Using GPT-4 to measure the passage of time in fiction
tedunderwood.com
tedunderwood.com
That is a fascinating detail in itself. My first thought is this is reflective of life moving faster in general.
For example, travel by horse vs plane is the difference between days (or weeks) and hours. If you're writing about this travel in modern era, the trip can be summed up in a partial sentence like "After landing in (new city), ..." but you can't really gloss over a days-long trip by horse in the same way.
Basically, "event density" has gone up: more things can happen in a smaller length of time, and number of words is reflective of how many things are happening.
> The average length of time represented in 250 words of fiction had been getting steadily shorter since the early eighteenth century.
Time passed per word has been decreasing. A page used to mean a week, now it means a day. In your horse vs plane example, this makes sense if they both gloss over it. EG "we rode 3 days to Rivendell" vs "I landed in Tokyo after a long flight".
I don't know if that generalizes to all writers, but it gives the impression that many don't really care about realism like they used to.
But yeah, it could be the mismatch between history and current travel methods. Writers of the past, even when they wrote about the past, largely would have still been limited to traveling at the same speeds as distant ancestors via sail and horse. Writers of today who write about the past have no personal experience when they've only ever traveled by plane and automobile.
Isn't your interpretation backwards?
The length of time is shorter now, meaning 250 words describes a shorter duration/less happening. They were glossing over more, historically.
An excellent portrayal of the dynamic and rapid nature of modern life.
these or some of the deeply embedded standard convictions we need to get rid of. We simply don't know.
The short answer is that of cause LLMs can come up with something new: They can learn to rationalise – the simplest example is learning to add numbers.
So pedantically speaking, the quote is straight up wrong.
It takes coaxing (and some cherry-picking) to get anything creative out of LLMs -- and some people suspect newer versions are so locked-down to avoid sounding "scary" that they actively avoid saying anything too interesting or philosophical.
The line between "creative output" and "hallucinated madness" is also somewhat blurry (for humans and LLMs alike...)
I assume by LLMs you actually mean "ChatGPT and similar commercial LLM services". Raw LLMs like LLaMA and Falcon will happily spout out the craziest, sometimes super creative ideas, it's just the lobotomising of ChatGPT and the like that destroys their creativity.
While trying to iterate on a name for a product, I had it output an internal monologue alongside outputting name ideas.
With each iteration of naming I'd ask it to probe its understanding in of human creativity and list how the usefulness/realism of the internal monologue.
After a few loops the internal monologue started modeling fleeting memories about things like "that time at camp where X happened".
Eventually it was modeling fleeting thoughts like "the warmth of the cup of coffee on the desk" as a source of inspiration, and random observations like "it looks like traffic is picking up outside, getting to yoga will take a while" in the middle of a name generation task
—
I made a website predicated on a similar idea after that: https://notionsmith.ai
Technically it's a hallucination generator, but I think you described why I made it (creativity and hallucinations really are similar for both LLMs and humans)
Sadly it didn't seem to be a fun, usuable, creative madness to me (a la a human on certain drugs), but YMMV.
> This may not count as plagiarism, but it won’t produce anything new.
First, it is plagiarism by definition. Any time one presents work as their own which they did not create themselves is plagiarism .
Second, and vastly more importantly going forward, students were never handing in “new” ideas on an assignment. Using an LLM to generate a “pastiche” that is then read by the student (hopefully) before being handed in is still a win in my book. Having students ask their questions to an LLM can be an amazing tool to further learning but when students and educators are terrified of this new tool they will hide their usage of it and then it becomes a tool for copying rather than learning and plagiarism is the only consequence.
There's a bit of a motte-and-bailey situation here - the student clearly plagiarized if they passed off ChatGPT's work as their own, but this has no bearing on the more interesting question of whether ChatGPT itself "plagiarizes" from its training set or creates new knowledge by synthesizing it.
EDIT: Fixed the spelling of Motte-and-bailey
Not sure this is true. If I hand in an assignment that is written strictly by scattering chicken bones into a grid representing the most common English words and writing down the result is that plagiarism? It will definitely be non-sense, but i'm not copying anyone else's work.
chatGPT is just a more sophisticated chicken bone grid.
