Welcome to Mostly Harmless. Here I (@apiad) will post interesting stuff about computer science, AI, math, programming languages, and random ramblings. Follow this channel if these things are of interest to you. I never spam. PM if you feel like sending some love.

Just started watching this video series hosted by Robert Downey Jr. on YouTube.
https://www.youtube.com/watch?v=UwsrzCVZAb8

It's a bit cheesy at times, and far more comercial than technnical. But anway, the production is amazing, and Downey is such an awesome host. I really enjoyed Episode 1, nothing really new there, at least for people in the field, but the presentation is very compelling. It worries me a little that they make AI seem like this Sci-Fi stuff that is kind of awesome and uncomprehensible, and also that it can solve all of our problems, when actually it's super hard to do anything of value and the more we advance the more we are discovering very concerning issues related to safety, privacy, and other complex stuff.

In any case, I think the series (at least Episode 1) is a very good marketing tool, so if you want to get your friends or parents impressed and excited about this new thing everyone is talking about, instead of showing them your dumb cat pictures, send them this link, and then add: "this is what I'm working on... kind of".

Enjoy!!

I've been watching Rob Miles' YouTube channel for a while and I cannot recommend him enough. He talks mostly about AI safety, which sounds kind of soft until you really start diving into it, and then it becomes a fascinating topic.

The basic idea is that AI is a super powerful technology, and as we develop it, we must also develop safety measures, for a number of reasons. Current, low-intelligence AI systems can fail catastrophically for many subtle bugs, including data biases and badly designed reward functions. Future, high-intelligence or super-intelligent AI systems can fail even more catastrophically if we fail to give them human-aligned goals.

Rob Miles' channel deals with these issues highlighting current efforts by giants of the field like Stuart Russel, and the teams at OpenAI, DeepMind, and many other places. The guy is also super fun in its own nerdy way.

Lately I've been working a lot of Automated Machine Learning (AutoML).

AutoML is a growing field that attempts to automate as much as possible the process of training, fine-tuning and deploying machine learning models on different problems. The basic idea is that researchers/practitioners spend a lot of time testing different models, tuning their hyperparameters or applying trying different types of preprocessing, feature engineering, etc. Most of that work is grunt work, it can be automated with some optimization. We are thus trading computer time for human time, which is far more valuable modeling than writing for loops to try different combinations. Plus, a principled optimization algorithm will be far more efficient searching in a large space of hyperparameter configurations than what any human can do.

Me and a few other collaborators have been working on a new AutoML framework in Python. It is still very experimental but it has already helped us create better machine learning models for a bunch of classic problems. I've written a blog post about it:

https://dev.to/apiad/introduction-to-automated-machine-learning-in-python-with-autogoal-45n4

Hey folks!

I'm starting a Discord server especially for newcomers to the ML world, for sharing and asking for feedback.

I'll try to give as much feedback and answer as many questions as possible.

I also hope some of the more experienced folks to join and help around.

In a couple of weeks I'll be participating (virtually) in the Python 🐍 Pizza 🍕 Cuban edition. Here's a (low resolution version) of my talk...

What Academia can learn from Open Source¶
InstantView from Source