Carlos Alberto Haro

I was asked in an interview about talks or public presentations I've given. Thought it might be useful to publish them here. Most of them are in Spanish, but here goes anyway: 1. Modern Data Stack @thedatapub An attempt to demistify the word "modern" by going through the cycles technology has had for analytics

After 5 years someone finally convinced me into reopening linkedin. I will mostly post the same as here, at least the political correct stuff, so here it is, let's connect: linkedin.com/in/h1sort/

I'd MCP my grandma if I could Johnny

And it all started with a formatter. Congrats @charliermarsh, thanks for such amazing tools.

Wake up babe, colab MCP just dropped

killer no credit card combo for teaching data now is neon + duckdb + databricks + opencode. > opencode: setup neon mock database + a write transactions .py > uv run transactions .py > opencode: write duckdb etl to databricks (via cli) > uv run etl .py And you just loop.

@sgmlabs Potato, potahto, let’s burn the 1M tokens together brother

@h1sort This is NOT correct. It is super dense in Math, whereas your estimate is based on pages full of english words. with the amount of math it has, 1M is likely right around how much is needed. But then you might need another 1M to actually interpret that meaning

1M context window is 2 copies of this bad boy stacked

@nicoritschel Oh haven’t thought about revenue go brrr Makes sense lol

@h1sort As per my chart, more cumulative context usage = revenue go brr. I imagine long context is more relevant for professional projects. Teams only allows Max 5x seats and then overage is billed at high margin API rack rate. Anyways…

Not unlike compound interest, folks underestimate token usage at long contexts. The labs want long context to work as much as we all do.

The side quest I took trying web dev for a year (and failing) is really paying off now that we have agents

@leonpalafox Yep, but with horrible accuracy according to that plot doc

@h1sort Didn’t Gemini accepted 1M context?

@nicoritschel Autoresearch should win that term war tbh

@h1sort Hold my beer x.com/nicoritschel/s…

Ralph loop is such a stupid name

It seems trivial to say but you really can impact and org if you just look at the data and take action

it feels to me like SKILLS will be split into pure skills and workflows

Sic semper erat, et sic semper erit

the thing you have to understand is that most applications were slop already

@leonpalafox Gemini sí. El instruction following es pésimo y como al 10mo turno se nota una clara diferencia con los demás modelos.

No tengo pruebas, pero tampoco dudas de que el 80% de los usuarios de IA no notarían la diferencia si les cambiamos entre GPT, Claude y Gemini