Pinned Tweet
I trained an AI model to autocomplete my piano prompts.
Here’s what happens when I give it the first few notes:
Here’s what happens when I give it the first few notes:
6
6
36
2,761
The Jev Doom demo by @CompleteSkeptic is really cool, so while I’m waiting for access I tried the same thing with Gemma 4 E2B from @googlegemma.
Mine takes 640×480 frames directly, with no intermediate text description of the scene. It shares the image prefill across batched ads the controls directly from logits.
It sort of works.
Mine takes 640×480 frames directly, with no intermediate text description of the scene. It shares the image prefill across batched ads the controls directly from logits.
It sort of works.
3
6
10
627
@CompleteSkeptice@googlegemmahared instructions as the prompt. Gemma processes that once, and we cache the resulting keys and values.
For each JSON field, we append a short question to an independent continuation of that cached prompt. All those questions run together in one batch.
Instead of generating the answers autoregressively, we read the logits for a fixed set of answer tokens and map them to the allowed enum values. A boolean is simply an enum with two values.
For each JSON field, we append a short question to an independent continuation of that cached prompt. All those questions run together in one batch.
Instead of generating the answers autoregressively, we read the logits for a fixed set of answer tokens and map them to the allowed enum values. A boolean is simply an enum with two values.
1
0
2
117
@CompleteSkeptic @googlegemma The schema here is very basic:
{
move: "none" | "forward" | "backward";
strafe: "none" | "left" | "right";
turn: "none" | "left" | "right";
fire: boolean;
use: boolean;
}
{
move: "none" | "forward" | "backward";
strafe: "none" | "left" | "right";
turn: "none" | "left" | "right";
fire: boolean;
use: boolean;
}
0
0
2
75
Training the next RollTab to play the piano bass line live while you play the melody, so you can actually jam together instead of waiting for it to autocomplete after you stop.
I used Gemini to score the last version because validation loss wasn’t a very good proxy, but it enough for bass lines.
So I’m back to blind listening: the same 25-second melody, three bass lines, over and over and over again.
I used Gemini to score the last version because validation loss wasn’t a very good proxy, but it enough for bass lines.
So I’m back to blind listening: the same 25-second melody, three bass lines, over and over and over again.
0
0
0
94
Say "sshhh" to drop your Mac volume 20%. 🔈 on!
Tried a tiny trained model and a dumb heuristic that looks for sustained treble energy.
With headphones, and very limited training data, the heuristic actually wins.
Tried a tiny trained model and a dumb heuristic that looks for sustained treble energy.
With headphones, and very limited training data, the heuristic actually wins.
4
1
7
221
RollTab now supports MIDI out! 🎹
Send the AI’s performances to external MIDI devices, or straight into iOS apps like GarageBand.
Send the AI’s performances to external MIDI devices, or straight into iOS apps like GarageBand.
0
0
3
107
Play a few notes on your piano → AI continues what you played in real time, on your iPhone. 🔊 Unmute!
New RollTab release: five new sampling methods alongside top-k. Mirostat v2 has been a surprisingly good fit for music.
New RollTab release: five new sampling methods alongside top-k. Mirostat v2 has been a surprisingly good fit for music.
0
0
5
139
Been testing different sampling methods for my MIDI generation model.
Top-k still works really well, but Mirostat v2 is surprisingly competitive.
Top-k still works really well, but Mirostat v2 is surprisingly competitive.
1
0
3
138
Then I realised: all my DPO training samples were generated with top-k.
So I may have accidentally post-trained the model to be particularly good at top-k sampling.
Needs more testing!
So I may have accidentally post-trained the model to be particularly good at top-k sampling.
Needs more testing!
0
0
1
55
5 days and my first App Store submission is still just “Waiting for Review” 🫠
Anyone else seeing long waits on first submissions lately? How long did yours take?
Anyone else seeing long waits on first submissions lately? How long did yours take?
2
0
2
208
The Power Mac G4 got export-controlled in 1999 for crossing some compute threshold. Apple put it in their ads.
Now Fable
too powerful to ship → too powerful to autocomplete.
Now Fable
too powerful to ship → too powerful to autocomplete.
0
0
0
95
Simon Edwardsson retweeted
V7 Go now connects to all your favorite apps via MCP Integrations
→ 2–3 click setup
→ Agents pick the right actions on its own
→ Results that look native to each app, not just a wall of text
→ Works at massive scale across agent tables
Learn more: v7labs.com
→ 2–3 click setup
→ Agents pick the right actions on its own
→ Results that look native to each app, not just a wall of text
→ Works at massive scale across agent tables
Learn more: v7labs.com
0
5
12
723
Simon Edwardsson retweeted
News! @airstreet has raised $232,323,232 for Fund III to back AI-first companies from the earliest stages in the US and Europe.
Now the largest solo GP venture firm in Europe.
Our third epoch begins today. Join us!
Now the largest solo GP venture firm in Europe.
Our third epoch begins today. Join us!
184
59
1,032
180,680