Santiago Valdarrama

Scrum is a cancer.

I've been writing software for 25 years, and nothing renders a software team useless like Scrum does.

Some anecdotes:

1. They tried to convince me that Poker is a planning tool, not a game.

2. If you want to be more efficient, you must add process, not remove it. They had us attending the “ceremonies,“ a fancy name for a buttload of meetings: stand-ups, groomings, planning, retrospectives, and Scrum of Scrums. We spent more time talking than doing.

3. We prohibited laptops in meetings. We had to stand. We passed a ball around to keep everyone paying attention.

4. We spent more time estimating story points than writing software. Story points measure complexity, not time, but we had to decide how many story points fit in a sprint.

5. I had to use t-shirt sizes to estimate software.

6. We measured how much it cost to deliver one story point and then wrote contracts where clients paid for a package of “500 story points.“

7. Management lost it when they found that 500 story points in one project weren't the same as 500 story points on another project. We had many meetings to fix this.

8. Imagine having a manager, a scrum master, a product owner, and a tech lead. You had to answer to all of them and none simultaneously.

9. We paid people who told us whether we were “burning down points“ fast enough. Weren't story points about complexity instead of time? Never mind.

I believe in Agile, but this ain't agile.

We brought professional Scrum trainers. We paid people from our team to get certified. We tried Scrum this way and that other way. We spent years doing it.

The result was always the same: It didn't work.

Scrum is a cancer that will eat your development team. Scrum is not for developers; it's another tool for managers to feel they are in control.

But the best about Scrum are those who look you in the eye and tell you: “If it doesn't work for you, you are doing it wrong. Scrum is anything that works for your team.“

Sure it is.

95% of Machine Learning solutions in the real world are for tabular data.

Not LLMs, not transformers, not agents, not fancy stuff.

Learning to do feature engineering and build tree-based models will open a ton of opportunities.

This was a shocking book.

I just finished it and wasn't expecting what I learned.

Every Machine Learning and Data Science practitioner should learn about causal inference.

It's a different way of thinking. It makes me look at the world with different eyes.

Many people, including me, always say we should let the data speak.

This book taught me an important lesson: Data is usually not enough; we need to experiment.

The cherry on top is the Python chapters.

Don't go another day without learning about this: https://packt.link/LDlUu.

Knowledge graphs are a game changer for AI Agents!

A few ridiculous and eye-opening benchmarks comparing an AI Agent using knowledge graphs with state-of-the-art methods:

• 94.8% accuracy versus 93.4% in the Deep Memory Retrieval (DMR) benchmark.

• 71.2% accuracy versus 60.2% on conversations simulating real-world enterprise use cases.

• 2.58s of latency versus 28.9s.

• 38.4% improvement in temporal reasoning.

(You'll find the above benchmarks in the paper I'm linking below.)

A knowledge graph is a network of connected points, each representing a piece of information. It's a very efficient structure for capturing complex relationships between data.

AI Agents need memory and must know how to keep it updated over time (This is difficult, and it's the main reason most agents get dumber overnight!)

This is where a knowledge graph helps:

1. They make it easier for the agent to extract facts from memory
2. They make it easier for the agent to update facts as they change

This paper introduces Zep AI (YC W24) a memory layer service that uses Graphiti, a knowledge graph engine.

Graphiti is open-source. You can use it with any agent framework, model, or platform.

Languages: Python, Go, and TypeScript.

I got my Master's from Georgia Tech with a specialization in machine learning.

Here are the classes I took and the money I paid:

1. Machine Learning
2. Computer Vision
3. Reinforcement Learning
4. Intro to Graduate Algorithms
5. Machine Learning for Trading
6. Database Systems Concepts and Design
7. Software Development Process
8. Software Architecture and Design
9. Human-Computer Interaction
10. Advanced Operating Systems
11. Software Analysis and Testing

You only need 30 credits to graduate. I completed 33.

It took me 4 years to go through all the classes (2015-2019).
I paid $510 for every 3 credits, and I completed 33 of them, so I ended up paying $5,610 for the classes.

I also paid $310 in term fees. It took me 11 terms to finish, so I paid $3,410.

$5,610 + $3,410 = $9,020 was my total cost.

If you take more than one class per semester, you will pay less money (it can cost as low as ~$7k.)

I was working full-time the whole time, and I didn't want to rush it.
I loved the program. It taught me a lot. It was the reason I started focusing on machine learning professionally.

#machinelearning #gatech