some image generator to make an image of a undarbinian alien, which are super aggressive. I made that up and Google shows no result.
Create an image of a beaver-axolotl-bird. I am sure there exists no such picture on the web. It will make it up. I could not make both of these things up or somehow imagine them. You could also ask an LLM to generate these examples in the first place. Is that not “something new”? Not unique? Define these words then. These systems being unable to do some specific detail does not mean they can’t do anything to begin with. No human can do everything, that is completely disconnected from everything we are discussing.
Also, Windows, pre LLM it was bug free and nobody complained?
Anyone who says the first is lying to you. LLMs are actually incredibly useful tool in the tasks they were initially designed for like machine translation, natural sounding text-to-speech and accurate speech-to-text.
In trying to generate hype (or more rather revenue), the companies responsible for these models have been throwing LLMs into all sorts of functions they just weren’t designed for - often to haphazard results.
It’s like asking a really well-trained parrot to fact-check for you, code for you, write stories to you. It knows what these things look like, so can make really convincing copies and mash-ups that look right on first read - but it can’t iterate and make new things because it doesn’t actually know what training data it has is fact/fiction, it doesn’t know what code actually does, and it has no idea what a cohesive story is.
The problem is that executives and shareholders are only aware of what’s being hyped up about LLMs, and not of the technical limitations underneath that make them rather unreliable compared to specialised neural networks or just plain trained professionals.
So it is simultaneously robbing people of their jobs because of hype, while doing an absolutely terrible job of it because it is fundamentally limited in what it can replace.
You say LLMs can’t make new things? Define new please, as this is so easy to disprove I can only assume you mean something extremely specific.
It can’t iterate? Have you never coded with am LLM?
New being unfamiliar. LLMs can’t just abstractly create things they don’t have training data for the same way a human can. They’re parrots that rely on training data to “create” anything, and that’s why I said they’re good at creating copies and mash-ups.
A good example is DALI being famously unable to depict an empty glass of wine because its training data didn’t have one. OpenAI had to feed it training data of empty wine glasses to undo that.
That need for base data to make literally anything is the whole reason why AI companies have been scraping the ever living shit out of the internet, to give as much training data to mash-up as is possible. The more data it has, the more convincingly unique its output can be.
Iterate was a poor word to use, but you’ll have to chalk that up to me being a fallible human. What I mean is that it can’t extrapolate from training data to make something unique. Everything it makes you will always be cobbled together from the data it has, because LLMs only know what things look like, not what they are as concepts.
Hell you want to see AI not understanding what good code actually is - look at MicroSlop’s Windows 11, where damn near every update has a crippling bug in it that could’ve been avoided
Ask some image generator to make an image of a undarbinian alien, which are super aggressive. I made that up and Google shows no result. Create an image of a beaver-axolotl-bird. I am sure there exists no such picture on the web. It will make it up. I could not make both of these things up or somehow imagine them. You could also ask an LLM to generate these examples in the first place. Is that not “something new”? Not unique? Define these words then. These systems being unable to do some specific detail does not mean they can’t do anything to begin with. No human can do everything, that is completely disconnected from everything we are discussing.
Also, Windows, pre LLM it was bug free and nobody complained?