Well, yes. They produce random output that adhere to statistical models.
I had ChatGPT 3.5 produce a wonderful whimsical portrait interview of a couple breeding Giant Tibetian Hamsters for home protection in Norwegian.
> They can learn to rationalise – the simplest example is learning to add numbers.
Certainly not? Composite systems that leverage LLMs can do a lot of things - but AFAIU LLMs will likely never rationalize or be able to "add numbers" in the normal sense; they can count only in as much they know that 1,2,3,4 is more probably coming after "count to four" than 10,2,4.
You might be interested in this experiment - https://thegradient.pub/othello/
The question is does a model trained on move sequences in a game of othello just learn "play X(t) after seeing the sequence X(t-3), X(t-2), X(t-1)" or does it build a representation of the board and use that to choose the next move to play.
This implies we know what “rationalising” actually is, which we don’t. There’s no reason to believe that our brains don’t operate in fundamentally the same way as LLMs. There’s no reason to think that the LLM approach to reasoning is any less valid than the “normal” way, whatever that even means.
Not really. Nothing random about the output of LLMs. https://www.nature.com/articles/s41587-022-01618-2
>Certainly not? Composite systems that leverage LLMs can do a lot of things - but AFAIU LLMs will likely never rationalize or be able to "add numbers" in the normal sense; they can count only in as much they know that 1,2,3,4 is more probably coming after "count to four" than 10,2,4.
LLMs can add numbers just fine. arithmetic is one of the easiest domains to test on data it'd never have seen in training.
I can see what you are getting at, but the underlying implementation of LLMs really is reliant on non-deterministic, random sampling of the underlying model. The random sampling is weighted to favor certain selections over others, but rare selections are possible.
Its intrinsically built on a degree of randomness. Though it is also unfair to say something like "produce random output" as the probabilities themselves make it less than fully random.
So given the right training data and algorithm, they should be able to learn to add numbers. Or any other computable function.
Also, we don’t know how GPT computes numbers really. I guess no one traced all the layers and understood the activations.
We also don’t do this with our children. We just assume that they understand it as soon as they don’t make many mistakes anymore.
Also, this theorem is almost useless in practice: it only tells you that, for any function and desired error rate, there exists some NN which would approximate that function within that error rate. But there is no proof that there is some way to train the weights for that NN based on any known training mechanism, even if we magically knew the shape of that NN (which we don't). Obviously, since we don't know if an algorithm to train the NN even exists, we have even less idea of how long it might take, or how large the training set would have to be.
So this theorem doesn't really help in any way answer whether our current AI techniques (of which training is a fundamental component) could be used to approximate human reason.
For a given error I guess. So it's a discussion about how good we can approximate such a function.
I do agree that it's fairly noisy though.
What does it mean to learn to add numbers? LLMs in general are not good at accurately portraying their own processes, and humans are equally bad at interpreting what LLMs do “under the hood”.
“To rationalize” means something like “to provide a reason (a rationale)” and I don’t think the current generation of models can do that.
What's awesome here, is that you often have to use the most advanced LLM available for generation (GPT-4), but you can use much simpler models to categorise the answer!
> But when I sat down with two graduate students (Sabrina Lee and Jessica
Mercado) to manually characterize a thousand passages from fiction
...
> It took the three of us several months to generate this data
...
> The total cost to my OpenAI account was $22. Of that amount, about $8 was testing the prompts and running on the (cheaper) Turbo API. The final run on GPT-4 cost $14; if I had completely replicated our work from 2017 on GPT-4 it would have cost $140.
Seems like coming up with good prompt itself is quite costly at around 1/3rd of the total cost. I wonder if there are any ways to iterate on the prompts cheaply?
In some sense, the author contradicts himself by using the model to do something non-trivial that parrots certainly can't do.
Anything involving numbers is red flag territory for LLMs, or at least math.
Wonder how it would do with medical lab data. I imagine it needs to ask questions of less than or greater than, which is math.
https://tedunderwood.files.wordpress.com/2023/03/screen-shot...
I wonder how sbert would do here given that it's a middle ground between words and full context.
regex = r"\d{4}"