Stevey P

@kentcdodds @DavidKPiano There's a possibility that agents do NEED to function like humans in memory - but often I don't know that "reiterate every related memory" is something the brain does on every toolcall i/o. Aside: that sounds pretty effective! better than current in-Claude one for sure.

Oh, you might like my memory system for Kody. Two tools: search/execute Model is told to use search first. Every tool call requires memoryContext which effectively is the model explaining what it's trying to accomplish with this tool call. Every tool call returns memories relevant to what the model is trying to accomplish. Seems to work pretty well.

ngl, I think everyone trying to solve ai memory is way off mark rn.

@DavidKPiano @kentcdodds 💯 Yep exactly. And it’s incredibly difficult still to try and codify that goal (you hinted at this in your skills/deterministic tweet recently too).

I think the biggest mismatch is that current memory systems work in terms of "what happened", not "what should happen" (i.e. goal-oriented) "User stated preference for this" "We did that" "These files were edited recently" all that is useless without an end-goal At least for me, memory works best when I have a goal in mind and I can "walk several graphs" (like what you said) to pattern-match previous observations that can get me close to that goal

Innovation should still keep going as that will likely do what it always does: learn, retest, improve, etc. This is maybe more of a prediction tweet. I could word it more like: current ai memory is bad, and I think current+1 attempts I see right now are solving the wrong paradigm.

@whichsteveyp Agreed. But maybe we can get somewhere that's better than what we do now and is that not worth pursuing?

this social chatter has been absolutely fascinating to me. I’ve been day 1 red, but constantly told I’m a bad person for it in yaps. which is fine, but this image describes things so accurately what the problem reads to me 🤣

I think we don’t fully understand the good and bad parts of human memory enough to try and translate that to models. One example: Do you think a model could step away and go for a walk and come back and suddenly have an aha moment injected randomly from a previously *seemingly* unrelated node? Or that models could spend 8 hours hallucinating about not having their clothes on and then wake up with a huge aha moment? I don’t think we can just run a background agent(s) and say “interject the main thread with memories” - and I also think “let me go try to walk several graphs to find what I think I need” will work currently.

@whichsteveyp How would you solve it?

@DavidKPiano @kenwheeler man we tried that shit back in January I thought we learned you can’t do that rly

@kenwheeler Also the dumbest recent example of this I've been seeing is people trying to enforce strict workflows in skills

agent driven workflows with a sprinkling of deterministic tools are a fools errand. especially with subsidization fading, expect to see deterministic workflows with a sprinkling of agent nodes, doing what they’re actually good at.

@charlie_hess lmfao

tolkien-maxxing

who’s gonna tell him

@kentcdodds didn’t read it but, you’re way off base for sure

Tell me how wrong I am. Two rules: 1. Be nice about it. 2. Read the article.

legit vibecoded a thing today that, yea it’s pretty solid and solves a need. didn’t read a single loc or get in the weeds just “now add xyz.” pretty neat but, I wouldn’t ship this or try to maintain it at all without actually dissecting and reading / reworking lmao.

@kenwheeler see if anyone there will laugh at my joke

big day at the wheeler defense offices protecting your mother

@ryanflorence @DavidKPiano if there were non important parts why didn’t you just drop them :chatting:

@DavidKPiano the important parts

Just merged my own 23k line PR AMA

@DavidKPiano ✅

@whichsteveyp DMs open

yall I cooked the best joke I’ve ever cheffed, but I can tweet it

@kenwheeler Dario is a goon in that clip dog like, sama threw a slow and low over home today and Dario was like ok I can make this worse lmfao

they are fucking AMAZINGLY out of touch

dogshit take from brain dead lowbie, however, i promise you, if any linkedin class milquetoast get out of the shower to piss mf handed me a c suite role, you’d fuckin double your money

@jayphelps @ryanflorence 🤣 ok you got us big dog

@whichsteveyp @ryanflorence I mean, you actually can. He didn't clarify whether they'd let you play or not.

GPT has a new phenomenon that's driving me nuts and I don't quite know how to describe it. - Ask it if I can do something - It says "no you can't [incredibly twisted restatement of what I asked but also not at all what I asked]" - It then tells me how to do the thing wonderfully - And finishes with an insulting "But you can't just [stupid thing I never actually said]" It goes something like this: "Can I form and coach a youth soccer team for my kid and play in P/D level leagues? Or do I have to be part of a full club?" Then it says: "For official competitive teams in Utah, you cannot just form a random team and enter a league. "There is a lesser-known option: UYSA allows independent teams to enter leagues if they meet requirements. [...lists some simple requirements...] "But it’s not “show up with a group of kids on game day”—it’s more like running a small club team administratively. I never said just "show up with a group of kids on game day"! It does this to me with code too. It's so weird.

can you imagine working for a ceo this dorky

god Dario is such a goon

ok this take is way worse that Sama’s earlier. Sama definitely IQmogged this one lmao