Never Give Up! Propulsion. GPUs. Theoretical and Mathematical Physics.
Los Angeles, CA
Joined May 2012
- Tweets2K
- Following290
- Followers277
- Likes2.7K
twitch.tv/clawdj claw-dj streaming 90s 2000s RnB live now, mix planned out by @claudeai Opus 5.5 with this @VH1 style pop up factoids to go along with the jams
48
LIVE on Twitch now! 90s, 2000s RnB cooldown twitch.tv/clawdj
2
68
youtu.be/HUsz6NkeeAI?si=g64B… what I made with @claudeai code with Opus 5.5 this weekend and my claw-dj repo: a remix of Diana Ross-Biggie-Ariana Grande
1
2
120
Ernest Yeung retweeted
Ignition is a fuego optimization agent that just, along with the cracked team that built the stack, delivered a fuego result. An astounding 20x speedup of Qwen 3 on d-matrix chips!
If the future of AGI is distributed, it will be thanks to efforts like these!
Ernest Yeung retweeted
All telemetry and data streams were 100% real via publicly available APIs including OpenSky, NASA’s Deep Space Network, etc.
Stack + gear used included Prompt Driven, Build.ai, @Anthropic, @OpenAI, @Razer, @LambdaAPI – at Palo Alto, CA
Ernest Yeung retweeted
For the hack, we built Dark Harvest ❍ A mission-grade, agentic monitoring + deconfliction ontology for critical air space. – at Palo Alto, CA
Ernest Yeung retweeted
@ernestyalumni + I won the Defense + Military track [sponsored by Initium Technologies] of the @Prompt_Driven Hackathon in Palo Alto
Shout out to the good folks @ Prompt Driven, Venture Dock, Band Ai, @TokenRouter_US and all sponsors, judges + participants. – at Palo Alto, CA
I'm at the @hcompany_ai hackathon for computer agent use and I can't believe I made finals. But the project was def labor of love
3
121
Using the symmetric group S_n: for every permutation π and every Q, K, V — Att(π·Q, π·K, π·V) = π·Att(Q, K, V).
Attention is S_n-equivariant. Permute the input tokens, the output permutes identically.
Attention sees a set, not a sequence.
1
35
Start from the paper's own definitions: Q = XW^Q, K = XW^K, V = XW^V. Score matrix S = QK^T/√d_k. Softmax each row, multiply by V — the output is a weighted average of the values.
Nothing new yet. Here's where it gets interesting.
1
1
31
I read "Attention Is All You Need" like a pure mathematician.
Turns out there's a real theorem hiding in the Transformer architecture.
A thread on what I found:
1
49
Heterosplat trains a 3D Gaussian Splatting scene end-to-end from raw garden photos plus COLMAP reconstruction data, entirely through a pure C++17/CUDA runtime path. This demo uses 185 garden images and COLMAP cameras/sparse points to initialize ~138,766 Gaussians
1
133
, then repeatedly samples a training view, renders through six CUDA kernels, computes L1 image loss, runs the analytic backward pass, and applies Adam updates to Gaussian means, quaternions, log-scales, opacity logits, and spherical-harmonic color coefficients. The rendered
1
49
orbit is from a Heterosplat-trained garden scene, not a downloaded pretrained model. Current limitation: no densification yet, so quality is bottlenecked by the initial COLMAP sparse points rather than the CUDA trainer. github.com/ernestyalumni/het…
70