Policy answers: What action should I take? For example, in chess/go playing, the policy network takes the board as input and outputs probabilities. 1. GPT policy: next-token policy (simplest) During normal GPT pretraining, there is no reward or environment. It is supervised learning:predict next token\text{predict next token}predict next token The loss is cross-entropy:L=−log⁡P(correct token)L=-\log P(\text{correct token})L=−logP(correct token) The optimizer then adjusts … Continue reading Policy vs Optimizer

MilestoneBig ideaAlexNet (2012)Learn from labeled examples (supervised learning).ResNet (2015)Train very deep networks effectively.AlphaGo Zero (2017)Learn through self-play with reinforcement learning.Transformer (2017)Replace recurrence with attention.Vision Transformer (2020)Apply Transformers directly to images. The evolution of modern AI can be understood as a series of breakthroughs in how machines represent information, learn patterns, and make decisions. 1. LeNet: … Continue reading From CNN to Transformer: How AI Learns to See and Reason

Abstract "We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision einforcement learning. Building on this multimodal foundation, … Continue reading Read Paper: KIMIK2.5: VISUAL AGENTIC INTELLIGENCE

When I first started learning networking, every acronym felt like another rabbit hole: HTTP, REST, TCP, UDP, SSE, WebSocket, RPC, gRPC, MCP... They all seemed related, but no one explained how they fit together into one coherent picture. Eventually, I realized they are simply different layers solving different communication problems. Level 1 — Two processes … Continue reading Computer Communication: From stdin to MCP

GPT-5.5 for hardest math and general reasoning.zerotwo Gemini 3.1 Pro for broad reasoning-heavy tasks.llm-stats Kimi K2.7 Code for agentic coding and long-horizon workflows.kimi+1 GLM-5.2 for another strong open-weight coding/agent option.llm-stats DeepSeek V4 Pro for huge context windows and document-heavy work.

In the AI era, more modern-day Da Vincis will emerge. People driven by relentless curiosity and an obsession with learning are no longer limited by time or access to knowledge. AI amplifies their ability to explore disciplines, master new skills, and turn ambitious ideas into reality at an unprecedented pace. The gap between curiosity and … Continue reading The Age of the Modern Da Vinci

Mixed precision means different parts of a model's computation use different numeric formats — e.g., weights stored in FP16/BF16, but certain ops (like accumulation) done in FP32 to avoid numerical instability. It's a speed/memory vs. accuracy tradeoff. This is a property of how you execute, not what the weights are. Native precision: the format the … Continue reading Terminology of LLM 01

If you know what you want and what you value, life gets simple. Not easy — simple. Naval said happiness is a skill. Skills require knowledge. Knowledge requires non-ambiguity. You can't master something you're vague about — and most people are vague about themselves. Ask anyone what they actually want, not the LinkedIn version. Most … Continue reading Remove Ambiguity Live a Simple Life

These five layers, each serving a different consumer at a different time: LayerFile(s)Who reads itWhenWhat goes in itAGENTS.mdAGENTS.md (repo root)Cascade/IDE agentEvery session, auto-loadedRules that prevent silent bugs. Short, imperative. "FF_V3 uses -R not -E", "deploy staging first", "docstrings <2KB". The guardrails.SKILL.mdSKILL.mdCascade/IDE agentWhen skill is invokedDeveloper's full project memory. Architecture, deploy ops, performance work, formula library, Snowflake patterns. The consolidated brain — everything an agent needs to work on the code.knowledge/kb/15 .md files + kb_index.jsonClaude Desktop (end user)Runtime, via get_index_knowledge MCP toolDomain methodology & templates. How … Continue reading Five Layers of Knowledge Files to Make My MCP Smarter

Claude Code has a full execution environment: Run Python/pandas between tool calls to transform data Read/write files for intermediate results Run bash (jq, csvkit, curl, etc.) Write and execute multi-step scripts Debug iteratively by printing intermediate state Claude Desktop has only: Call MCP tool → get response → call next MCP tool Every workflow that requires intermediate data processing (filter, join, dedup, add columns, merge results) … Continue reading Claude Code vs Claude Desktop