Andrew Ng

Ng argues coding agents are collapsing the value of language-specific syntax: a backend and ML engineer now ships frontend TypeScript. The counterweight he names is conceptual depth — learn one language properly, then let the model translate the concepts into the rest.

A contrarian case that carbon, electricity-price, and water objections to data centers are overstated — hyperscaler PUE near 1.2 against 1.5-1.8 on-prem, a Berkeley Lab finding that load growth has tended to lower retail rates, water use dwarfed by golf irrigation — with local planning as the real c

Agentic Document Extraction adds field-level extraction: point it at an invoice or medical form and get named fields back against a schema. If you'd rather not write the schema, one sample document plus a plain-language prompt generates it.

How the Model Context Protocol removes per-use-case integration work: an MCP client inside the application talks to servers that expose tools, resources and prompt templates, run either as a local subprocess or as a remote process. Plus what building one involves.

With coding agents compressing build time, the scarce resource becomes product judgment. Ng describes teams pulling engineer:PM ratios toward 1:1, then argues that even that leaves a handoff bottleneck — the fastest AI-native teams have engineers who understand users and decide directly.

Ng argues LLMs are general but far less general than the hype implies: after pretraining consumed the public web, progress has become piecemeal, with labs targeting individual tasks. The practical consequence is that breadth of competence is now assembled, not emergent.

Data silos stop being a procurement footnote once agents can correlate signals across systems. Ng describes a SaaS vendor pricing an API key to his team's own customer data at over $20,000, and argues for selecting software by whether your agents can actually reach the data.

A course on document extraction that starts from OCR's failure modes — lost reading order, orphaned table headers, uncaptioned charts — then builds agentic pipelines that parse pages visually, map fields to a schema with bounding boxes for validation, and feed event-driven RAG.

Ng pushes back on the advice to skip learning and just ship, citing candidates who rebuilt standard RAG chunking and agent evaluation techniques badly. His prescription: courses plus deliberate building, with research papers as an optional third leg.

A course on web agents that reason over both screenshots and DOM structure to act on pages. It centers the compounding-error problem in long action sequences and covers AgentQ's self-correction, with builds that scrape sites, fill forms and chain multi-step tasks.

StatQuest's Josh Starmer takes attention down to implementation level: word and positional embeddings, the Query/Key/Value matrices, the math of self-attention, and how masked self-attention in the decoder differs from the encoder's context-building attention — coded in PyTorch.

An argument that running agents in parallel is emerging as a distinct scaling axis next to training and test-time compute — cheap tokens plus latency pressure are pushing research fetches, multi-agent coding across worktrees, and other workflows from sequential to fan-out.

Databricks and a DSPy co-lead teach signature-based programming, MLflow tracing to see how each pipeline stage performs, and the optimizer that tunes prompts — few-shot examples included — against evals, then re-tunes the system when you swap in a newer model.

A short course with Jay Alammar and Maarten Grootendorst tracing language representation from bag-of-words through Word2Vec to transformers, then dissecting a transformer block stage by stage with runnable code for attention and the feedforward layers.

LangChain's Harrison Chase teaches memory architecture for agents: semantic, episodic and procedural stores, updated either in the hot path or by a background helper agent. The build is an email assistant that learns user preferences and evolves its own instructions.

Ng's AI Startup School talk on building fast with AI: why speed is the startup advantage, where opportunity sits across the AI stack, the shift toward agentic systems, and the specific execution habits AI Fund uses to shorten the loop from idea to running product.

Ng argues the fastest-moving agent teams are the ones running disciplined evals and root-cause error analysis rather than chasing new tooling. Part one of a two-part letter on locating where an agentic pipeline actually breaks before attempting fixes.

Multi-agent systems break when every team's framework choice needs its own integration. ACP puts agents behind a unified REST interface hosted in an ACP server, with a defined agent lifecycle — and a comparison against MCP and A2A that clarifies which layer each occupies.

A Windsurf-built course on the inner workings of IDE coding agents: how they combine human-action tracking, tool integration and context management to stay aligned with intent, and why that — not the shared underlying model — is what distinguishes one assistant from another.

Ng proposes the Turing-AGI Test: a subject, human or machine, gets a computer with internet, browser and Zoom, then a judge designs a multi-day work experience — training, task execution, feedback — and the system passes if it performs like a skilled hire.

LlamaIndex's VP of developer relations teaches event-driven agent workflows — branching, looping and concurrent execution triggered by completion events — applied to filling a PDF form: parse the blank fields into questions, answer them from a RAG index, review with a human in the loop.

Ng's read on two access decisions landing at once: Claude Fable 5 shipping terms that bar building competing LLM technology, alongside reports of silent, unannounced output degradation for users detected doing LLM research. His conclusion is that buyers are hedging toward access nobody can revoke.

A walk through the three common post-training methods and what each one costs you in data: supervised fine-tuning on ideal outputs, DPO on chosen-versus-rejected pairs, and online RL on reward scores. Post-training is where a token predictor becomes an assistant.

The open-source aisuite package now handles tool use with a single command instead of provider-specific scaffolding, generating function descriptions from Python docstrings and behaving uniformly across multiple LLM vendors.