Hi, I am Akash. I am a designer and creative technologist. I love working at the intersection of product design, creative, art direction, and technical writing.

October 2026

vLLM Paper Notes

Simple notes on vLLM and PagedAttention for efficient LLM serving with paged KV cache memory management and sharing.

October 2026

DeepSeek V4.1 Flash Paper Notes

Simple notes on DeepSeek V4.1 Flash and how it shrinks KV cache for million token agents while improving quality.

September 2026

DeepSeek Harness Paper Notes

A runtime theory for live plugin composition with undoable effects and reactive dependencies, built as Cordis and proven on thousands of Koishi plugins.

September 2026

Ingressing Minds Paper Notes

Michael Levin on hidden patterns that shape bodies and minds, from bioelectric memories and Xenobots to sorting algorithms that act smart.

August 2026

Continual Harness: Agents improve their own tools, skills, and memory mid-run

Continual Harness: agents that improve their own tools, skills, and memory mid-run, studied through Pokemon gameplay.

August 2026

From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement

RLSVR and SpyRL: turning open-ended tasks into self-verifiable games so models improve without judges or reward models.

July 2026

DeepSeek - V4 Notes

DeepSeek-V4: making million-token context practical with hybrid compressed attention, hyper-connections, and the Muon optimizer.

July 2026

Purified On-policy Self-distillation Paper Notes

Why plain OPSD damages long chain-of-thought reasoning, and how purified OPSD preserves reflective epistemic markers.

July 2026

Reinforcement Learning via Self-Distillation Paper Notes

SDPO: self-teacher policy optimization that fixes the sparse credit-assignment bottleneck of RLVR for code and math.

July 2026

On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes

GKD: distilling large models into smaller ones by training on student-generated outputs instead of fixed datasets.

June 2026

Introduction to Reinforcement Learning Notes

A ground-up primer: MDPs, TD learning, Q-learning, policy gradients, actor-critic, GAE, and the RL taxonomy.

June 2026

RL Foundations and PPO Notes

From RLHF pipelines to the PPO clipped objective: why RL follows SFT and how PPO training loops work for language models.

June 2026

On-Policy Self-Distillation for Large Language Models

OPSD: one model acting as its own teacher and student for better reasoning, using privileged contexts for dense token feedback.

June 2026

Self-Distillation Enables Continual Learning

Beating catastrophic forgetting with self-distillation fine-tuning that preserves in-context abilities while learning new tasks.

June 2026

Adapting the Interface, Not the Model

Notes on LIFE-HARNESS: fixing the model failures by adapting the runtime interface instead of retraining the model.

June 2026

AGENTS.md Paper Notes

Do AGENTS.md-style repo context files actually help coding agents? Notes on the paper and its AGENTbench evaluation.

June 2026

CaMeL: Computer Use Agents

CaMeL: system-level security for computer-use agents via a dual-LLM design that separates planning from acting.

June 2026

DeepDive Paper Notes

DeepDive: training open-source deep-research agents with auto-generated hard questions and multi-turn RL.

June 2026

DeepSeekMath Paper Notes

DeepSeekMath: how a 7B open-source model approached GPT-4 math performance, and the GRPO algorithm behind it.

June 2026

DeepSeek-R1 Paper Notes

DeepSeek-R1: teaching a model to reason through pure RL with no supervised traces, then distilling it into smaller models.

June 2026

Natural-Language Agent Harnesses

Natural-language agent harnesses: writing agent runtime logic as editable text instead of code, and what the experiments show.

June 2026

Recursive Language Models

Recursive Language Models: handling huge contexts by letting the model explore the prompt with code and call itself on pieces.

June 2026

SFR-DeepResearch Notes

SFR-DeepResearch: the Salesforce recipe for autonomous single-agent deep research via single-turn reformulation and RL.

October 2025

Process Reward Model in RL Training

Outcome vs. process reward models in RL training: why rewarding each reasoning step helps, and why steps, annotation, and reward hacking make PRMs hard.

July 2025

Model parallelization techniques

Pipeline, tensor, and sequence parallelism for faster LLM inference, and how combining them scales large models to longer inputs.