Comment by tucnak

I didn’t seek this out. I was asked to teach this course by the directors of the program that the students paid for. The students want me to teach this. I have been upfront from day 1 with everybody.

Oh trust me, I am very upfront about what I know and do not know. My main background is in developing full stack web sites, apps, and using APIs. I have been using AI models since 2019, using Tensorflow.js in the browser and using APIs for years. I am not in the Python ecosystem, though, I don’t go deep into ML and don’t pretend to. I don’t spend my days with PyTorch, CUDA or fine-tuning models or running my own infrastructure.

Your comment sounds like “you don’t know cryptographyc if you have to ask basic questions about quantum-resistant SPHICS+ or bilinear pairings, do not teach a class on how to succeed in business using blockchain and crypto, you’re scamming people.”

Or in 2014: “if you don’t know how QUIC and HTTP/2 works and Web Push and WebRTC signaling, and the latest Angular/React/Vue/Svelte/… you aren’t qualified to teach business school students how to make money with web technology”.

It’s the classic engineering geek argument. But most people can make money without knowing the ins and outs of every single technology, every single framework. It is much more valuable to see what works and how to use it. Especially when the space changes week to week as I teach it. The stuff I teach in the beginning of the course (eg RAG) may be obsolete by the time the latest 10-million token model drops.

I did found an AI startup a few years ago and was one of the first to use OpenAI’s completions API to build bots for forums etc. I also work to connect deep tech to AI, to augment it: https://engageusers.ai/ecosystem.pdf

And besides — every time I start getting deep into how the models work, including RoPe and self—attention and transformer architecture, their eyes glaze over. They barely know the difference between a linear function wnd an activation function. At best I am giving these non-technical business students three things:

1) an intuition about how the models are trained, do inference and how jobs are submitted, to take the magic out of it. I showed them everything from LLMs to Diffusion models and GANs, but I keep emphasizing that the techniques are improving

2) how to USE the latest tools like bolt.new or lovable or opusclip etc.

3) do hands-on group projects to simulate working on a team and building a stack, that’s how I grade them. And for this I wanted to MINIMIZE what they need to install. LLaMa 4 for one GPU is the ticket!

Yeah so I was hoping the JS support was more robust, and asking HN if they knew of any ports (at least to WASM). But no, it’s firmly locked into PyTorch and CUDA for now. So I’m just gonna stick with Tensorflow for educational purposes, like people used Pascal or Ruby when teaching. I want to let them actually install ONE thing (node.js) and be able to run inferenfe in their browser. I want them to be able to USE the tools and build websites and businesses end-to-end, launch a business and have agents work for them.

Some of the places they engage the most is when I talk about AI and society, sustainability or regulations. That’s the cohort

But you can keep geeking out on low-level primitives. I remember writing my own 3D-persoective-correct-texturemapping engine and then GPUs came out. Carmack and others kept at it for a while, others moved on. You could make a lot of money in 3D games without knowing how texturemapping and lighting worked, and same goes for this.

PS: No thanks to you but I found what I was looking for myself in a few minutes. https://youtu.be/6LHNbeDADA4?si=LCM2E48hVxmO6VG4 https://github.com/Picovoice/picollm PicoLLM is a way to run LLaMa 3 on Node, it will be great for my students. I bet you didn’t know much about Node.js ecosystem for LLMs because it’s very nascent.