38 comments

[ 2.2 ms ] story [ 75.6 ms ] thread
America's second finest news source.
This has definitely got an "Onion" flavor, and makes me wonder what the future of that publication is going to look like, if AI is this good at writing absurdist humor articles.
Of all the all the applications of LLMs I've seen so far, I think generating endless amounts of humour is still the best thing they can do.
Seriously, the «best thing they can do» is probably to "make it a module" and build the interfaces to integrate it in systems where it is controlled.
I use it for brainstorming ideas for writing. The fact that it can come up endless ideas incorporating whatever surface level details you demand makes it extremely effective for this purpose. At least so far I've not seen it produce quality you would want to use that wasn't filtered through a creative human (even the good humor and poetry it produces on its own is cherry picked), but we shouldn't underrate the value as a brainstorming tool. And the day is probably not too far away where it can filter for itself.
> even the good humor and poetry it produces on its own is cherry picked

This is so funny in itself, and speaks so much of human expectations :)

Yes, the really good examples are cherry picked, but there's a loooot of "good enough" examples and some cherry-picked bad examples of ChatGpt.

To me this is mind boggling. My ML knowledge stopped at 2018 levels and ChatGPT makes me think picking up ML once again.

I remember training a ML model in 2018 in a yandex course for image description. Up to the last few epochs the output was gliberish, then a few epochs later booom, readable english.

I've used it to help come up with metaphors in Japanese. Have to filter out its misuses of words, but still, very helpful source of ideas.
I like it, but I wish I could zoom out, the font is obnoxiously massive and ignores the browser's zoom level
Cmd + Shift + - on Mac
Reminds me of the line in VeggieTales, In the future, entertainment will be randomly generated. [0]

[0] https://youtu.be/j4Ph02gzqmY

Having seen far too much Veggie Tales, I'd say that show already feels randomly generated. With bible inputs too heavily weighted.
Interesting post this of today, of the «treadmill that runs away from you»:

when somebody claims that we have with "LLMs" a potential quasi-sentience we sometimes reply "but we have already invented the cat", and the idea of an AI based treadmill which you have to chase around to exercise makes us note - "that would be a dog".

The treadmill that runs away from you also to me showed a weakness of the approach; I think it's pretty obvious that the treadmill that chases you is more obviously funny. You get to touch on ominous, push more psychological buttons, and still have the fundamental hardware humor of a treadmill designed to move rather than just be still. I think any serious humor writer would make that change.

On the other hand, this does open up some interesting spaces no human would normally explore, for that sort of reason. https://www.garlicgpt.com/post/shopify-ceo-quits-after-losin... is just... not quite right... and almost funny just for that. At least to me. But that's probably not a very reliable measure of humor. A human humorist could probably riff off of the ideas in that that fairly effectively. But as written, it just never quite lands.

` One might question what drives the customer to pursue the treadmill, but Bill observed that "With a price tag of $49,999, customers are concerned about their investment getting away from them."

`

The Shopify CEO joke has the sophistication of a 6 year old with the breadth of knowledge of an adult.
Whats the difference between a 6 year old with the breadth of knowledge of an adult and an adult with the breadth of knowledge of a 6 year old?

Voting rights.

I didn't realize I was reading the garlicGPT post (I often open in new tab a bunch of stories and then flip through them)

Honestly I was very surprised by this, checked the date to make sure it wasn't April fools, and then took it as gospel.

Well done.

In case this isn't available in the future, here are the hallucinated article contents:

Peloton launches new treadmill that runs away from you

by Leo Lightfoot

Peloton, the fitness company known for its interactive bikes and treadmills, has announced a new product that aims to challenge and motivate users like never before. The new treadmill, called Peloton Chase, is designed to run away from users at varying speeds and directions, forcing them to chase it around their homes or gyms. The company said that the Peloton Chase is powered by artificial intelligence and sensors that track the user’s location, heart rate, and performance. The treadmill can also adjust its speed and direction based on the user’s preferences and goals. “Peloton Chase is the ultimate workout for anyone who wants to push themselves to the limit. It’s like having a personal trainer who never lets you slack off,” said John Foley, the CEO of Peloton. Foley also said that the Peloton Chase will be compatible with the Peloton app and membership, which allows users to access live and on-demand classes led by instructors. The Peloton Chase will be available in early 2023 for $4,999. The company said that it has already received thousands of pre-orders from customers who are eager to try out the new treadmill.

by Leo Lightfoot

Deciding to use that name makes the whole article even better.

Can you add a RSS feed?
Seconded! I would love to get daily updates.
The word you're looking for is not hallucinate, it's "confabulate".
I think "hallucinate" is the term-of-art that the AI community adopted for that behavior, but I do agree after having looked up the definition that confabulate is much more accurate and I wish they had picked it instead
> much more accurate and I wish they had picked it

In comedy?!

Err, sorry, by "they" I meant the AI community which settled upon "hallucinate" as a term to describe the LLM's behavior. I think this site was 100% accurate to reference the same term that the AI folks do
The AI community already had picked confabulate, but hallucinate is winning out in actual usage now.
Damn, I was really hoping to find the prompt that generated the story in the page source -- be entertained and learn something simultaneously
ChatGPT has continued:

The release of Tyson's new vegan chicken nugget comes as more and more consumers are seeking out plant-based alternatives for health, environmental, and ethical reasons. According to a recent survey, over 60% of Americans are now reducing their meat consumption, with many citing concerns about animal welfare and the environmental impact of animal agriculture. By offering a vegan option made from real chicken, Tyson may be hoping to tap into this growing trend while still appealing to those who prefer the taste and texture of meat. However, it remains to be seen how the market will respond to this unique approach to veganism.

I wish they had similarly run with the idea of Netflix watching you.
If there was another LLM for applying “internal consistency” across generated hallucinations this could easily work as a unique story series generator.

Throw in an updated Stable Diffusion and we have an infinite supply of tv entertainment.

Don’t like the current series it’s showing you, start fresh with a new random seed…

Side note:

The IKEA maze article was spot on and shockingly “close but not quite” realistic.

I’ve often referred to the “IKEA time dilation field” you walk in and suddenly it’s 6 hours later and all you’ve picked up is a set of coffee cups and a lamp shade.

ChatGPT is surprisingly good with humor:

Me: I want you to act as Jerry Seinfeld, the comedian. Give a stand up set, with a punchline, talking about people who get out of their seats before the plane has stopped at the gate.

ChatGPT: You know, sometimes I feel like these people who are in such a rush to get off the plane are the same ones who take their time ordering a coffee at Starbucks. They'll stand there for ten minutes debating between a latte and a cappuccino, but as soon as the plane touches down, they're like "Get out of my way, I've got places to be!" It's like they only have patience for the things that don't matter.

I just released the new version of my site https://aidev.codes which will allow you to easily create your own "hallucination" pages including images, since it has stable diffusion integrated. https://news.ycombinator.com/item?id=35086436&fbclid=IwAR2Sr...

For example you can write "Please invent an onion-style article in a web page about the Hacker News website, Elon Musk, and cheese. Include a few appropriate images." Then "great, but make it more dramatic and scandalous" and you get this:

https://aidev.codes/runvnc/trip1/

which admittedly is not great so I think that it kind of proves that GarlicGPT has a good prompt.

This needs an RSS feed. Why doesn’t it have an RSS feed?