At risk of being too pedantic even for HN, I have to say that a news article written by GPT-3 isn't "Fake News". Maybe you could called it "artificially authored news" or something, but nothing about synthesizing and regurgitating words is inherently fake. "Fake News" is a loaded term that fact-checkers tend to avoid for its ambiguity. Generally its usage refers to disinformation, which is the use of media to intentionally deceive the reader for political or social motivations.
Its terrifying to imagine artificially authored disinformation, but from my sparse understanding of GPT-3, it wouldn't be the right tool for crafting novel disinformation without a lot of inputs for its user. Disinformation is dangerous when represented as truth by platforms with credibility, and no content creation tool can garner and wield credibility on its own. That said, it could certainly wreak havoc with mass commenting campaigns and such.
"GPT-3 is a marvel of engineering due to its breathtaking scale. It contains 175 billion parameters (the weights in the connections between the “neurons” or units of the network) distributed over 96 layers. It produces embeddings in a vector space with 12,288 dimensions."
I don't know much about AI, though I do know about programming, and to me this vaguely smells like "our program is so great because it has 1 million lines of code!"
Does the number of parameters, dimensions, etc, really have anything to do with how breathtaking and marvelous something like this is?
The whole schtick of GPT-3 is the insight that we do not need to come up with a better algorithm than GPT-2.
If we dramatically increase the number of parameters without changing the architecture/algorithm its capabilities will actually dramatically increase instead of reaching a plateau like it was expected by some.
"To the surprise of most (including myself), this vast increase in size did not run into diminishing or negative returns, as many expected, but the benefits of scale continued to happen as forecasted by OpenAI."
I love articles about the cutting edge of AI - because it always ends up putting the spectacular glory of nature front and center.
Any time you see an AI promo with really big numbers claiming to be close to replicating human cognitive functions, just remember that the human brain has 100 trillion synapses. So much of our discourse around AI right now is nothing more than chimpanzees scribbling on a wall with a crayon and calling it a "self portrait".
It _can_ be an analog to your million lines of code. But often times a given architecture won't scale to greater capabilities just by adding more parameters in the same pattern as before. (Oversimplifying, but) the signal from the training data gets weaker and weaker the "further away" the parameters are from it in the model. It can take actual ingenuity to figure out how to get those further layers in a network to contribute anything useful.
If nothing else, actually training an AI algorithm that large is an extremely large engineering challenge.
Googling around, it looks like most neural networks have somewhere in the neighborhood of tens of thousands of parameters. If nothing else GPT-3 is much, much bigger than most of its peers.
We can see a very strong correlation between brain size, and overall intelligence within the animal kingdom. The larger the brain, the smarter the animal.
If read this correctly, it does hold in very narrow sense, when compared within hominid evolutionary path only.
In general size of brain is proportional to body mass, more sensors bigger brain, has nothing to do with intelligent effective behavior, to large extent, arguably.
To best of my knowledge there is no known method that would estimate minimum amount of neurons even for simple problems, let alone complex one, however there is convergence of some form cerebral cortex between species, but octopi break this model to large extent (might we wrong here). Things get even more complicated when you account for space between individual neurons.
Even measuring “intelligence” is actually quite tricky, as it’s not a simple attribute to assign a metric to. Even comparing the relative intelligence of say, gorillas to chimpanzees involves a lot of nuance, and that’s before we dig into brain size.
The biggest bulkshit, to me, is people confusing pattern matching with intelligence. Sure, the model is outputing coherent text, but it has no fucking clue what it's talking about.
Well, this could be explained in a few meta-principles or just principles of a proper (non-abstract) philosophy.
1. A map is not a territory. Weighted connections are not semantic relations.
2. Environment and its laws and constraints comes first.
3. Language is a tool of describing What Is, not a tool of producing what could be.
4. Like untyped lambda calculus, applying anything to anything produce bullshit.
5. A proper use of a language require a type discipline which reflect the laws and constraints of the environment, and reject sentences with is not type-correct.
Everything else will produce a bullshit. Theoretical physics and other abstraction based fields are thus flawed.
For years we've been hearing how AI will build robots and wipe out humanity and the bollocks that is the trolley problem. I remember when I was reading the Unsupervised Cross-Domain Image Generation[1] paper and my immediate thought was "yep, I can see this going south". And sure enough, not long after deepfakes became a thing. GPT-3 is absolutely astonishing in terms of it's capabilities and I'd love to be able to dig into it's inner workings and scroll through it's code. The truth is there are three stoppers for the large majority of people who would love to exploit it.
1. Data. For better or worse obtaining a dataset that big isn't a great deal if you really want to. Gutenberg project, wiki corpus, the reddit dumps - difficult but definitely doable.
2. Costs - training the model costs $5M which is a considerable amount of money by anyone's standard(rich people will also have a second thought when they hear that number). But there is a catch - the hardware is becoming more and more accessible. Remember when a server grade GPU like P100 was ~10k a piece? And now the high end 30X series are 1/10th of that and have better specs... Adjust those numbers and you get something close to 20x price decrease in the course of 4 years(iirc p100 came out 2016).
3. Finding the people with the adequate knowledge to build something like this. This I think is the only blocker at this point. Realistically we are talking a few dozen people on earth that have the mental capacity to build something like this.
If there is one thing that I see as a potential threat in this field: information losing credibility.
Regarding point number 3, I wonder if there might be more than we think, but they are sequestered in various defense projects in different countries around the globe?
I think there are probably more than a dozen people, but they are working other projects. GPT-3 is cool, but it's not really commercially viable yet. There are lots of more immediately profitable projects to work on in the large enterprises that employ these capable people.
Surely there are more but we are talking about a very limited number: Think top 20 athletes in any given sport. And considering the Pareto distribution(which I'm absolutely convinced is equally valid in this scenario), we are most likely looking at double digit numbers.
We tried GPT-3 for our email startup (Mailscript), initially as a fancy way to detect and understand the content of an email. That didn't work out great, because it's really prone to false positives and ultimately requires more work than doing fancy Regex. We're hopeful it will solve other problems we'll run into though, but we're not going to push for it before we find the problems.
I'll share the lessons learned from the implementation if that's interesting to people here.
Maybe bullshit could wind up being useful in unexpected ways.
It could be an interesting tool to use to workshop ideas.
For example, if you were trying to work your way through an idea, you could throw various aspects of your idea at it, and it would throw back a slightly different take.
Rubber-duck debugging, but the duck talks back, and you can throw your fundamental assumptions about life at it.
Its terrifying to imagine artificially authored disinformation, but from my sparse understanding of GPT-3, it wouldn't be the right tool for crafting novel disinformation without a lot of inputs for its user. Disinformation is dangerous when represented as truth by platforms with credibility, and no content creation tool can garner and wield credibility on its own. That said, it could certainly wreak havoc with mass commenting campaigns and such.