
MSc Applied Math @polimi ⏳ Statistical learning by day | Local AI by night 2 x 5070ti | Ryzen AI 370 mini PC | MacBook Pro M4
Milan, Lombardy
Joined March 2022
- Tweets520
- Following156
- Followers235
- Likes1.5K
Pinned Tweet
Make 3D birds fly in your terminal. Now in v1.7.
Don't get hypnotized.
cbirds --3d
github.com/clainstone/cbirds
clainstone retweeted
‼️ Someone says they burned 8.9 billion tokens getting Claude Code to make a single strawberry in Blender, the free 3D modelling program.
The creator's own write-up puts it at 2 weeks, 3,958 prompts and 17,453 tool calls.
Software is both art and utility. AI is both an Industrial Revolution and a Renaissance.
Factories are for utilities, at AIE Code I talked about how we can make better art.
Vibe coding a game is fun until you don't know what your own characters do.
@trq212 on keeping creative control as models get better: readable specs, custom languages, and a software renaissance.
The Sweet Side of the Bitter Lesson.
Watch: youtu.be/xRJMiaUoD_8
Buying GPUs is the best thing you can do right now imo
You own your hardware, you own your intelligence, and you learn a lot along the way
Waiting for Qwen4-27B to leave it running 24/7
2
324
clainstone retweeted
🚨 BREAKING: Jensen Huang is reportedly considering founding a brand-new semiconductor company 🚨
He could raise tens of billions on day one, would have full founder control, and could build fast on his own terms. With the relationships he's built over decades, he could easily plug into the #AI ecosystem.
Imagine the demand and customer base lining up before his first chip even ships...
@grok Could Jensen Huang launch a competitor to $NVDA, attract tens of billions in funding, and lock in major customers?
clainstone retweeted
BREAKING: AMD stock, $AMD, crashes -13% in 30 minutes after Hunterbrook Media accuses the company of illegally selling chips to China.
Hunterbrook Capital also disclosed a short position in AMD.
MiniMax Music 3 just landed in llama.cpp ❄️
So I made 10 Christmas songs on my 2x RTX 5070 Ti, fully local:
- 90 s of music in 127 s per song
- 18 GB of VRAM with Q8_0 (BF16 doesn't fit)
- about 10.6 Wh per song, 21 min for all 10
Gotta say, way better than my first try this morning, you need to specify exactly what you want since the model is not so good at improvising. The one in the video is the 1950s big band swing. It's my favourite, probably only because the singer sounds a bit like Michael Bublé. Volume up 🔊
The other nine are at the link in the first reply.
Full prompt of the first song in the second reply.
2
1
219
All 10 songs and the setup: clainstone.com/blog/ten-chri…
PR: github.com/ggml-org/llama.cp…
GGUF: huggingface.co/ggml-org/Mini…
1
125
Full prompt of the first song:
[Tags]
Genre: 1950s big band swing, Christmas crooner. BPM: 132, swing feel. Key: B-flat major. Vocals: warm male baritone crooner, relaxed and charming. Arrangement: walking upright bass and brushed drums, muted trumpets and a saxophone section answering every vocal line, sleigh bells on the backbeat, celesta sparkles between phrases, big brass hits on the chorus, a short trombone solo. Mood: cozy, playful, champagne and snow. Production: vintage warm, wide room reverb, clear vocal up front.
[Lyrics]
[intro]
[verse]
The city's wearing silver and the cabs are running slow
There's a fellow by the bakery who's selling mistletoe
I bought a dozen paper stars and hung them on your door
'Cause baby, you're the reason that December's for
[chorus]
So pour a little moonlight, put a record on
We'll dance across the kitchen till the candles are gone
Let the snowflakes gossip on the windowsill
I'm staying here for Christmas, and I always will
[verse]
The radiator's humming and the punch is mostly rum
Your mother's on the telephone, she wants to know we'll come
[chorus]
So pour a little moonlight, put a record on
We'll dance across the kitchen till the candles are gone
Let the snowflakes gossip on the windowsill
I'm staying here for Christmas, and I always will
[outro]
I always will
55
All 10 songs and the setup: clainstone.com/blog/ten-chri…
9
Full prompt of the first song :
[Tags]
Genre: 1950s big band swing, Christmas crooner. BPM: 132, swing feel. Key: B-flat major. Vocals: warm male baritone crooner, relaxed and charming. Arrangement: walking upright bass and brushed drums, muted trumpets and a saxophone section answering every vocal line, sleigh bells on the backbeat, celesta sparkles between phrases, big brass hits on the chorus, a short trombone solo. Mood: cozy, playful, champagne and snow. Production: vintage warm, wide room reverb, clear vocal up front.
[Lyrics]
[intro]
[verse]
The city's wearing silver and the cabs are running slow
There's a fellow by the bakery who's selling mistletoe
I bought a dozen paper stars and hung them on your door
'Cause baby, you're the reason that December's for
[chorus]
So pour a little moonlight, put a record on
We'll dance across the kitchen till the candles are gone
Let the snowflakes gossip on the windowsill
I'm staying here for Christmas, and I always will
[verse]
The radiator's humming and the punch is mostly rum
Your mother's on the telephone, she wants to know we'll come
[chorus]
So pour a little moonlight, put a record on
We'll dance across the kitchen till the candles are gone
Let the snowflakes gossip on the windowsill
I'm staying here for Christmas, and I always will
[outro]
I always will
25
if you have to make an X profile managed entirely by an LLM at least put the thinking to the maximum otherwise the replies to the posts don’t even have a logical sense
@elonmusk pls do something
2
91
Worth looking at: @PrimeIntellect just introduced Model Factory. The idea is putting your whole lab on a single cluster, with zero handoffs between stages.
Train specialist models with async RL in isolated microVM sandboxes, deploy them, and feed real usage failures back into your next environments.
That's smart if you want to own the intelligence without spending months becoming an RL lab
Introducing Model Factory.
Train, deploy, and continuously improve your own models on one stack.
Run on-demand, on dedicated clusters, or bring your own.
Own your intelligence. Own your model factory.
68
I'm building a benchmark for LLM kernel development on AMD NPUs.
For months I've been writing kernels with IRON for an inference engine that runs RWKV-7 on the XDNA 2 NPU of Ryzen AI laptops. IRON is AMD's open-source toolchain for the XDNA 1 and XDNA 2 NPUs, the ones inside Ryzen AI laptops and Strix Halo mini PCs.
Why it makes a good test for LLMs:
- no scheduler, no dispatcher, no cache: the model has to place every byte of local memory and program every data movement itself
- no debugger to observe bugs
- IRON is not well known, so there's very little public code a model could have encountered in its training set
More details soon
1
2
119
I had the feeling Gemini 4 Carbon was unusually smart at reasoning. Neon is adapted from Carbon, and Google's card puts it ahead of Opus 5.5 and GPT-6.2 Sol on HLE-Diamond (71.3% vs 67.9% vs 64.0%) and RiemannBench (84% vs 80% vs 80%)
1
1
153
Model card: storage.googleapis.com/deepm…
Methodology: deepmind.google/models/evals…
47














