I hope the future generations of models will stop with the fluff/textual clutter/overexplanations - I don't know the right term. They easily fix those once you tell them, but they still do them by default. What I mean on this page, for example:
> A WORLD, FOUR BILLION YEARS IN THE MAKING
> DEEP TIME · CONTINENTS IN MOTION
> Distance becomes an ocean.
> Continents drift apart. Shallow seas spread across land that will one day be dry.
> EXPLORE 4.54 BILLION YEARS OF EARTH HISTORY
> EARTH THROUGH TIME
> Scroll to travel through time Drag to orbit
> THE PAST IS ANOTHER WORLD. IT IS ALSO OURS.
> ILLUSTRATIVE TRANSITION
> Ga = billion years · Ma = million years / Event spacing is not linear
It gets to the point where if you ask a model to make a tiny toy "OS", it'll write stuff like "REAL MODE. REAL MULTITASKING. DRAG WINDOWS / TAB TO SWITCH" directly in the background of the desktop of said toy OS.
The term that Anthropic is now using is "mannered prose". If the creator of this example simply prompted "Remove all mannered prose" then the entire experiment would suddenly become normal sounding. In the fable 5.1 prompting guide, they have a longer prompt for removing mannered prose too if needed. I find it works on Astra as well.
At some point I wonder where and when AI will just be taken as a means of exfiltration of knowledge from the traditional cathedrals/bazaars/silos - I mean, obviously, the more that knowledge is shared the more valuable it becomes .. but couldn’t this also mean that the training data included apps like this?
A giant for-profit corporation based in Silicon Valley, funded by titans of finance, and in bed with the government is about as "traditional" as it gets.
Fewer prompts just means that you are getting something closer to the stolen goods in their entirety. LLMs are made to obfuscate the source of their responses to avoid copyright infringement lawsuits. Fewer prompts is not an indication of breadth of "knowledge".
The overwhelming majority of the universe contains zero human input. I'm not saying that makes this interesting (though obviously it was entirely built on human input, if involuntary), but I find physics and chemistry and astronomy and space and energy and so on extraordinarily interesting, despite humans being irrelevant to their construction or behaviour.
I wonder if we will get to the point where this is all so easy that we'll need something like Flickr to organize all the random things one has vibe-coded.
It's already starting to feel like the days when your uncle would make you sit in a dark room and look at his vacation slide show. Sure, they're fine photos, but we quickly got to the point where a photograph was so easy to create that it needed to be extraordinary not just in execution, but in vision, timeliness, originality, etc. And even then, if you couldn't get it in front of the right editor, it just sat in a pile.
It is amazing, that just a few years ago this would have been a pretty amazing, portfolio-worthy project and today it's "pretty cool", said with a shrug.
It has to do with prompter. If you goal is just the result, make more sense for the AI to do that better. If you structured your prompts, then the resulting code is closer to how senior engineers will write it. Decades ago, probably before you're born, a lot of shareware codes are written way messier than what you see with AI.
This is objectively wrong. Over the north pole, Greenland looks as large as North America does when it is the center of the viewport. This is some strange mercator-esque projection applied to a globe.
I’ve wanted something like this for a long time. However, I wanted it with extensive fact-checked information and AI slop is the opposite of that so I feel this is kinda pointless.
I've been seeing lots of demos like this, especially with games. Yes, they are impressive, but I'm surprised nobody is talking about some of the things I've noticed with GPT-6: the code it writes, by default, is surprisingly messy and obviously unmaintainable. It's also very...not human. No human would write code like it does. I get the sense that the optimization of these models on benchmarks is causing their behavior to morph into a very brute-force type of approach. It makes you think: when code is suddenly very cheap, does good style and organization even matter as much? Or were those just important for humans? (I think they still do, for the record).
I haven't looked at the code honestly, but it's a interesting observation. My impression of the way OpenAI is going is instead of making a larger model smarter, they're going wide with just extremely high effort and a lot of parallel agents to get towards a goal.
For example, for design and architecture, I prefer fable, but for pull requests I like GPT because it's so pedantic about things.
35 comments
[ 0.18 ms ] story [ 14.2 ms ] threadI'm more interested in seeing how others are prompting and working with LLMs than the result itself.
Care to share?
Publish the prompts or re-run them today to see if you get the same quality result.
> A WORLD, FOUR BILLION YEARS IN THE MAKING
> DEEP TIME · CONTINENTS IN MOTION
> Distance becomes an ocean.
> Continents drift apart. Shallow seas spread across land that will one day be dry.
> EXPLORE 4.54 BILLION YEARS OF EARTH HISTORY
> EARTH THROUGH TIME
> Scroll to travel through time Drag to orbit
> THE PAST IS ANOTHER WORLD. IT IS ALSO OURS.
> ILLUSTRATIVE TRANSITION
> Ga = billion years · Ma = million years / Event spacing is not linear
It gets to the point where if you ask a model to make a tiny toy "OS", it'll write stuff like "REAL MODE. REAL MULTITASKING. DRAG WINDOWS / TAB TO SWITCH" directly in the background of the desktop of said toy OS.
- rhetoric of universal humanity as backdoor to software anthropomorphism
- commercialistic exposition of features (in sentence fragments) and foolproof instructions for use
"Put 'em together and wadda ya got?" TURN OFF YOUR BRAIN
https://platform.claude.com/docs/en/build-with-claude/prompt...
Funnily enough even the image model does it: gpt-image just throws little bits of text all over the picture wherever possible.
Edit> on second refresh it worked
It's already starting to feel like the days when your uncle would make you sit in a dark room and look at his vacation slide show. Sure, they're fine photos, but we quickly got to the point where a photograph was so easy to create that it needed to be extraordinary not just in execution, but in vision, timeliness, originality, etc. And even then, if you couldn't get it in front of the right editor, it just sat in a pile.
It is amazing, that just a few years ago this would have been a pretty amazing, portfolio-worthy project and today it's "pretty cool", said with a shrug.
https://terrashift.io/ (it doesn't work on mobile right now, lame I know)
What a fun time to build!
- I prompt better when I understand it
- Good code helps the AI too
- I can catch wrong assumptions it made when it's readable
For example, for design and architecture, I prefer fable, but for pull requests I like GPT because it's so pedantic about things.
(Use its dropdown.)
And then I clicked the “Explore” button on the top right and there are even more things that you can see! Pretty neat.