Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
My amazing story with AI today. My order of the sinking star savegame was corrupted after opening my steam copy save with the pirated one which was on lower version. There was no backup and there was severe data loss. So I just described what kind of state the world was in and to try and rebuild it when recovery was impossible - and the bastard (sol 6.1 low) one shotted it.
We are in uncharted territory, and I just can't predict the way the dice will roll. So in which direction did your P(doom) moved this week?
I've been vibevibecoding, by which I mean issuing only a very general outline that the AI turns into a full plan for actual execution (after asking me questions). Then I tell it to execute steps of its plan and change things if I don't like the output.
Back in the day, you had to say in detail what you wanted and even if you didn't understand the code you at least knew what files did what. I had to manage the context of what the machine saw. Whereas now I don't look at the file structure at all.
This is really more for starting a project than finishing (that last 10% of polishing, detail, refinement takes 90% of the time) but you can get pretty far doing things this way. I recommend people, especially if they're on a $100 or higher plan, to write out a halfbaked 5 page plan, ask the AI to ask detailed questions, turn that into a real plan and then just execute. Then you see the true power of the models.
Also this just popped up, Claude Sonnet 5.5 fighting battles in Rome Total War. I thought Sonnet was a bit crappy in my own experience but it seems to be faring pretty well in warfare:
https://youtube.com/watch?v=Mhv8XwmA8R4
The phase change for me was around ChatGPT 5.5 (where I updated my take that was previously, "Yeah, this is useful for some things but is not generally good at solving problems independently"): I had a winning lottery ticket ($20) that I had stuffed into a pocket while it was slightly wet. Then it dried and kind of stuck to itself, and unfolding it ended up tearing some chunks out of the scannable barcode. On a whim I handed it to GPT 5.5 in Codex and asked it to find the actual payload using the information in the picture and repair the barcode. (No online or offline tool I could find could read the barcode even with some half-assed GIMP work to straighten it out.)
The problem was actually trickier than expected: the string above the barcode is in the barcode, but there's also a metadata prefix, so GPT's initial attempt to beam search through character permutations and check the generated code (which it identified as PDF417) wasn't enough. It eventually decided to reverse-engineer the error correcting modules and reconstruct the payload from that; this worked, though surprisingly
zintcouldn't be coerced to produce a precisely matching output (presumably the company didn't use zint to generate the barcode in the first place or were passing some exotic arguments) even though it could produce a valid PDF417 with the right dimensions and payload, so then GPT finished by just writing the bitmap rendering code itself and producing the perfect, repaired barcode as a png. Which I opened up in an image viewer and scanned from my phone into the lotto app, and it just worked.I can't even estimate how long it would've taken me to do all of this; it's obviously not impossible, but the 30 minutes I'd spent de-skewing and adjusting contrast and sending the image through random tools had already surpassed the effort it was rational to expend for $20. Implementing Reed–Solomon error correction alone would probably be a (not so fun) weekend project, and that's before actually extracting the cell data from the distorted image.
/images/17936624789209785.webp
/images/17936624791162748.webp
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link