There's only so much context you can pack in the semantic model. The rest of it needs to come from a clean data model and a well designed AI workflow.

Probably makes a lot of sense for Fivetran to buy an orchestration company next.

AI coding companies are just feeding our dopamine addiction now. Every two weeks something arbitrary like Claude Skills or Cursor Composer is launched and everybody is like "this changes everything".

When you feel overwhelmed by all the hype, thankfully Daria is always there to cool things down a few notches. youtu.be/GhJK46g-G_s?si…

In many such situations of shared living in Europe, we even had to clean after others fearing that if someone saw the mess, they would default assume it was the Indians.

Using text snippets from Raycast/Alfred for mini prompt instructions works better than fighting with Cursor/Claude rules files. You can inject snippets anywhere.

It helps to have an AI detox week once in a while. No vibe coding or AI auto completion. Just raw dogging everything.

1 year late to the party perhaps but LangGraph looks pretty interesting. Nice clean abstractions, decent state management and unlike some other frameworks that confuse the definition of Agents, LangGraph is very clear about it, which I liked.

My tips for using ChatGPT advanced voice for practicing German speaking. Sharing these here for other language learners: 1. Set the context about your level, expectations and speed. Give it a persona (ex. you are Anna, my German teacher... etc). 2. Create a project in chatGPT

Purple is the new black

Just like calling yourself "hard-core" non-vegetarian doesn't make you an animal abuser, calling yourself "pure" vegetarian doesn't make you a casteist. It's a cultural definition of vegetarianism in India where multiple forms of vegetarianism exist.

Not as smart as the people in this thread (I don't know what encoder-decoder or multi attention layer mean). But working in this field for the last few months on the application layer and there are 5 key dimensions that we have learned about this problem: 1. Users need insights,

Software consultants relabeling themselves as Forward Deployed Engineers in 3...2...1...

Here's an example of the type of question people ask in a text-2-insights system. No current AI system in the world can one-shot this with accuracy. And you'd be wrong if you thought this was just a text-to-sql accuracy problem.

Not sure why the algorithm is showing me so many tweets from Indians living in SF who disapprove of Mamdani.

Nice talk but I never thought of FDE as building custom features. That would be the worst way to build anything. For me FDE is about configuring and tailoring the context and learning enough from staying close to customers to let you incrementally add generalizable features.

Staff training for enterprise AI? It's a friggin chatbot.