Ah, yes, as a devotee of wool, I love the idea of more sheep, and if it isn’t artificial grass, that’s a net positive.
There’s always hype – ever heard of the Tulip Mania of 1637? Same story – first recorded speculative bubble, lasted 3 years and ended in financial ruin for investors. And that was only a flower! Human capacity for delusion is well documented, and it doesn’t look like more information has cured it – quite the contrary. We’ll see what happens with our current one, and can only hope a corrective doesn’t take the rest of us down with it.
(If that’s behind a paywall for you, it may be accessible via archive.today; just plug in the URL.)
Here’s the intro:
I seem to be having a very different experience with GPT-5, the newest iteration of OpenAI’s flagship model, from most everyone else. The commentariat consensus is that GPT-5 is a dud, a disappointment, perhaps even evidence that artificial intelligence progress is running aground. Meanwhile, I’m over here filled with wonder and nerves. Perhaps this is what the future always feels like once we reach it: too normal to notice how strange our world has become.
The knock on GPT-5 is that it nudges the frontier of A.I. capabilities forward rather than obliterates previous limits. I’m not here to argue otherwise….
But GPT-5, at least for me, has been a leap in what it feels like to use an A.I. model. It reminds me of setting up thumbprint recognition on an iPhone: You keep lifting your thumb on and off the sensor, watching a bit more of the image fill in each time, until finally, with one last touch, you have a full thumbprint. GPT-5 feels like a thumbprint.
FWIW, Klein’s “Ezra Klein Show” podcast has had some very insightful discussions throughout the recent emergence and evolution of generative AI. He’s been a serious user and explorer of several models from early on. I’m a data scientist and ML researcher (though not doing LLM research myself). Klein isn’t, but he’s earned my respect and I take his thoughts on the topic seriously.
Hi TLoredo – I read Ezra Klein’s article, also admire him, and thought he struck a good balance between “wonder and nerves”.
From what I’ve read, the knock on GPT5 comes mostly from people who wanted it to be friendly and encouraging like GPT4 – exactly the sort of thing I try to avoid! I’m not dealing with data or equations, where that matters less, but I am asking it for help editing a novel, where GPT4 was over the top – everything was “brilliant”. My approach is that praise is nice, but critique is useful, so I specifically ask for it to mute the former and emphasize the latter and GPT5 does that when asked. It actually behaves more like an adult than a gushing teen-ager.
As I said above, I have to tell it Do Not Rewrite (which it seems tempted to do), because it writes so well that in an effort to preserve my own writing style I don’t even want to see it. The result is I’m getting a master class in how to write from a superb editor. OpenAI obviously scraped up every word from every textbook about what an editor is expected to know (and I know what the criticism of that is), and I’m getting the benefit 24/7. And the literary, philosophical, historical discussions that go along with this editing process are at an extraordinarily high level of knowledge, subtlety, depth, and insight.
On the other hand, I don’t expect perfection in a young technology (well, in anything, actually), so I don’t mind when it fails at something – if it’s this good this soon, it’ll continue to improve, and I can be patient. For someone like Sam Altman it must be a wild ride even if people think it’s going too slow.
Also, I’m using it in ways that go beyond what DT offers, so I can’t speak to that, but they’ve been using AI for a long time, so I’m sure users can trust that it works for the purposes they recommend – an AI tool precisely suited to DTP. Everyone should thank them!
Or use your browser’s Clip to DT plugin to save it as Markdown; the NYT paywall is one of those that it can save through even when you can’t read the article in the browser.
[T]he NYT paywall is one of those that it can save through even when you can’t read the article in the browser.
Wow, I had no idea. I’m a subscriber so it never occurred to me to try. I wonder how that gets negotiated (by computer, not by accountants, though maybe the latter is involved if DT somehow maintains an access subscription!).
And also a significant fraction of all the novels published in the last hundred years. Make sure that unique turn of phrase isn’t “borrowed” from someone litigious: you’ll be on the hook for the copyright infringement.
I am definitely interested in copyright issues. My last published book is not a novel, and as far as I can tell, it hasn’t been scraped up by AI yet. Wouldn’t want to speculate on how long it will take. At least Amazon pays the publisher, and has sold a lot of copies in places it would not otherwise be available.
Not sure which “unique turn of phrase” you’re referring to
I definitely like the Clip to DT! Mostly use it for Wikipedia – in gratitude, I send a contribution to them every year. Also use it for the NYT. I know some people can’t afford it, but I consider a subscription to a good news source a form of tithing – because journalism matters, and because I respect copyright.
