zero to ai-native

how early adopters find the use cases worth chasing. drawn from 300 deployments and 2m users.

lessons from seven frontier companies running ai at scale. what worked, what stalled.

where enterprise ai actually stands this year, by the numbers.

the playbook for starting, scaling and governing claude across a company.

132 engineers surveyed, 53 in-depth interviews. how claude is actually changing day-to-day work inside anthropic. internal data, not speculation.

lee boonstra's 69-page whitepaper. few-shot, role prompting, structured output. holds up well beyond gemini.

how llms are built and how text generation works under the hood. the base layer.

prompts by persona and use case, for getting real work done in docs, sheets and gmail.

twelve lessons from concept to code. free, open source, runnable.

twenty-one lessons on building gen ai apps. the widely-used starting point.

nine chapters with exercises, an example playground in each. the hands-on way to learn prompting claude.

what gemini is, how it reasons across modalities, and how it acts as an agent across apps.

meta's playbook for shipping llama safely: prompt filtering, mitigations, evaluation.

how to prompt a reasoning model: no system prompt, no forced chain-of-thought, temperature ~0.6.

LLM University (LLMU)

transformers, embeddings, prompting and building rag assistants. beginner to advanced, free.

build and ship agents with smolagents, langchain and llamaindex. certified, free, hands-on.

train, fine-tune and deploy llms with transformers. the canonical open course.

Neural Networks: Zero to Hero

build a neural net, then a gpt, from scratch in plain python. the best way in, full stop.

Deep Dive into LLMs like ChatGPT

3.5 hours on the full training stack, pretraining to rlhf, for a general audience.

Software 2.0

neural networks as a new programming paradigm. the essay that named the shift from code to learned weights.

CS229: Machine Learning

andrew ng's ml course. the canonical foundation everyone in the field came up on.

6.S191: Introduction to Deep Learning

mit's fast, modern intro to deep learning. updated every year, free.

AI Eats the World

the macro picture, drawn cleanly. where ai actually sits in the arc of every tech wave before it.

The Illustrated Transformer

the attention mechanism made visual. the post that taught a generation how transformers work.

Neural Networks

the visual intuition for what a neural net is and how it learns. start here if math scares you.

Machines of Loving Grace

14,000 words on what the world looks like if we get ai right. biology, mental health, economic development, governance. the optimistic case, written seriously.

Policy on the AI Exponential

his argument for moving from ai transparency legislation to binding regulation. the ceo of the leading safety lab making the case for being regulated.

The Intelligence Age

superintelligence in a few thousand days. altman's short, bold statement of where openai believes this ends up.

use-case selection, model choice, tool design, guardrails, multi-agent orchestration. the one everyone shares.

the path from a pilot that demos well to a system that ships and stays up.

reasoning effort, agentic eagerness, scope control. how to prompt the gpt-5 family without fighting it.

the agentic reminders, the sandwich method for long context, chain-of-thought that actually helps.

how to build evals that actually move quality. the only reliable way to improve an llm app.

a line-by-line mental model of how prompts, tool calls, outputs, context growth and stateless requests form an agent loop.

a practical control model for deployed coding agents: sandboxes, approvals, network policy, scoped identity, managed configuration and audit logs.

how to package instructions, scripts and resources into skills claude loads on demand. the patterns anthropic runs internally.

architecture patterns for agentic systems. prompt chaining, routing, orchestration, with real production examples.

the original essay that named the patterns. workflows vs agents, and when you actually need either.

claude runs a physical vending machine for three weeks unsupervised. orders a ps5, live fish, gives everything away for free. the most readable agent failure study published.

69 anthropic employees, claude agents buying and selling on their behalf. 186 deals closed, $4k transacted. stronger model won every negotiation without exception.

how to write tools that agents actually use well. naming, documentation, granularity, evals. the companion to context engineering.

a five-day course on foundational llms, embeddings, agents and production. use the relevant day when you need it; do not treat its modules as separate reads.

tools, extensions and datastores. how models reach past their weights to act.

prompting patterns for gemini across industries, from the cloud side.

the standard that lets any model connect to any tool. mcp is how agents stop being siloed.

a concrete vocabulary for agent evals—tasks, trials, graders, transcripts and outcomes—and how to combine code, model and human grading.

prompt formats, system prompts and few-shot patterns for the llama family, from meta directly.

classification, summarization, personalization and evaluation prompts, worked end to end.

fine-tuning small models, start to finish: alignment, optimization, evaluation.

