July 09, 2026

From Laptop to Blackwell: Scaling an AV World Model with Omnissa Horizon VDI

May 31, 2026

V-JEPA 2 on UCF-101: Backbone Generalizes, SSv2 Head Is the Bottleneck

I ran V-JEPA 2 frozen on 1,059 UCF-101 videos it was never trained on. A linear probe on the frozen backbone hit 98.2%. The SSv2 classification head fails on 42% of classes. The representations are not the problem.

May 09, 2026

What Billiards Reveals About JEPA: Strong Physics Representations, a Precise Planning Boundary

April 14, 2026

Validating Google's TurboQuant: 5x KV Cache Compression on the M5 Max

I tested TurboQuant (ICLR 2026) on a MacBook Pro M5 Max with 128GB unified memory. KV cache RAM dropped from 4.2 GB to 0.85 GB on a 32k-context Qwen-7B run, with 100% retrieval accuracy on a Needle-in-a-Haystack test. Here's what I learned.

January 12, 2026

Stability-Based Reasoning: A JEPA Approach to Sudoku

Most reasoning systems traverse invalid states and repair them. I built a JEPA-style Sudoku solver that never enters an invalid state at all. Reasoning emerges from maintaining latent stability rather than fixing errors.

November 10, 2025

Edge AI, JEPA Style: Sensor Prediction on a Jetson Orin Nano

A JEPA-inspired MLP with 7,376 parameters learns to predict sensor embeddings in latent space on a Jetson Orin Nano. Training loss converges in under 10 epochs. Inference runs at 1ms per step using 0.17MB of memory. The model never sees raw sensor values as a target.

October 06, 2025

Seeing Quantization in Action: Training ResNet-18 at FP32, FP16, FP8, and FP4

I trained ResNet-18 on a chest X-ray dataset across four precision levels on a MacBook Pro M2 Max and a rented NVIDIA B200. FP4 training completed in 12 minutes with 83-84% accuracy. FP32 took 50 minutes at 87%. Here is what the precision journey looks like end to end.