Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- > It would be foolish to deny the effectiveness of image recognizers based on generalized adversarial networks (GANs), the key neural network technology underlying LLMs
I could be misreading this, but I hope the author doesn't think GANs are used in LLMs. They are cool, though.
by ainch - Yeah, it's hard to take the rest seriously.by chpatrick
- He hallucinated. This is how you verify non AI nowadays.
AI is so good that if you see an hallucination as obviously wrong like this one it’s a sign it’s written by a human.
- Things that are vaguely GAN-shaped might play some role in post training. Though no one would call them GANs or identify them as the key underlying technology.by scarmig
- It’s unfortunate - he’s an author, it would have been fine to stick with an authors perspective that LLMs can’t write (which is true), as well as the copyright stuff (which I don’t agree with but he certainly has standing to give an opinion on).
But he’s made the error of trying to come at it from a technical perspective, when he clearly knows nothing about that side of things, which discredits the rest.
by andy99 - Dude's an artist. He should do what his artistic intuition tells him to. I'm a software developer. I'll follow my own experience. It'll all work out.by tptacek
- It’s likely it won’t all work out.by slopinthebag
- Roughly a third of the market is digging a $1T-a-year-deep-hole with <5yr amortization schedule on all of it and passing around the same $100B like it's in all of their bank accounts simultaneously. Yeah, shit's going to work out for sure, in all of the same ways that flying a plane into the side of a mountain has a clear ending.
- He says exactly this in the comments on his page. This article is specifically about writing/art.by sodapopcan
- For some value of "work out" anywhere between "human utopia that spreads across the stars" to "we destroy everything in a ball of nuclear fire and only microbes survive", yes, it will. The universe will keep on going either way.by esperent
- This is very interesting coming from an author in whose writing autonomous, super-powerful AIs have been a common theme.
Consider his book Accelerando (which I'll take the opportunity to plug again, especially as he's made available for free here: http://www.accelerando.org/fiction/accelerando/accelerando.h...) Not only did I find it quite engaging and thought-provoking, it is also proving rather prescient, and even helpful in decoding some of the things that are happening today.
For instance, in the very first chapter the protagonist spawns agents to go research something in the background and report back to him. And then a year ago, I randomly became curious about a rather involved topic (how fast could we feasibly replace all human labor with robotics), but I did not want to spend time researching so I outsourced it to Google Deep Research which churned away for almost half an hour and came back with a 30 page report with 49 citations via "actual internet searches for verifiable sources." (If you're curious about the conclusion: not for a very long time partly due to critical supply chain constraints.)
After I went through the report, it suddenly struck me: my agent may have executed in a GPU cloud instead of a cybernetic brain, but holy crap I had literally just lived a SciFi scene! A scene that I did not expect to experience in my lifetime!
Which is why I found TFA a bit unexpected. TFA says a few things that I would disagree with. Like, no, AI is not a "stochastic parrot" and I'd assume he'd be primed to realize it. And AI is not competing with authors -- other authors with AI are, and using AI trained on real text to aid writing has been a thing since the days of red squiggly lines in word processors, which TFA even acknowledges.
If you see his last few comments (https://news.ycombinator.com/user?id=cstross) it's clear he's pretty negative on the tech industry and Capitalism, which I tend to agree with. I can see how that could color his thinking.
That may also mean he's avoided LLMs to the extent that he is not aware what the frontier models have become. I would encourage him to put aside his distate and re-engage with them deeply; maybe he'd be at least a little bit excited to see some of his writing turn out to be prophetic in some good ways besides the bad.
by keeda - > I'd quite like a tool (running entirely locally on my own hardware, with no cloud service and no copyright-thieving grifters making bank on it via subscription fees) that digests a manuscript and derives a scene-by-scene timeline, that I could then query interactively and use to plan my next round of edits. Being able to map out where and when each protagonist and minor character shows up, and see a frequency distribution heat map of names in the manuscript, would be useful.
Interestingly enough, I do this with various books I'm reading. E.g, I'm currently rereading Accelerando and had Claude generate a wiki-like timeline of key events, characters, and salient plot points. That makes it easier to jump around when I want to re-read a section and grok a plot thread that is scattered across chapters. Ironically, it also exposes inconsistencies (or "hallucinations" as some might call them) in the text because the author didn't have an AI proofread the text.
by sxp - Are you able to share? Both your process and the specific timeline for Accelerando. I would love to build one for A Fire Upon the Deepby derektank
- > Ironically, it also exposes inconsistencies (or "hallucinations" as some might call them) in the text because the author didn't have an AI proofread the text.