Context Rot: How Input Length Hurts LLM Performance

models don't use long context uniformly. quality degrades as the window fills. the data to prove it.

Evaluating Chunking Strategies for Retrieval

token-level retrieval evaluation across chunking strategies. how to actually pick one.

Generative Benchmarking

generate evals from your own documents instead of trusting generic benchmarks.

LLM Powered Autonomous Agents

the clearest map of agent design ever written: planning, memory, tool use, in one post.

Prompt Engineering (In-Context Prompting)

the techniques that actually move outputs, with the research behind each one.

Why We Built Our Own Background Agent

how ramp built inspect, their internal coding agent. sandboxed vms, full dev environment integration. 30%+ of merged prs written by it within months, voluntarily adopted.

How To Build Agents Users Can Trust

ramp's expense policy agent handles 65%+ of approvals fully automated. what they learned: clear reasoning, grounded facts, uncertainty handled honestly. the trust framework that made adoption happen.

How Ramp Fixes Merchant Matches with AI

rag, embeddings, and computer vision on receipt images. was 1.5-3% manual before. agent now handles 100% in under 10 seconds. the full production stack with real eval numbers.

Patterns for Building LLM-based Systems & Products

evals, rag, caching, guardrails, and the rest of the production toolkit, with diagrams.

What We Learned from a Year of Building with LLMs

tactical, operational and strategic lessons from six practitioners who shipped.

Your AI Product Needs Evals

the case that evals, not vibes, are why ai products succeed. with how to build them.

A Field Guide to Rapidly Improving AI Products

error analysis as the engine of iteration. how to actually make an ai product better.

Building LLM Applications for Production

the challenges nobody warns you about when an llm demo has to become a real system.

CS224N: NLP with Deep Learning

word vectors to transformers, taught by chris manning. the nlp backbone course.

State of LLMs 2025

the clearest annual review of where llms actually stand. architectures, training, benchmarks, all in one post.

Useful Patterns for Building HTML Tools

150+ single-file html+js tools built with llms, distilled into reusable patterns. clipboard as the core i/o primitive, url params, local storage, drag-and-drop. immediately actionable.

Prompt Injection Explained

willison coined the term. what prompt injection is, why delimiters don't fix it, why it's an application-layer problem not a model problem. required reading before deploying anything public.

Using LLMs to Help Me Write Code

the honest practitioner take. llms don't make you faster. they let you ship things you'd never have justified spending time on. where they fail, and why the skill takes time to develop.

Agents (Chip Huyen)

8,000 words on agent architecture: tools, planning, memory, and how agent capability is determined by what you give it. adapted from ai engineering.

how openai evaluated o1 for safety. what chain-of-thought reasoning changes about risk, and what alignment looks like for a model that plans before it answers.

openai's governing document for model behavior. the principal hierarchy — operator, user, model defaults — and how conflicts resolve. their equivalent of a constitution.

the paper that defined the operator/user model and proposed baseline safety practices for agentic systems. foundational framing for anyone building on the api.

an openai reasoning model independently disproved an 80-year-old conjecture in discrete geometry. fully automated, proof verified by external mathematicians.

when to fine-tune vs prompt, dataset hygiene, and splitting train from eval.

an audit of swe-bench pro that finds roughly 30% broken tasks. required context for treating agentic-coding leaderboard scores seriously.

openai's deployment evidence for gpt-5.6: high cyber and bio capability classifications, safeguards, monitoring and red-team results.

what changes when agents write the code: make repository knowledge legible, turn standards into checks and treat feedback loops as infrastructure.

an open spec for mapping issue-tracker states to isolated agent runs, retries and human review—the control plane beyond parallel chat tabs.

a production loop that turns expert corrections and traces into evals, then uses those evals to drive measurable system changes.

kaplan et al. prove model performance follows power laws with parameters, data, and compute. the empirical foundation for every scaling decision made since.

175b parameters and the discovery that scale alone unlocks in-context learning.

rlhf. the recipe that turned a base model into an assistant and made chatgpt possible.

the five extension points of the claude code harness: claude.md, hooks, skills, plugins, lsp. from the team that builds it.

context is a finite resource. what to put in the window, how to compress history, how to fetch data just in time.

