Senior software engineer Automation + applied AI (pharma) Contributor: MLX-Serve, TensorFold, ds4 Local inference · Real hardware · Builders
Joined September 2026
- Tweets1.1K
- Following491
- Followers2.2K
- Likes7.7K
Pinned Tweet
Building automation and applied AI systems for pharma. Contributing to local inference projects (MLX-Serve and TensorFold).
Obsessed with making models run faster and more reliably on real hardware.
Tinkerers and builders — what are you working on? I’d love to connect.
A world model is a simulator only if it ranks your policies correctly. Pretty video is not the test.
PrimeSim (open weights) ordered 8 robot policies exactly like the LIBERO physics sim: Spearman 1.00, mean gap 2.8 points.
RL trained only inside the model lifted real-sim success from 61.7% to 73.2%.
arxiv.org/abs/2611.04870
Parallel drafters waste most of their guesses because token 9 has no idea what token 8 picked.
Tencent's DSpine passes each neighbor's guess forward through all 5 draft layers. On Qwen3.8-27B it accepts 7.20 tokens per round on GSM8K vs 5.69 for DFlash2, and agent replays run 2.7x to 3.8x faster than plain decoding on GPU.
Apache 2.0: huggingface.co/tencent/Qwen3…
Asahi seems more like a weird vanity project for the developers than a genuine effort to bring Linux to Mac for the masses.
Asahi Linux is moving everything from the MIT licence to GPLv2-or-later.
Asahi is the volunteer project that reverse-engineers Apple Silicon Macs so they can run Linux. In today's Linux 7.3 progress report, developer James Calligeros explains why: people have been using the permissive licence to take the team's work. "In some extreme cases, in-development feature code was being lifted straight out of our personal forks and published without attribution or lazily rewritten with LLM assistance to obfuscate its origin."
The report points to WINE, which switched from MIT to LGPL in 2002 for the same reason: "Those who do not learn from history are doomed to repeat it."
The same report brings real wins too. M3 Macs now install without the EXPERT flag, Thunderbolt support (USB3 and networking for now) is shipping to users, and the team got GTA IV and Fallout: New Vegas running outside the muvm VM using Google's per-process page size patches.
Next-token prediction is an expensive way to learn what the world looks like.
Balestriero and LeCun trained a 109M text model on pairs of descriptions of the same scene. It lines up with a vision model (DINOv2) better than Qwen2.5-72B: 0.184 vs 0.154.
On the same corpus it needed 42x fewer samples than next-token to reach equal alignment.
It measures alignment, not chat. Still, the bet is clear: better training pairs beat more parameters.
arxiv.org/abs/2610.39302