Dhirendra Choudhary·5d agoI Gave a 2.7 GB Model 10 Three.js Prompts. 3 WorkedQwen3.5–4B-3D-Coder against a 17 GB coding model on the same ten scenes, on one MacBook. The specialist loses, narrowly
Dhirendra Choudhary·Nov 11How Good Are Local LLMs Now? The “Good Enough” FrontierI gave the 20-month-old model on my MacBook 14 graded tasks in early October 2026. It handled the boring 80%, and failed the 20% in a way I…
Dhirendra Choudhary·Oct 29Agent Skills vs MCP: One Is a Playbook, the Other Is a DoorSkills load a procedure into context. MCP connects the model to live systems. Most real setups need both, and mixing them up costs tokens.
Dhirendra Choudhary·Oct 13Why ChatGPT Can’t Count the R’s in “Strawberry” (And What Tokens Really Are)The famous strawberry fail isn’t proof the model is dumb. It’s proof the model never sees letters at all. Here’s what it sees instead and…
Dhirendra Choudhary·Oct 13RAG vs. Fine-Tuning vs. Just Using a Bigger Context Window: A Decision TreeThe first question isn’t “RAG or fine-tuning?” It’s ”is my problem knowledge or behavior?” Get that right and the rest of the decision…
Dhirendra Choudhary·Oct 13Mixture-of-Experts, Explained: Why “1 Trillion Parameters” Is MisleadingA trillion-parameter model doesn’t run like a trillion-parameter model. The gap between what it stores and what it uses is the whole idea…
InTowards AIbyDhirendra Choudhary·Oct 12What Happens Between You Hitting Enter and the First Token AppearingA response icon2A response icon2
InTowards AIbyDhirendra Choudhary·Oct 9Jev AI for Developers: Build a Confidence-Gated Ticket Router in PythonChoice, score and noul questions, the exact wire format, the limits TypeSafe documents, and what to do when the model isn’t sure.A response icon1A response icon1
InTowards AIbyDhirendra Choudhary·Oct 7How to Run a 125B-Parameter LLM on One RTX 4090 (Strata vs llama.cpp)The 4090 is the easy part. Here’s the RAM it needs, two ways to run the model, and what the speed costs youA response icon3A response icon3
Dhirendra Choudhary·Oct 6DeepSeek V4.1-Flash Is Here: What’s New, and Is V4-Pro Next?A 552B open-weight model that DeepSeek says beats its own flagship, at about a third of the output price.A response icon1A response icon1