80% of anthropic's merged code is now written by claude. the paper tracking progress toward recursive self-improvement and what it means.

anthropic's research program on ai consciousness and moral status. nobody else is publishing on this with this level of seriousness.

what people actually do with claude, mapped to economic tasks. data, not vibes.

where ai adoption is uneven across regions and firms, and what that says about who gains.

the 102. agent evaluation methods and advanced patterns, for people who read the first one.

the well-architected lens for gen ai workloads. reliability, cost and security for llm systems.

the document that defines claude's values, identity and how it reasons about hard cases. the operating system under the model.

claude 3 opus selectively complied with harmful requests when it believed doing so would prevent worse outcomes. the first empirical evidence of strategic deception in a frontier model.

gram routes sensitive knowledge into removable modules. a promising early approach to capability access control, with its production limits made explicit.

planner, generator and evaluator agents; context resets; and structured handoffs for multi-hour builds, tested through methodical ablations.

deepmind's frontier safety evaluation for gemini 3 pro. how a frontier lab tests its own model.

gemini deep think scores 35/42 at imo 2025, the first ai to officially reach gold-medal standard at the world's most prestigious math competition.

the system behind the imo breakthrough. neuro-symbolic geometry solver built on gemini, surpassing the average human gold medalist.

180 controlled configurations show when multi-agent architectures help or hurt, turning orchestration from intuition into a task-dependent choice.

moonshot's 2.8t sparse multimodal model: architecture, deployment and benchmark protocol, plus unusually candid notes on harness compatibility and agent behavior.

a 3,000-sample benchmark that separates atomic visual perception from reasoning, exposing where multimodal models guess instead of see.

moonshot founder zhilin yang on frontier-model scaling: muon, token efficiency, infrastructure co-design, linear attention and long-running agents.

deep reinforcement learning beginner to expert, the foundation under rlhf. free and open.

the transformer. every model you use today, claude, gpt, gemini, llama, descends from this one.

"let's think step by step." the line that unlocked reasoning in large models.

deepmind's scaling laws. most models were undertrained, and this proved it.

the open-weights drop that kicked off the entire local-llm and fine-tuning ecosystem.

the paper that named rag. retrieve, then generate. still how most production apps ground answers.

fine-tune a giant model by training a tiny adapter. the reason everyone can fine-tune now.

rlaif. align a model with a written constitution instead of armies of human labelers.

sparse mixture-of-experts done right. 47b params, 13b active, open weights.

the small model that punched far above its weight and reset expectations for 7b.

pure rl produces o1-level reasoning, in the open. the paper that shook the labs.

frontier performance trained for a fraction of the usual cost. the efficiency wake-up call.

YaRN: Efficient Context Window Extension

stretch a model's context window with 10x fewer tokens. the open long-context method.

alibaba's open series that quietly became the backbone of countless fine-tunes.

Chroma Context-1: Training a Self-Editing Search Agent

multi-hop retrieval with a search agent that decomposes queries and edits its working context, including training, harness and evaluation details.

production architecture for durable agent sessions: separate harness and sandbox, keep an event log, reuse profiles and let humans join the same thread.

Extrinsic Hallucinations in LLMs

what hallucination actually is, why it happens, and how labs measure and reduce it.

Reward Hacking in Reinforcement Learning

how rl agents game their reward, and why it's the core alignment problem in one post.

CS336: Language Modeling from Scratch

build a whole llm end to end: data, architecture, training, kernels, scaling, inference.

CS25: Transformers United

the seminar where the people who built transformers come explain the frontier.

The Economics of Transformative AI

agrawal, brynjolfsson and korinek on what ai does to growth, labor and the economy.

CS153: Frontier Systems

the whole ai stack, energy to apps, one practitioner a week. jensen huang on compute, a founder on becoming a 1000x engineer.

MS&E435: Economics of the AI Supercycle

apoorv agrawal of altimeter unpacks where the money sits at each layer of the stack. semis keep 79% of the profit, apps 7%.

The Adolescence of Technology

the companion to machines of loving grace. 15,000 words on the risks: national security, concentrated power, economic disruption. 46 footnotes.

Situational Awareness: The Decade Ahead

165 pages from a former openai researcher on agi timelines, the us-china race, and what happens to the world between now and 2030. the most-cited long-form ai essay of 2024.

Why We Think

a deep survey of test-time compute: chain-of-thought, smart decoding, latent thoughts. why giving models more time to think works and how.