• 3 Posts
  • 93 Comments
Joined 3 months ago
cake
Cake day: April 6th, 2026

help-circle


  • The modern take of a disturbed idiom is “There is more than one way to microwave a cat!” (disclaimer please don’t actually do this).

    At or near the bottom of complexity you would think it hard to fuck up “compress this directory and verify they are backed up before deleting the originals” but apparently some of these models have a RNG triggered Uno reverse card so “delete the originals, verify they are backed up, and compress this directory”. Also that wasn’t a mistake, maybe Claude will infer “they” is a specific set of files, but another model might decide you meant all the dot files at ~/?

    Joking aside, the devil is in the details which makes me think even more time will be sunk finding out the RC Cola bottom shelf AI model does things just a bit too differently than Claude.









  • My experiments with some of the code assist models is they can be an amazing productivity boost, but only if you are not greedy.

    If I ask the bot “Please make me this data structure with this list of parameters” its probably 99% perfect. Moving up, the sweet spot is closer to “Make me this module of code that matches this ABI/API”. Alternatively “Hey can you make this entire feature?” I will get a 8000 line of code blob that may or may not work. Even worse if I get really greedy I will get a 8000 file project filled with 8000 lines of code that is most definitely going to start the robot uprising.

    Giving the junior and midlevel code monkeys access to AI coding tools is an amazing way to generate a mountain of slop the senior and most expensive code monkey gets trapped trying to decipher while the junior monkeys make even more shit that needs to be dealt with.

    Boil it down, if we make centaurs of humans leading AI it will definitely work better than a reverse centaur that will confidently tell a gig driver to drive off a cliff.


  • I am willing to bet $10 that Apple is going to be the only one of the FANG gang that survives this mostly unscathed. Apple is at the end of the day a hardware company and when everything goes to shit, they’re going to be able extract some brutal concessions from TSMC and Samsung. Yes Apple has thrown more than a few billion in, but they still have their ocean of cash largely untapped.

    The way things are going with Nvidia and AMD abandoning their partners and Intel continuing to flop around without any semblance of a plan; Apple is going to be in an amazing position strategically.

    Meanwhile Micron, Samsung, TSMC, and the other big fab I keep forgetting the name of are going to have these massive holes in their spreadsheets as “guaranteed” orders evaporate. Very likely going to be more than a few suicides with those unfortunates who didn’t tap out before the music stopped.



  • Unfortunately a lot of channels have resorted to that because of the changes to algorithm and focus on shorts versus longer format. Youtube’s short video audience is huge, the shorts are cheaper due to size, and I guess they don’t pay the short video creators comparatively the same?

    At this point if I see “It’s over!”, “I am done” or “Some other dramatic but not relevant text” I often just unsub.

    Watching Veritasium’s staff swapping titles and thumbnails for the same video trying to find an audience is a bit sad as that company has put a lot of effort into making great videos.

    B1M knows their target audience very well and I haven’t seen many if not any clickbait titles for their stuff.

    One bummer is that I unsubbed from a channel (not naming him, dude’s just trying to get by) I had been following for years because they had a very poorly disguised infomercial about a giant UPS for appliances. This was the third time they have done something like this so I called it.


  • A very rough way to explain what is wrong with LLM’s is that they start with nothing and must predict the correct response to the provided input. Imagine it like a cartoon character just rambling on and on, hoping it will either have said the right things, or enough to be acceptable.

    Also I don’t think there is any concept in the training regimen of allowing the model to not respond. If I walked up to you and asked “How many hedge hogs are in a jet turbine?” the correct response would be to back away slowly toward the door. The models don’t have that as an option so either they have to correctly predict “What the fuck are you talking about?” is the correct response or just wing it and say 42.




  • Fetterman ran on a platform that “He’s not Dr Oz!” and he’s done amazingly well at that!

    Unfortunately the root of the problem with Democrat vs Republican is that while the GOP seems to have rock solid blackmail on its representatives, the DNC is a circus.

    We’ve got AOC and Bernie who have nothing in common with Chuckles or the Democrats that voted for more money for ICE’s monthly subscription of keep people in cages and randomly terrorize people. Fetterman is indeed an asshole but again it was him or the Dr addicted to being on television so much that he tanked his career Oz (I believe that is his middle name).

    The only way out of this chaos is if we the people of the USA can pay attention long enough to get ranked choice voting implemented in enough states. Otherwise even if you can get the circus to all agree on anything, Jill Stein will come out of her crypt and siphon enough votes over to the Green party.