Ilia Breitburg
banner
breitburg.com
Ilia Breitburg
@breitburg.com
Game engines are O(n) in scene complexity. Diffusion models are O(1), so the same cost whether you’re rendering an empty room or a million polygons. What if you made the engine differentiable and optimized the diffusion model against it directly, rather than sampled frames? Has anyone tried it?
January 17, 2026 at 2:16 PM
Reposted by Ilia Breitburg
I recommend reading The Math of Why You Can't Focus at Work.

Excellent visualization about how we lose valuable focus time because of interruptions and what we can do about it.

justoffbyone.com/posts/math-o...
The Math of Why You Can't Focus at Work
Interruptions, recovery time, and task size: three numbers that determine if you'll get real work done. Interactive visualizations show the math behind bad days.
justoffbyone.com
January 15, 2026 at 1:57 PM
The perfect Claude Code machine acquired
January 14, 2026 at 10:09 PM
My latest PR to pebble-tool adds the ability to press emulator buttons via CLI, so now Claude Code can build complex apps and navigate PebbleOS fully autonomously! @repebble.com @rebble.io @ericmigi.com
January 6, 2026 at 12:25 PM
@ericmigi.com Congrats on the launch! I was wondering why you guys ditched the speaker in the new Round? This is a dealbreaker for me :(
January 2, 2026 at 3:17 PM
Here's to the crazy ones.
January 2, 2026 at 1:56 PM
Releasing Wrist & Biases for @repebble.com. Now fellow ML engineers and researchers can obsess over eval/loss right from their wrist.

apps.repebble.com/en_US/applic...
December 31, 2025 at 3:56 PM
December 22, 2025 at 11:56 AM
Reposted by Ilia Breitburg
ignore the title about caching, this is the best explanation of how LLMs work, period
Prompt caching: 10x cheaper LLM tokens, but how? | ngrok blog
A far more detailed explanation of prompt caching than anyone asked for.
ngrok.com
December 21, 2025 at 3:23 AM
As sexy as toolbars get
December 21, 2025 at 2:57 AM
It turns out that if you add Recents to your dock, the icon hasn’t been updated for over a decade
December 20, 2025 at 9:11 PM
That's what I call a "base" model (of 2025). Models pre-trained on data post-2022 are magically aligning themselves
December 19, 2025 at 2:52 AM
Reposted by Ilia Breitburg
Someone is training neural nets on their new Pebble 2 Duo, of course...

github.com/breitburg/pe...
October 28, 2025 at 3:27 AM
Designed a fun thing :)
July 9, 2025 at 4:06 PM
Should I be seeing this? lol
June 18, 2025 at 5:11 PM
Claude 4 claims to have subjective experiences
May 26, 2025 at 12:12 AM
Claude 4 is (not) conscious
May 25, 2025 at 11:04 PM
I tried switching from Claude to open-source LLMs for 10 days. It's really disappointing. On benchmarks they're great, but in most practical tasks the alignment is really not there.
May 9, 2025 at 7:06 PM
insect54
www.flickr.com
May 6, 2025 at 6:58 AM
I estimated the token consumption for my Claude usage within the last month to figure out what the API cost would look like:

251 chats in April
Total output: 7,704,378 symbols ≈ 1,926,094 tokens
Total input: 66,016,919 symbols ≈ 16,504,229 tokens

So is $20/month for AI worth it?
April 29, 2025 at 4:09 AM
I've long searched for a decent LLM-agnostic chat app. @raycast.com is excellent, but is proprietary. Open WebUI comes really close, but doesn't feel native and lacks a mobile app. Yesterday I got fed up with that, so I recreated ChatGPT desktop app but without vendor lock-in.
April 26, 2025 at 12:08 AM
Not wrong
April 1, 2025 at 1:45 PM
Gotta prepare for iOS 19 in advance
March 29, 2025 at 10:46 PM