Shorter thoughts from X and LinkedIn, collected in one place. The longer ones become blog posts.
Initially I was primarily learning how to integrate LLMs into software like RAG, AI Agents, Evals, etc. Now realized you can do so much better if you have the knowledge of ML, both theoretical and applied. Starting this course and will share updates here: aiengineeringfromscratch.com
Most AI tutorials end when the demo works. Real AI engineering starts after that. What happens when retrieval fails? How do you evaluate an agent? What breaks at scale? How do you design systems instead of chaining APIs? I've been collecting books that help build the mental models for those questions. Here are 9 worth knowing about. Which book would you add?
Most prompts don't fail because they're bad. They fail because they get buried. Keep what matters close to the attention. Visibility = reliability.
90% of AI coding mistakes = bad instructions. This repo fixes that. Most AI repos try to make models smarter. This one does the opposite. It fixes how they behave. Andrej Karpathy “skills” repo = instruction layer for coding agents: → stop guessing → ask before coding
Initially I was primarily learning how to integrate LLMs into software like RAG, AI Agents, Evals, etc. Now realized you can do so much better if you have the knowledge of ML, both theoretical and applied. Starting this course and will share updates here: aiengineeringfromscratch.com
Most prompts don't fail because they're bad. They fail because they get buried. Keep what matters close to the attention. Visibility = reliability.
Most AI tutorials end when the demo works. Real AI engineering starts after that. What happens when retrieval fails? How do you evaluate an agent? What breaks at scale? How do you design systems instead of chaining APIs? I've been collecting books that help build the mental models for those questions. Here are 9 worth knowing about. Which book would you add?
90% of AI coding mistakes = bad instructions. This repo fixes that. Most AI repos try to make models smarter. This one does the opposite. It fixes how they behave. Andrej Karpathy “skills” repo = instruction layer for coding agents: → stop guessing → ask before coding