2023: The best AIs score in the top 10% of law, medical, engineering exams, and outperform licensed professionals on real-world tasks.
Please stop speculating. You don't know what we are (or aren't) "close" to.
We know how to improve an AI
Beginning of 2023: "GPT4 appears to demonstrate the ability to reason within certain constrained problem spaces."
There is no human being on this planet that actually understands how general intelligence fundamentally works. It's incredible to see people so confidently setting concrete timelines on its rate of development.
> As far as I'm aware they were still hallucinating pretty hard and incredibly biased and overly agreeable.
Certainly. But those aren't blockers to displaying the ability to reason within limited contexts. There are hundreds of thousands of human beings that fit that description.
The reason I'm bringing this up is that these issues, in humans, are solved by redundency (i.e. hiring someone else to look over their work).
As long as errors persist and, more importantly, are impossible/hard to be warned of, competent humans will need to oversee and validate every word they type, making their "advantage" much more nuanced.
It kind of turns into having a friend that can google pretty well answer your questions.
I'm not an expert, feel free to show me wrong - I'm reading the paper now and expecting it to be at least a touch troubling :)
I'm on my phone at the moment. But maybe someone else can link it.
"Competent" humans make errors too. This is a non-argument.
The author argues that with current tech we'll eventually hit a bottle neck providing additional training data, which will limit any rapid self improvement.
It's equally as speculative to assume we'll magically manage to overcome this challenge.
Like with the crypto hype cycle, people repeated all kinds of unfounded claims, but they themselves didn't come up with the idea. That the idea existed was all the "evidence" they needed for these things to be apparently true.
[+] And if it does good luck collecting on that ;)
If someone was to begin arguing with you that it was going to end tomorrow and started raining you with blog posts and papers, what's the sensible thing to do?
Just ignore em
We can learn really fast with limited data. There are no laws of physics that are preventing us to do the same with an artificial LLM, or the next platform.
There are AI research efforts that are seeking to directly map and mimic the structure of the brain; however, they are not at a place where they can remotely demonstrate practical results. To the best of my understanding, it would take several further orders of magnitude more processing power to simulate a full human brain at anything close to speed.
It is not at all impossible that we will be able to do this someday. But, as TFA says, we are not close.
Why do you believe that this is relevant? Our brain is just the product of natural evolution; running on hardware that is already incredibly inferior in several critical aspects (processing speed, bandwidth and information storage density), and the only holdout--computational power/watt--exists because we have no good way to compare it.
Why would it be necessary or even helpful to COPY that biological implementation to achieve superhuman cognitive capabilities?
Guidelines contemplation for you.