But if you’re an engineer evaluating this stuff, ask: what is this actually doing under the hood?
Nine times out of ten, the answer is simpler than the marketing suggests.
#AIEngineering#LLMs
The AI Infrastructure Buzzword Translation Guide
Most of the AI infrastructure layer is Unix pipes and stored procedures in new packaging, wrapped in YAML, and sold with a landing page.
Are the new abstractions useless? No. MCP solves a real interoperability problem. RAG is genuinely useful. “Cognitive architecture” matters in complex systems.
Tool use - function calling, which is just… calling a function
Cognitive architecture - design decisions about routing and reasoning, often reduced to a config file
MCP - a JSON-RPC wrapper around function calls
Guardrails - input validation
Memory - saving stuff to a file or database between sessions
Multi-agent systems - multiple prompts calling each other
Agentic workflows - pipelines where one LLM call’s output shapes the next
Orchestration layer - if/else statements deciding which LLM call to make
RAG - search before answering
The translations:
Skills - tools with business logic baked in
Recipes - predefined sequences of tool calls with a prompt template
Agents - scripts where the LLM decides the control flow instead of you
It planned the tech stack, wrote 437 lines of code, and spun up a server to test it. On my machine, not a cloud.
Fair warning: you need 32GB of RAM for this. The video explains why.
Watch it Here: https://t.co/w9Hw4v7oxU
#LLMs#ClaudeCode
Made a follow-up to the Ollama video. This one goes further.
Instead of just chatting with a local model, I connected it to Claude Code and turned it into an autonomous coding agent. It reads files, writes code, runs commands, and tests its own work.
All locally, no API key, no subscription.
The video walks through the full setup: pulling a coding model, configuring the 64K context window, and launching Claude Code with one command.
Then I gave it a single prompt and let it build a todo app from scratch.
Nothing fancy, just a straightforward setup guide if you've been curious about local LLMs but haven't tried yet.
Link: https://t.co/c1YZ0unSrZ
#llm#localai#ollama#opensource
Made a short walkthrough on running AI models locally with Ollama.
It covers the practical stuff: hardware requirements, installation, and how to increase and verify the context window so the model doesn't forget what you pasted in.
What is a vector space - really? (Machine Learning)
A useful way to think about it:
- Field: a kitchen with reliable arithmetic
- Vector space: a kitchen where mixing and scaling always make sense
- Vector: a specific recipe
https://t.co/jGgYgI4Q5m
With the introduction of S3 Vector Buckets, you can now store, index, and query embeddings directly inside S3 — enabling native similarity search without the need for a separate vector database.
Check out the my video here: https://t.co/2bpPjAynDh
#aws#S3#s3vector