Beyond Tokenmaxxing
saved because it is worth a closer look.
readmanav's notebook
tensor boy's picks: hackathons, repos & reads worth your time
a running notebook from manav. what i'm reading, building and betting on, plus the odd hackathon i'd clear a weekend for. the repos and reads carry the date i last checked them. no algorithm, just me.
karpathy: zero to transformer
checked july 4, 2026: still the cleanest fundamentals spine from neural nets to transformers before you start wiring agents.
readkarpathy/nanochat
checked july 4, 2026: small assistant-era training repo. useful because you can read the whole thing instead of hiding behind a framework.
readanthropic: building effective agents
checked july 4, 2026: practical taxonomy for workflows, agents and tool use. read this before turning every feature into a fake autonomous loop.
readopenai agents sdk
checked july 4, 2026: first-party Python SDK for agents, tools, handoffs and tracing. good when the hackathon sponsor stack already points at OpenAI.
readlangchain-ai/langgraph
checked july 4, 2026: graph and state machines for agents that need retries, branches and human stops. heavier than a weekend script, useful for serious demos.
readpydantic/pydantic-ai
checked july 4, 2026: structured Python agents with Pydantic-shaped outputs. good when correctness and typed handoffs matter more than multi-agent theater.
readbrowser-use/browser-use
checked july 4, 2026: browser-agent primitives for demos that touch actual websites instead of pretending a tool call happened.
readbrowserbase/stagehand
checked july 4, 2026: browser automation with AI-assisted actions on top of Playwright. useful for visible browser-agent demos, but keep deterministic fallbacks.
readollama/ollama
checked july 4, 2026: fastest local-model demo path. use it when offline-first, privacy or cheap iteration matters more than top benchmark scores.
read