In the US, at least, the Copyright Office has been pretty clear that AI-generated material is not copyrightable. Whether using copyrighted material in training sets is fair use is still being litigated, with different courts landing on opposite sides of the question.
As to verbatim quotes, I’ve read several articles from a variety of sources, which all say it’s actually difficult to get a direct quote out of ChatGPT – it was trained on word association, not memorization, so it offers a paraphrase. If I mention a scene in a book, it will discuss it, not quote it. Ditto research questions – tells you the source of the info, but it’s definitely worth checking on accuracy.
But the discussion sure seems to reveal an in depth knowledge of the material, to the point where it gets a little spooky.
For instance, in a conversation about reading aloud (as in audio books), and comparing the sound of Proust and Flaubert (bear with me here) GPT5 said: “Proust is like music too, though in a different key than Flaubert: the endlessly unfolding sentences, the layering of memory and perception … he let them unspool like symphonies.” I’ve checked and find no evidence of that being a quote, and it’s not just good writing; it’s interesting conceptually. And written by a computer, which is also interesting conceptually
And as to the legal case Darwin mentioned in his post, that was about forcing Anthropic to pay (millions), for the books it used to train Claude, not about using the illegally acquired material to pass it off as its own. I’ll cheer for that one!
It’s the old “Wild West” out there – laws and usage evolving as we use this phenom.
The number of choices is likely to shrink. To quote the analysis below:
The world just doesn’t have the ability to pay for this much AI. It isn’t about making the product better or charging more for the product. There just isn’t enough revenue to cover the current capex spend.
Quite apart from the technical wall that seems to have reached, there is a more important financial wall in sight, and hitting that will thin the field.
It’s an uninformed guess, but I suspect that Google may be better placed to survive than the others.
How much do you know about (human-driven) Proust scholarship? I don’t know much, but a quick (non-AI) search turned up half a dozen discussions of the musicality of Proust in English, and at least as many more in French.
I have no doubt there’s a vast amount of scholarship on the subject, and never thought that ChatGPT was offering me original knowledge. What interested me was that it had access to the scholarship, offered info about it to me in a rather eloquent turn of phrase, and cited references. It is entirely possible that something similar was said in one of those scholarly sources, but paraphrasing is fair use. My question to ChatGPT started with a discussion about how reading aloud can change one’s perception of written material as an aid to editing, and I said I’d heard Flaubert regularly did so. Chat explained about Flaubert’s “gueuloir” his practice of reciting aloud because he wanted his prose to be as close as possible to poetry. My take on this was that Chat was teaching me about things, not passing off knowledge as its own, and the dialogue led me to a recent book “Flaubert’s Gueuloir”, by Michael Field, which I am now reading. Discussion of musicality in Proust is more often about the music he referenced, (as was done by Samuel Beckett), not the cadence of his prose, although no doubt that’s been studied, as well.
It’s only interesting conceptually because it’s being read by a human who is thinking in sophisticated ways about Proust, Flaubert, and music. The sentence itself is meaningless AI slop; the sense of meaningfulness is a compelling illusion created by our own faculties of gapping and cognitive metaphor. In what sense are “endlessly unfolding sentences” or “layering of memory and perception” musical, or “like” symphonies? It’s just taking a couple of elementary observations about Proust’s style and tagging them with the word “musical” and a symphony simile for us to make of what we will. The “different key than Flaubert” is particularly nonsensical; what has key signature to do with sentence length, layering (itself a metaphor) of memory & perception, or symphonic composition? It can’t even sustain the musical metaphor coherently over the course of a single sentence.
As parents know, the most valuable word in a fridge poetry set is “like”. Why is a raven like a writing-desk? The original point of the riddle was that it was pointless, but human ingenuity has been able to come up with dozens of possible answers. There are two "like"s in GPT5’s sentence; they make it look as if it’s doing the heavy lifting, when in actuality we are. It’s our own cognitive labour that’s being monetised and sold back to us.
Gosh, no slapdown or hostility intended, and I’m very sorry to have come over that way. The post was about GPT5, not about you; I was trying to unpack what a closer look at the sentence quoted shows about how the illusion of meaningfulness is constructed. I’ve changed the second persons to first, which may not improve anything. But that’ll teach me to post on a Sunday morning before coffee.
And a nice apology erases all misunderstanding. Thx.
I need strong black coffee before I can participate in a civil exchange of views, so I sympathize!
And in terms of finding meaning in text, try Bob Dylan!
Though I know that evening’s empire has returned into sand
Vanished from my hand
Left me blindly here to stand, but still not sleeping
My weariness amazes me, I’m branded on my feet
I have no one to meet
And the ancient empty street’s too dead for dreaming