LLMs will not get us to AGI because you can’t prompt engineer and autocomplete your way to Data from Star Trek or C3PO.
This is fairly self evident but leads to multiple paths of false thinking.
1. LLMs are a scam or overhyped: We can and are getting a lot of value from this technology even if it’s not artificial general intelligence.
2. We can’t achieve AGI: LLMs are just one of many AI approaches and simply the most popular today. OpenAI already has smarter models using different approaches
What if this supposed value is just more hype? (it is).
It doesn't matter what they do, it will be useless, because the way we are going we won't have arable land in 2070.
@j_bertolotti @carnage4life The main value so far is as wage pressure against human workers.
For sure, this creates a lot of value for shareholders, but it is not actually valuable.
@j_bertolotti @androcat @carnage4life
I agree. It's fine to make a deontological or consequentialist argument against LLMs more broadly. They are littered with ethical issues. But I get daily value from them in my work. To pretend they don't facilitate a particular subset of tasks is silly.
There are better tools for those tasks and they have been around for decades.
Because a tool that was made for that task will always outperform an ad hoc application of stolen snippets, provided without understanding or intelligence.
Just go with a dumb script.
@androcat @j_bertolotti @carnage4life
Simple example: I want to speak at SQLBits. They tend to have somewhat verbose speaker instructions that change from year to year. https://sqlbits.com/speak/
I pasted the text and asked for a summary. ChatGPT identified that 20 minute sessions are now limited, which changes my strategy of what I will submit.
https://chatgpt.com/share/6772a836-c294-8012-a806-bf232df0beff
I looked myself to verify. I still intend to read it manually, but this reduced the risk of human error on my part.
What decades old tool would you recommend to accomplish this task?
@Sqlgene @j_bertolotti @carnage4life
Just read the damn thing.
You're describing a situation where the only "benefit" would be if you go with the untrustworthy summary.
You still have to read it to actually know what it says.
And now, rather than reducing your risk of error, it has infected you with the risk of confirmation bias.
Utterly worthless.
@androcat @j_bertolotti @carnage4life 🤷♂️. I think LLMs are best treated an unreliable drunken intern. There are a number of tasks I'd happily hand off to such an intern for a first pass, understanding that they very easily could be mistaken or fabricating.
I think we fundamentally disagree on things on a deep philosophical level and I don't see any value for either of us in engaging further. I think your viewpoint is reasonable and understandable, I just don't see us reaching any common ground.
Be serious. If an intern showed up drunk, you'd throw them out.
Don't treat a world-destroying tech toy better than you would a human.
@androcat
You are being overly aggressive with people who agree with you 90% of the way. That is seldom a wise strategy.
You're probably right, but I am distressed.
These are fucked up times, and this particular hype-wave is just making things worse all over.
Those tasks are not hard to do, just because people don't like doing them.
Doing those tasks for people provides no value, and instead impoverishes the people so "convenienced".
But seriously, spending a small city's worth of power to do something that you could just use a template for? In what world is that "valuable"?
It isn't.
The whole industry is rotten.
@carnage4life 1. LLMs *are* overhyped and *are* a scam. They are not economically or environmentally viable, *and* the overfitting from being large makes them *worse* than smaller models for things that might actually be viable applications.
2. OpenAI is lying. They do not have any smarter models. They are not trying. All their people are high on their own supply and the only point is extracting as much value as possible from everything they ruin along the way.
@carnage4life Can we achieve AGI? Absolutely. The existence of human intelligence and ability of computers to simulate physical systems is a (non-efficient but constructive) existence proof.
Is any of the research headed in this direction? Absolutely not. Because capitalism ensures all the money goes to whatever is most attractive for scamming people. Artificial beings that would be entitled rights are completely unattractive to capitalism. And it's not happening without giant funding.
@carnage4life In 1952 Stanley Miller and Harold Urey conducted an experiment in which they simulated the primitive atmosphere and electrical storms in an isolated capsule, causing the spontaneous generation of aminoacids. This proved that it was possible for inorganic minerals to produce organic materials spontaneously.
In the years after, chemists predicted they would be able to synthesize DNA, livers for transplant and even artificial life. None of that happened.
This is where LLMs are now.
@carnage4life yes, the limitations of LLMs do not prove that AGI is impossible. But the burden of proof goes the other way around. LLMs are being hyped as an argument that an AGI breakthrough is just around the corner. THIS is what is driving hype. Nobody wants to pay millions for a 40% medical test score, the unwritten expectation is that owners of this tech will soon get 90%-scoring new models that let them fire the human physicians.
Like the Miller-Urey livers, there will be no such thing.
But, have we achieved "API": Artificial Partial Intelligence?
"Partial" as in: does well in some niche silo.
Bill de Worde
Dare Obasanjo
Androcat
j_bertolotti
Eugene Meidinger
FeralRobots
Cassandrich
Andrew France
Cogito ergo mecagoendios
Daniel Costa
DanSmithCreates