it used to be that you had to draw the rest of the owl, now you can just slop it out and see if it works or not
English
difficultyang
2.9K posts
difficultyang
@difficultyang
More social alt of @ezyang
Katılım Nisan 2022
63 Takip Edilen3.5K Takipçiler
the cool thing about ai slop is you can rapidly test hypotheses on what happens if you build the program around a core abstraction built one way, and then just keep tweaking it and seeing the implications until you're happy
English
holy shit, i've seen the tweets about backrooms, but why haven't I tried it myself before
English
Gemini truly [redacted] as always: "I have determined the prompt includes musical notes along with a filename, potentially revealing the song's melody. I've determined I have music notation: C C D E C E D, indicating the song's melody."
English
An even funnier eval is to give the LLM a wav of the tune but with the wrong filename and see what they answer.
English
New eval: "Identify this song: g4 | c8 c8 c4. d8 e8 f8 | a8 g2"
English
me and my boy claude be identifying music from notes and then Anthropic me hits me with that "API Error: 400 Output blocked by content filtering policy"
English
Do LLMs truly understand C++ ownership rules or have they simply memorized enough examples to present a simulacra of understanding
English
@zeewahee Idk man, after I give up trying to incrementally code review I just sit down and figure out what the correct data design is lmao
English
@difficultyang have you found any techniques that help it do better?
English
An interesting data point is that Codex 5.5 cannot be trusted to design good data structures purely from behavioral prompting. (I'm sure it can come up with good ideas if you prompt it, but not if it's incidental.)
English
@enjoyingthewind I have a pretty good design doc to try one shotting this from so we can try it
English
@cosminnegruseri This post was prompted by Codex coming up with a terrible internal data representation for an autograd tape with some special checkpointing behavior
English
@blueprintsmb22 We ended up buying during COVID because there wasn't any rental stock we were happy with LMAO. But yeah, we did spend a lot of time touring houses that we weren't excited about. If there isn't inventory in your locality, that's probably the more important problem.
English
i think the value from $500k to $3mm is very bad given lack of inventory. My wife keeps dragging me to open houses but I'm fatigued touring homes I have no desire living in that cost too much. Much happier renting, keeping our $$$ in the market and calling the building when the dishwasher breaks but we are fine with apartment living having lived in NYC for so long
English
@blueprintsmb22 I do! I can believe premium areas are several million. But there's a lot of IMO reasonable towns all over NJ that are not that.
English
@difficultyang You live in NJ? We rent a small apt w our daughter and the math to me makes no sense to buy. Homes are selling for $1.2mm with no bathroom on the first floor and cabinets from pre the Great Depression x.com/blueprintsmb22…
Blueprintsmb@blueprintsmb22
No bathroom on first floor meets no inventory => crazy home price x.com/nypost/status/…
English
@_seemethere I feel the correct terminal state is that you are only giving the agent tasks that it will reliably one shot
English
Agentic coding is amazing but it’s also dangerous in that it basically has the same reward function that gambling has.
You push a button and then it spins, lighting up with visual cues rewarding your brain with good productivity.
So many people I know are getting addicted.
English
It is about value. The cost doesn't matter, but getting a good value for money does matter.
VEO@vrexec
Is anyone else’s wife like this? I could make $10,000 in a day on a quick stock trade and my wife would have no reaction… wouldn’t care at all. But if I came home with a new jacket that cost $250 to replace the one I’ve been wearing for over 10 years… I’m jeopardizing our entire family’s financial future.
English
@tenderizzation *bows deeply* I apologize for falling for the engagement bait again (om nom nom)
English
