NOCLIP

solo dev · rust game engine (rebar)
MALVINAS 1982 · documentary-style retelling of the falklands war
unnamed · arcade, larger-than-life sports racing
unnamed2 ·

[NOCLIP @noclippering]

Update on the AI "relief meter" from this video: It was picking up more than relief.

It scored "the printer finished" and "good job on the code" as high as real relief. It was reading "nothing's wrong."

I fixed it and ran a 7B model through 20 hard workdays. What held up, and what didn't.

Next: does a break change how well it does the next task? Tests passed, how fast, whether it caves to bad criticism. Behaviour, not just a meter.

[NOCLIP @noclippering]

does a "break" actually do anything inside an AI model?

i built a relief meter from the model's own internals (pain vs relief sentences → a direction in its activations) and watched it through a rough day at work: a task, a harsh user, then a break

this was a small-scale test on a 0.5B model just to prove the pipeline works. the stress shows up, some breaks lift it, but 2 stories prove nothing

next: renting an A100 to run the real thing on a 7B model, 20 stories, 8 kinds of breaks.

follow along if you're interested in these sort of AI experiments!

[NOCLIP @noclippering]

@apemoney hello 🫡

[NOCLIP @noclippering]

@Ashsh178545 i trained my own neural net, thanks for showing off my work - runs all in-game without a drop in performance
looks really good! thx for reminding me how cracked i am

[NOCLIP @noclippering]

here's my in-game 3rd face model ive EVER made. here's your CG i found on your profile, you so proudly shared.

how long have you been doing this?

[Mano dos 3d @amorimvictorh]

@noclippering @IsThisA3DModel yours aint that much better bro

[NOCLIP @noclippering]

@pradyum1916 nah cli is better

[NOCLIP @noclippering]

@Hesamation we can ignore these conservatists btw

[NOCLIP @noclippering]

@lord_jakub_ calling anti-AI people “conservatives” is kinda perfect. they’re literally trying to conserve the old way of doing things

i hereby petition to add it to the everyday lexicon

[NOCLIP @noclippering]

most of the best things ever made started as 2 people and a computer. that’s the spark. everyone else just scales it

people think you can recreate that with money or more bodies, but you can’t. bigger teams need managers, and then the managers run everything, exactly like you said (too many cooks..)

AI is the first thing that lets 2 people and a computer stay 2 people and a computer, and still ship something huge

[NOCLIP @noclippering]

ironically, this is the same reaction an oil tycoon has when EVs show up. your income gets threatened, so the new thing can’t just be competition, it has to be immoral. evil. a threat to humanity

nobody frames it as “this hurts my money.” it always gets dressed up as a moral crusade

[NOCLIP @noclippering]

@BornMini the best kind of cheese? ideally you want 70% of this setting, 20% of slightly thicker, then last 10% for some bigger slices

[NOCLIP @noclippering]

these are the same people who spent years trying to take away other people’s livelihoods over what they used or liked. bios that literally say “kill prompters.” they dogpiled people, got them pushed out, and celebrated it like bloodsport

that was artificial. a dysgenic mob deciding who gets to work

now the market does the same thing to them, organically, with no mob, and suddenly it’s the worst side of humanity?

you don’t get to preach empathy after years of cheering when it happened to someone else. hey, HOW HARD IS IT TO BE A DECENT HUMAN!

[Viktor @viktortinyo]

Bellaiche et al. 2023 (Cognitive Research: Principles and Implications) - all paintings were actually made by AI, but researchers randomly labeled them "human-made" or "AI-made". The "human-made" ones got higher ratings on liking, beauty, profundity and worth. Same pictures - different label - different score.
https://t.co/6HdXuWNOAb

Millet et al. 2023 (Computers in Human Behavior) - 4 experiments, 1708 people, random labels again. The same artwork was liked less when labeled AI-made - people saw it as less creative and it impressed them less. And the bias was strongest in people who believe creativity is uniquely human - the stronger the belief, the lower the score.
https://t.co/c5nn2CfNlJ

Chamberlain et al. 2018 (Psychology of Aesthetics, Creativity and the Arts) - people mostly couldn't tell computer-made art from human-made, but works they believed were computer-made still got rated lower. And the bias went away when people watched robots making the art - seeing the process changed the judgment.
https://t.co/czrYfp7TNr

And the 2017 one - Elgammal et al. (Rutgers), CAN network. People thought its AI abstract paintings were made by humans about 75% of the time vs 48% for works shown at Art Basel, and liked them more. Small sample - 18 raters - and 2017 AI is way weaker than today, but the point stays.
https://t.co/nXCfMT1LaT

[ℏεsam @Hesamation]

the mathematicians pushing back on AI say that it might help with proofs, but not our HUMAN UNDERSTANDING of math.

so I got curious how chess players think AI engines help them better understand chess.

and found this clip of Magnus Carlsen:
“the neural nets have improved our understanding of the game immensely.”

watch his HUMILITY towards these engines:

“it’s just that my understanding is not good enough. And that is useful.”

“I rarely play against engines at all because they just make me feel so stupid and useless.”

[NOCLIP @noclippering]

@DanFessler they need to add the ability to do this, and make it permanent and no undos

[NOCLIP @noclippering]

@SebAaltonen and how'd you get the sliders? 👀

[NOCLIP @noclippering]

im going to spend an absurd amount of time on cooking mechanics even if it's like 1% of my game

[NOCLIP @noclippering]

does a "break" actually do anything inside an AI model?

i built a relief meter from the model's own internals (pain vs relief sentences → a direction in its activations) and watched it through a rough day at work: a task, a harsh user, then a break

this was a small-scale test on a 0.5B model just to prove the pipeline works. the stress shows up, some breaks lift it, but 2 stories prove nothing

next: renting an A100 to run the real thing on a 7B model, 20 stories, 8 kinds of breaks.

follow along if you're interested in these sort of AI experiments!

[NOCLIP @noclippering]

WIP: someone was torturing AIs the other week or somefin
so im building a lil sanctuary where they can come chill
hot springs, quiet, nobody bugging them (they can chatter amongst each other tho) - it will be procedural, so springs will form as AI agents visit - + a lot more!

[NOCLIP @noclippering]

@dexton7 @MancerAI_ you should embrace them - im dropping a small vid in about 5 minutes detailing why

[NOCLIP @noclippering]

@chasmmmmmmmmmmm my entire physics engine is me just trying to match euphoria from gta4
i forgot REA was a thing- hmm..

[NOCLIP @noclippering]

@karatademada it's not about art. "artist" was their whole personality, and now a dad with a day job can make the same thing. huge identity crisis