GPT-3, OpenAI's Language Generator, Writes Eloquent Essay in Defense of AI

a reply to: ChaoticOrder

Here's another article about GPT3 which is pretty amazing. Make sure you read to the end to see why.

I started getting a suspicion about a third of the way through.

a reply to: zosimov

HOLMES IV
That's where we're headed.
I should advise the Space Force, Congress, and U.N.
that when Luna is colonized,
DO NOT under any circumstances use it as a penal colony.

We have been warned:
The Moon is a Harsh Mistress

PS
You know how some people have uncontrollable urges to buy Catcher in the Rye (Conspiracy Theory), well, I have been known to buy copies of Heinlein's book, at least 5 times.

And I don't know why!!!!

originally posted by: ChaoticOrder
a reply to: pthena

Creative writers will now have to get AI ghost writers to do their work, if they ever want to win again.

Just saw this thread, that short story by GPT2 is fairly impressive but it took me around a dozen attempts before GPT2 generated something that coherent and original. It's still relatively easy to tell when something is written by AI but it gets much harder when human editors fix all the mistakes. Even in this essay by GPT3 you can see there are a few mistakes and contradictions. At one point it says "Robots are just like us. They are made in our image." as if it were a human talking. However the reasoning and comprehension capacity of GPT3 is very impressive and approaching human levels.

GPT-3 has ingested a huge amount of human written text, that's where its 'smarts' come from, but underneath it's just correlations. It's replicating some version of what humans have written---yes of course humans do the same thing most of the time, but there is a difference. Humans can make more coherent arguments, having been trained on far less data than GPT-3, which has read essentially all of the machine-readable text available to download.

originally posted by: mbkennel

GPT-3 has ingested a huge amount of human written text, that's where its 'smarts' come from, but underneath it's just correlations. It's replicating some version of what humans have written---yes of course humans do the same thing most of the time, but there is a difference. Humans can make more coherent arguments, having been trained on far less data than GPT-3, which has read essentially all of the machine-readable text available to download.

Absolutely. Deep-faking a few paragraphs of text is one thing, but putting hype, inaccurate headlines and massive computational cost aside, GPT-3, even in its most recent iteration (and similar systems like Deepmind's Gopher) is essentially stupid. A few tricks are implemented to mask some of the more obvious shortcomings, but a superficial and short-term conversational engagement immediately reveals that nothing approaching comprehension is taking place.

I revived this topic to modify the opinion I expressed above.

I've been viewing some of the YouTube videos of Dr Alan D. Thompson conversing with GPT-3 (via an avatar called Leta) and if these are, as claimed, unedited replies, the technology has come further than I realised. Whether we call it 'understanding' is a matter of semantics when it can sustain meaningful, coherent arguments which are genuinely informative.

From what I've watched so far, it doesn't stray far from the chatbot formula, albeit providing answers more precise, relevant and accurate than previous chatbots. It would be more impressive to hear it sustain an argument on a specific question over several minutes without repetition and by building upon and deconstructing points previously raised.

Impossible questions:



Ungoogleable questions:



Towards the end of this one, when asked what's changed since the start of the conversation, Leta says that Dr Thompson has become smarter by chatting to her - a good reply. However, it highlights a major limitation of the program as it stands. Leta does not store and process her conversation with the Dr., so she gains nothing from it. She cannot reference it in future conversations, nor learn from the positive and negative responses her answers elicited.