5. Latent Space
Models store patterns in compressed abstract representations. That’s why they can remix ideas and generalise surprisingly well.
6. Fine-Tuning vs Prompting
Prompting changes the output for one conversation. Fine-tuning changes the model’s behavior itself.
3. Hallucinations
LLMs are designed to generate believable output, not guaranteed truth. That’s why they can be confidently wrong.
4. RAG (Retrieval-Augmented Generation)
RAG gives models access to external data at runtime instead of forcing everything into the prompt.
14 minutes of content I think is worth your time.
How to develop soft skills and interview well (from an ex recruiter). Presented in a way that won’t send you to sleep. 100% free to watch.
Link below if you’d like to.
Become an AI Engineer by 2027 👾🤖
This role is exploding, and there will likely be more of these than machine learning engineers. This how to become one and build a SaaS
Video on YouTube out this week (more detailed, i.e including LangChain and MCP)
npm install is now a cybersecurity issue.
Vercel = hacked. Axios = hacked.
Hackers rode the trust developers place in OAuth apps, maintainers, and package registries.
If something can be trusted, it can be weaponised.