That's not irony, or even cute. He's literally describing a tool he'd like to use to plan edits to his writing.
by magicalist - You could build a tool like that and on a MBP with 48+ GB of RAM all of that can happen locally in terms of keeping your content locally. I am sort of an outliner and planner and I heavily use AI when writing documents for work. I don't see why fiction / prose would be any different? Indexing your content with a vector RAG and having a little 27B model locally could do everything he wants with some help from a big model to implement it all. It sounds like a weekend of coding for a functional prototype to me.by bitexploder
- Not being able to see the forest for the trees I suppose is a prerequisite for success in Silicon Valley, because it one had even a fraction of Stross's perspective on the matter they'd be in a different line of work.by Fricken
- It kinda looks like we can only train AI on dead people's data.
I wouldn't be too upset about that. They write better anyway.
by andai - Should maybe rather than can?by tim333
- I don't think it ever gets old to restate that the singularity and all of the gradiose promises of a glorious 21st century have last mile problems.
But this post seems to mostly channel Harlan Ellison. And while he was amazing in his prime, with some absolutely legendary rants, he didn't age well in the end.
I do like his challenge to make LLMs useful to himself and to other authors though. Someone needs to make that happen and nearly exactly how he described it.
by LogicFailsMe
I'm sure I've read about Hollywood scriptwriters having tools like this.I'd quite like a tool (running entirely locally on my own hardware, with no cloud service and no copyright-thieving grifters making bank on it via subscription fees) that digests a manuscript and derives a scene-by-scene timeline, that I could then query interactively and use to plan my next round of edits. Being able to map out where and when each protagonist and minor character shows up, and see a frequency distribution heat map of names in the manuscript, would be useful.(Or is this just the Gell-Mann amnesia effect striking again?)
by ben_w- > They're word-association mechanisms with no embodiment and no way to associate the text vectors they manipulate with real-world phenomena.
Doesn't most of this also apply to a guy living in The Matrix?
by andai - Searle* strikes again!by ericpauley
- Yes, but also for humans who learn of distant lands by reading (books or news, just so long as it's reading).
And also these models have been associating with real-world phenomena from the first moment their training data did, and also those text vectors are (to varying degrees) associated with corresponding image vectors in multimodal models.
Of course, the Plato's cave critique would still be valid.
by ben_w - Embodiment doesn't exist, even in regular humans. We do not have direct access to reality; we have sensory inputs that are much lower bandwidth than one might expect but correlate with external events, and our brain uses them to form rich world models that are usefully predictive. Most of the world we experience is just in our head.by scarmig
- Stross complains about people from China hammering his blog and stealing his work, but I feel like China releasing all these free models is a lot more defensible. Yeah, they're pirating a bunch of stuff and stuffing it into a blender but they're giving away the resulting soup for anyone who's hungry. (And though running it on your own hardware is a stiff proposition, I would bet Kimi K3 can figure out the scene-by-scene timeline thing.)
- From the comments:
> I will concede that using LLMs in software development is a different category from fiction, insofar as the languages and APIs are much smaller and more tightly constrained and in software you actively don't want long-range dependencies between elements (something you often do want in fiction, i.e. put a gun on the mantlepiece in part 1 of the book, pull the trigger in part 3).
by Apocryphon - It can definitely plan such things. Because agentic systems don't have to output a book all at once. The mainstream still lives 1-2 years in the past.
Agents can plan out a narrative draft, write it out section by section, maybe out of order, then tweak it, plant clues if it wants to, then trigger it in another chapter file etc. Then read the whole thing, reconsider, etc. It can iterate. I'm not saying it will be great literature, but the problem isn't the inability to create long range references.
My bigger point would be this: arguing from such mistaken technical ideas, when not having any idea about the tools is a sign that people's real problem is none of those technical limitations, but Western culture has lost the vocabulary to express their true sentiment. They are grasping for words but can't reach for "soul" and other religious terms because that would sound ridiculous in a secular modern world. But they clearly want to express that and we'd be ahead if we could discuss in such actual terms. Would you want the machine to write literature if, hypothetically it could write indistinguishably from any great writer? I guess not. Because there is something mysterious and important about the human spirit. (the same vocabulary-inaccessibility is plaguing other culture war issues too). People will rather talk about fake water consumption problems than touch the topic of the human soul or spirit.
by bonoboTP