anik jha·1d agoI Fit a 122B AI Model on a 16 GB GPUI removed three quarters of Ornith-1.5–122B’s “experts”. It started repeating itself. Two fixes got it coding again.
anik jha·4d agoYour AI Sleeps Through Its Own Reset. One Line Fixes It.A tiny hello that keeps your Claude Code and Codex five-hour windows running back to back, so your next reset is always closer than you…
anik jha·Oct 18I Deleted 6,912 Experts From NVIDIA’s Nemotron Bolt for a 16 GB GPU93–94% of two coding benchmark scores retained. A 13.6 GiB peak. Download the weights or explore the pruning code.
anik jha·Sep 29The Most Expensive Word in Claude Code Is “Continue”Stop letting it resume itself after a usage limit. Make it start over. One resume wrote 578,016 tokens before doing any work.A response icon1A response icon1
anik jha·Jul 30There’s a New True Local King in TownEvery few months a model shows up that’s technically open‑weight and genuinely too big to actually run locally — this month it’s Kimi K3…
anik jha·Jul 10The $1 Trick That Beats $10M in AI Fine-TuningMicrosoft Research just proved you don’t need to touch the model at all. A self-optimizing text file does it better — and it runs while you…
anik jha·Jul 5NVIDIA Just Made Diffusion Models Practical for Text: The TwoTower BreakthroughWhat if your LLM could generate text the way Stable Diffusion generates images — not one pixel at a time, but by refining an entire canvas…
anik jha·Jul 4DeepSeek Did It Again: The Speculative Decoding Revolution of DSparkDeepSeek did it again. Just when we thought we had hit the physical limits of LLM serving, this insanely resourceful team has ignited a…
anik jha·May 25Four Ways to Make LLMs Stop Being So Slow: A Comparison of Inference Acceleration TechniquesYou’ve trained your model. You’ve fine-tuned it on your data. You’ve even gotten it to stop apologizing for everything. Now you want to…
anik jha·Apr 22Stop Letting Your AI Agent Browse Unsupervised. Here’s WhyThe Setup: A Robot That Reads Papers So I Don’t Have To