Varun Nuthalapati·13h agoSiri Now Reads SKILL.md: Apple Turned App Intents Into an Agent Tool APIThe iOS 27.3 developer preview gives Siri AI the same building blocks every agent harness uses (tools, model-filled arguments, skills and…
Varun Nuthalapati·1d agoThree Mac Inference Servers Built for Agents, Not Chat — and No One Has Benchmarked Them Against…inferd, Locals and Rapid-MLX all shipped notable releases in the last two weeks. They make three different bets about what a Mac running a…A response icon1A response icon1
Varun Nuthalapati·3d ago64 Agents, One Mac Studio: How inferd Trades Single-Stream Speed for Fleet ThroughputA new Apache 2.0 engine for Apple Silicon is built for the case where agents arrive in crowds, come back every minute, and share most of…
Varun Nuthalapati·Nov 8Airbnb Raced a Fine-Tuned 2B Model Against GPT-6.1A production benchmark on one real classifier, what “max F1” hides, and the request rate at which self-hosting stops paying off
Varun Nuthalapati·Nov 8Stripe’s Agents Found 8% More Fraud. The Hard Part Was Knowing Which Labels to Trust.Inside Radar’s training scale-up, the delayed-label trap that forced a rollback, and why Stripe fine-tunes a 4B model instead of prompting…
Varun Nuthalapati·Nov 6One 740M Model to Search Your Code, Photos, Voice Memos, and Videos — LocallyEmbeddingGemma 2 puts four modalities into one 768-dimensional space. The interesting engineering is in what it lets you leave out.
Varun Nuthalapati·Nov 4Pinterest Replaced a Five-Stage Clustering Pipeline With a 4B Model That Writes Your Goals as JSONHow a distilled Qwen3 student, a prompt that acts as policy, and ~100 L40S GPUs turned “what is this person trying to do?” into a…
Varun Nuthalapati·Nov 1Mistral Large 4 Is Open. Running It on Macs Takes 338 to 590 GB, and the Math Explains WhyA trillion-parameter MoE meets MLX: three community builds, one unmerged PR, and the 512 GB Mac Studio that went on sale the day before the…
Varun Nuthalapati·Oct 27Netflix’s New Recommender Is an LLM That Never Writes a WordGenRec turns your viewing history into a prompt, scores the whole catalog in one forward pass, and never decodes a single token. Here’s how…
Varun Nuthalapati·Oct 25Claude Code Now Lets Plugins Rewrite the Agent From the Inside.Mods put your own code between the model and the permission system. That’s powerful, and it moves the trust boundary.
