Deep Learning
Deep learning is the subfield of machine learning built around neural networks with many layers — "deep" because of the stack of transformations between input and…
42 notes on LLMs, deep learning, and building things — written to be understood.
Deep learning is the subfield of machine learning built around neural networks with many layers — "deep" because of the stack of transformations between input and…
Chinchilla (Hoffmann et al., 2022) is the paper that corrected a mistake the entire field was making about how to train large language models. It asked a deceptively…
DPO (Rafailov et al., 2023) is the method I now reach for by default when aligning a model. It delivers what RLHF gives you — a model that prefers responses humans like…
These are my notes on the TCP chapters from Computer Networking: A TopDown Approach (Kurose & Ross). TCP is the protocol that makes the unreliable IP layer feel…
A while back I worked through a set of JavaScript exercises designed to expose the parts of the language people gloss over: function binding, named vs. anonymous…
Constitutional AI (Bai et al., 2022, arXiv:2212.08073) asked a pragmatic question: can we make a model harmless without relying on armies of human raters labeling toxic…
Computer vision (CV) is the field of getting computers to understand images — classify them, find objects, segment regions, describe scenes. For a long time it was…
A friend in the US asked me to help with a parentchild education website. During our conversation I learned there are still a fair number of IE users over there, so he…
Reflexion (Shinn et al., 2023, arXiv:2303.11366) solves a frustrating problem with LLM agents: they make the same mistake twice. A standard ReAct agent that fails a task…
Toolformer (Schick et al., 2023, arXiv:2302.04761) asked a deceptively simple question: can a model learn to use tools on its own, without humans demonstrating how? The…
ReAct (Yao et al., 2022) is the paper that gave us the recognizable shape of an LLM agent: think, act, observe, repeat. Before ReAct, reasoning (ChainofThought) and…
RetrievalAugmented Generation (RAG, Lewis et al., 2020, arXiv:2005.11401) is, hands down, the most widely deployed LLM architecture in production. If you've used a…
DeepSeekR1 (DeepSeek, 2025, arXiv:2501.12948) is the release that showed the world you can get o1level reasoning from an openweight model — and, more importantly, that…
ChainofThought (CoT) prompting (Wei et al., 2022) is probably the highestROI prompting technique ever published. It costs you almost nothing — a sentence in the prompt —…
InstructGPT (Ouyang et al., 2022) is the paper that made LLMs useful as assistants. Before it, GPT3 was a brilliant nexttoken predictor that rarely did what you asked.…
GPT3 (Brown et al., 2020) is the moment the world stopped thinking of language models as "text completers" and started thinking of them as programmable via prompts. The…
BERT (Devlin et al., 2018) is the model that made "pretrain on a huge corpus, then finetune on your task" the default playbook for NLP — and for a few years, almost…
Every large language model you use today — GPT, Claude, LLaMA, Gemini — is built on a single 2017 paper that, frankly, I did not fully appreciate when it came out.…
Scaling Laws (Kaplan et al., 2020) is the paper that turned "make the model bigger" from folklore into engineering guidance. Before it, nobody could say how much bigger,…
Tree of Thoughts (ToT, Yao et al., 2023, arXiv:2305.10601) takes ChainofThought and makes it deliberate. Instead of generating one linear reasoning path and hoping it's…
AutoGen (Microsoft, 2023) is a framework that treats multiagent conversation as the unit of computation. Instead of one prompt solving a task, you spin up several agents…
Generative Agents (Park et al., 2023) is the famous "AI town" paper — 25 LLMpowered characters who go to work, gossip, throw parties, and form relationships in a…
RLHF — Reinforcement Learning from Human Feedback — is the technique family behind every helpful modern chatbot. While InstructGPT (covered separately) popularized the…
Machine learning explainability is the set of methods for answering "why did the model predict that?" As models moved from simple linear regressions to opaque neural…
Mixtral (Mistral AI, 2023) is the model that made mixtureofexperts (MoE) practical and popular in the open ecosystem. The idea behind MoE is old, but Mixtral shipped it…
LLaMA (Meta, 2023) is the release that kicked off the openweight LLM ecosystem we now take for granted. Before it, if you wanted a capable model you used a closed API…
FlashAttention (Dao et al., 2022, arXiv:2205.14135) is one of those papers that looks like a systems optimization but quietly reshaped what models are even possible. The…
LoRA (Hu et al., 2021) is the technique that made it practical for ordinary people — not just labs with clusters — to specialize a large model. If you've ever downloaded…
is the userspace tool for configuring the Linux kernel firewall (netfilter). It looks arcane at first, but it's really just a set of tables, each containing chains of…
Softmax is the function that turns a model's raw scores into a probability distribution. It is tiny, it is everywhere, and it has a numerical trap that bites almost…
CLIP (Radford et al., 2021, OpenAI) is the model that connected vision and language in a way that actually transferred. Trained on 400 million (image, text) pairs…
Core Idea: The project is a specialized Model Context Protocol (MCP) server that connects AI agents directly to the WeChat Developer Tools. It transforms the miniprogram…
SQL is easy to start and surprisingly deep to master. The "advanced" parts — window functions, CTEs, query planning, indexing — are what separate "I can fetch rows" from…
Reinforcement learning (RL) is the branch of ML where an agent learns by trial and error, receiving rewards for good actions and penalties for bad ones — rather than…
Notes from Mankiw's Principles of Economics. The book's framing is deceptively simple — ten principles that supposedly explain most of what economics studies. What I…
SelfConsistency (Wang et al., 2022, arXiv:2203.11171) is, in my experience, the best accuracyperdollar trick in LLM reasoning. It builds directly on ChainofThought:…
Rotary Position Embedding (RoPE, Su et al., 2021) is the positional encoding used in LLaMA, Mistral, Qwen, and most openweight models you'll actually deploy. If you've…
Geospatial analysis is working with data that has a location — coordinates, regions, routes. It shows up everywhere once you notice it: delivery zones, disease spread,…
Part of the Tencent Cloud developerlab series.
In this tutorial, we'll detail how to deploy the OpenClaw AI assistant platform on different Raspberry Pi models. Whether you have a highperformance Raspberry Pi 4B or a…
In the rapidly evolving world of smart homes and IoT devices, one of the biggest challenges developers face is the fragmentation of platforms and protocols. Whether…
Over Kaggle competitions and real datascience projects I've collected a set of practical techniques covering data cleaning, visualization, and timeseries feature…