knowledged.to

Selected notes of Subhash Chandran on ML, engineering, and assorted curiosities. Generated from a knowledged-managed repository.

LoRA (Low-Rank Adaptation) in AI

A new section for Fine-Tuning Techniques is created to hold the LoRA document, and the LLM Architecture section is de-duplicated.

Function Calling Support in LLM Models

Explains function calling (tool use) in LLMs: how models emit structured requests to invoke external functions, the request-execute-return loop, provider support, and practical reliability notes.

Clarity: The New Programming Language in the Age of AI

Rainer Stropek argues that as AI automates coding syntax, the new core technical skill is clarity — the ability to precisely describe intent, set constraints, and verify results. Programmers are the "canary in the coal mine" for all knowledge workers facing this shift.

What is Speculative Decoding?

Explains speculative decoding, which pairs a small draft model with a large target model to accelerate LLM inference without changing outputs.

Mixture of Experts in AI

Explains sparse Mixture-of-Experts (MoE) architecture with conditional computation, router/gate mechanisms, load balancing, and trade-offs vs. dense models.

Local + Frontier Model Collaboration Patterns in Open Source Harnesses

As local LLMs improve, harnesses are learning to pair them with frontier models. A look at the four collaboration patterns already shipping in open source.

Autoregressive Image Generation

Explains autoregressive image generation as sequential visual-token prediction using Transformer-style next-token modeling.

What is a Diffusion Model?

Explains diffusion models as generative AI systems that learn to create data by reversing a noising process.

RLVR vs. the Agent Loop: Training-Time vs. Inference-Time

Distinguishes RLVR as training-time weight updates from inference-time agent verification loops.

The Modern LLM Training Pipeline

Explains the four-stage modern LLM training pipeline from pre-training through verifiable-reward RL.

The Modern LLM Training Pipeline

Explains the four-stage modern LLM training pipeline from pre-training through verifiable-reward RL.

Where RL Fits: Training vs. Inference in the LLM Pipeline

Explains that RL in LLMs is a training/alignment stage, not inference, with pipeline context.

Reinforcement Learning (ELI-Teen Explainer)

Teen-friendly explainer of reinforcement learning agents, rewards, exploration, delayed rewards, and applications.

Coding LLM Training with SFT and Verifiable RL

Explains scripted coding-LLM training with teacher traces, synthetic bugs, tests, SFT, and verifiable RL.

Prefix Caching in AI

Explains prefix caching for reusing attention KV computations to speed up shared-prefix AI inference.

Shiva Nataraja: Lord of Dance

Image-focused guide to Shiva as Nataraja, explaining dance posture, symbols, Apasmara, and ring of fire.

Jevons Paradox

Explains Jevons paradox: efficiency gains can increase total resource consumption through rebound effects.

MCP vs Plugins in Claude Code: which to use when

MCP vs plugins: layered distinction (external systems vs assistant-extension packaging); both cross-surface across Claude Code, Cowork, and the Claude apps

World Models in AI

Explains AI world models as internal predictive representations for planning across RL, LLMs, and robotics.

Jailbreaking LLMs: A Security Researcher's Field Guide

Field guide to LLM jailbreaking attack surfaces, threat modeling, defenses, and responsible disclosure.