Thanks & what I learned

Yes — if you can already follow the engineering but keep getting caught out on what to build and in what order. That is a different skill, and almost nobody teaches it.

Jennifer Liu runs it, and the structure is the value. She takes you through a way of approaching AI product building rather than a list of technologies — the six levers and the 4D Method — and that is what set the tone for me: think in a structured way instead of scattering prompts at a problem and hoping. The exercises carry it the whole distance, from incubating an idea, to translating it into a PRD, to turning that PRD into a product.

Rohan Varma leads product for Codex at OpenAI, and what he described about the pace they build and ship at reset my sense of what "fast" means.

Henry Shi is technical chief of staff to a co-founder at Anthropic, and brings a dimension you will not find elsewhere.

What makes it work is that these are people with their hands in the actual work, at the companies building the models most of us use every day. The level at which they think about AI is genuinely different, and being in the room for it is worth it on its own.

Yes — particularly if you already use coding agents daily and suspect you are getting maybe a third of what they can do. That was me.

I have taken it more than once, and it cost nothing to go back. You enrol once and can then join any later cohort you like — there is no second fee. That makes it closer to a community than a course with an end date. Each cohort brings a different dimension: new tools, and a fresh run at taking something from a prompt to a product that is actually live. I went back for the second one deliberately, and it was worth it again.

They take the gloves off and work at ground level. They run the tools in front of you rather than describing them, and that is what makes it land: you see what is genuinely possible, and you start asking why you would not build your own tooling. Once you know how to prompt, how to set up an agent’s tools and how to serve MCP, that stops being a fantasy and becomes a plan.

It also let me understand an expensive lesson. At one point I had 71 MCP agents running badly, and in seven days they burned through roughly 18.3 billion tokens — the API equivalent of about $15,000. I did not pay that: it ran on a Claude Max plan, which is the only reason a mistake of that size was survivable. I use it hard, and usually reach the weekly ceiling before the seven days are out, then wait for the window to reset. The mistakes were still mine. What the course gave me was how to regroup: how to channel MCPs, how to structure and disperse agent tooling when you are building more than one product, and why context input will bury you if you run many agents without a real pipeline and control over it. That is a simple thing to say and an expensive thing to learn on your own.

The part I did not expect is that everyone takes something different from the same mentor. It is a spark, not a script. Learning is learning; how you interpret it and turn into a builder is where people diverge — and that divergence is the whole point.

Yes — and sooner than you think you need it. Take it the moment you have an LLM feature in front of users and cannot answer "is it any good?" with a number. It changed how I think about building on AI more than anything else on this page.

The idea it dismantles is that building an AI product is just prompting — that because anyone can write a prompt, anyone can build the product. What replaced that, for me, was a discipline: look at your failures before you choose a metric, set your gold labels deliberately, and judge a judge through error analysis and true-positive and true-negative rates rather than by feel. That is the difference between something that demos well and something you know works.

It goes well past building the application. It taught me to stand up a clear evaluation pipeline and then measure and monitor it in production, and to see how the input data itself moves the results you are grading — retrieval in a RAG system being the case I hit first. That is what finally made the whole landscape legible to me.

The teaching is real throughout. Shreya and Hamel work through real applications rather than toy examples, and the guest speakers are people who have actually built these systems and run real evaluations on them — the founder of LangChain presented to our cohort. A lot of teams shipping AI right now are running on vibes. This is the course that fixes that.

Yes — and if you are starting from the beginning, start here. For the statistics and the machine learning properly, in order, with the maths actually done rather than gestured at.

The curriculum is genuinely extensive — statistics and Python, data science, machine learning, neural networks — with a great deal of material and a great many labs. Your hands get dirty. There is no shortcut through it: you can work the whole thing over two or three years, or level up to whatever depth you actually need. It holds at either.

I came to it after twenty-one years of enterprise work — mainframe development, then SAP analytics, then Salesforce architecture — with the last several years on AI. The foundation I built here is what I am working from today, and everything after it was easier because of it. Since then it has been a matter of finding the right courses and the right people to build on top of it.