tim (@NERDDISCO) header
tim (@NERDDISCO) profile picture

cto @kalecomputer building a computer for curious kids

germany
Joined December 2011
Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)Photo by tim (@NERDDISCO)
Pinned Tweet
kale-img-1 fine-tuning results 1. @ostrisai / 8.10 / moss 2. @AmandineFlachs / 7.61 / moss 3. @suhaaspk / 7.20 / moss 4. @Assu_Bhadhu / 6.90 / moss 5. @max_pe2002 / 5.35 / hdo1 6. @hasnain__dev / 3.51 / sdxl (same prompt, seed, resolution)
paying $5k to whoever fine-tunes kale-img-1, an img model for kids dm me i'll send u the dataset u pick a base model we can license commercially u comment with ur result here ur result gets judged on: - recreating 5 imgs from the dataset almost 1:1 - making 5 new imgs in our style
4
2
2
21
2,178
my agents ping me via telegram
82
my agent desktop is getting bigger
1
172
kids are curious & creative
we put a camera on @KaleComputer so kids could ask questions about the world. they used it to take selfies and draw mustaches on each other. turns out curious kids are also really creative.
81
i added a canvas to my @t3dotcodes that u can only control by voice (local tts / stt) can access this from my phone via @netbird in tandem with the t3 code mobile app this is me working on @KaleComputer mixed with few mini games to play while waiting for agents
1
204
everyone starts on a blank canvas add what ever u need by saying it out loud this is ur new home, ur own desk
1
53
anyone wants to play with this?
33
holy moly
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2
155
it’s not about coding anymore, but shipping what ever u want without any limitations but your own mind
coding with ai is incredibly boring, change my mind developers are problem solvers, there is no fun or gratification in writing prompts for a machine to do the work coding exercises the brain, and when the novelty wears off from having ai generate features quickly... are we just glorified copywriters with tech knowledge?
2
2
208
everything @KaleComputer is based on the wayfinder skill
2
121
i always wanted to dj now i can bonus: it connected with my old hercules djcontrol compact over midi via usb, which lets me control the ui with an actual mini dj controller
1
182
i never had any use for this until now, as all the dj tools are complicated now i have one that just fits me 100% & i am mixing these two songs for the last 2h non-stop
43
just keep having fun
in light of the "programming is dead" talk going on, I really recommend listening to this @dhh episode with @lexfridman. here are my main takeaways, which left me motivated and extremely optimistic: - it's ok to grieve the way we used to code. it's part of accepting the new reality - don't waste time being pissed about it. put that energy into making the most of it - writing code isn't the bottleneck anymore (!!). humans are, so communication and coming up with ideas are crucial - one person can now build any app they always wished existed - if you took a year off from coding now, you'd probably be able to catch up in two weeks (remember how we had to keep up with react component lifecycle?) - don't try to predict model releases ahead, you'll drive yourself crazy. focus on what's possible right now - coding by hand was actually painful while we were doing it! looking back, it feels like an accomplishment, and that's why we tend to romanticize it dhh is very based and also an OG who's been programming for over 20 years, and despite that, he said the last few months were the most fun he's ever had with computers. so if he can be this optimistic about the future, you should be too :)
1
1
147
who is going to come with me to big berlin hack? will be there from 27 to 30
📣📣📣 Big Berlin Hack is back 🚀🚀🚀 And this time, even bigger: 🗓️ 28 & 29 November 👥 350+ builders from all over Europe 💰 Bigger prize pool Last time we had >1500 applications for Big Berlin Hack, so better be fast 👀 Registration is now open: luma.com/bbh Thanks to all of our partners, it's gonna be huge‼️
2
8
378
this time for real 🙈 teapaw.org/NERDDISCO/status/21111…
should i go to gtc in berlin next week? who would like to hang out?
33
this is how it literally feels like to code with agents
I think I found the perfect analogy for how to code with agents
1
2
256
ever wondered how to create a new voice? we had to figure this out last week we were using a voice from @ElevenLabs, but after talking to their team, we discovered that we couldn't use their service for our computer, which is designed for kids aged 6–12 so we tested a bunch of open-weight models to see which one delivered the best quality in the end, we picked drama-3 from @FishAudio here's how we created our own voice: 1. design a voice by prompting how it should sound 2. pick your favorite voice & clean up the audio until it sounds right 3. clone the voice using the cleaned-up audio (30 seconds worked best for us) 4. use the resulting voice.json to serve your new voice we're really happy with the result & hope the kids will love it as much as we do
4
1
9
488
finally doing inference optimizations again
2
5
262
making sure that our voice & img endpoints go brrr
25