🪐 Do planets listen to fire Nep-tunes?
We have a faked AI software engineer, the AI Index report and a tool to analyse your git repos! I’ve been on a few adventures since last time, and we talk about deep learning for natural disaster response. To round it off, I even share some career advice I will try to follow myself!
Late to the Party 🎉 is about insights into real-world AI without the hype.
Hello internet,
I had two hectic weeks travelling while struggling with my health a bit. But that won’t stop us from more machine learning!
In this issue, we have a faked AI software engineer, the AI Index report and a tool to analyse your git repos! I’ve been on a few adventures since last time, and we talk about deep learning for natural disaster response. To round it off, I even share some career advice I will try to follow myself!
So, let’s dive right in!
The Latest Fashion
- Turns out Devin, “The AI software engineer to replace you” demo is fake
- Stanford published their AI Index Report for 2024
- Automatically analyse the “truck factor” of your git repositories
Worried these links might be sponsored? Fret no more. They’re all organic, as per my ethics.
My Current Obsession
I was in Glasgow for the Dimension 20 live show, and it was SO fun! I don’t want to spoil it, but almost the entire cast was there and it was incredible seeing them improvise live. Also got to see an Ally Beardsley Nat 20 live, which was the coolest! As far as I know, they are recorded and will be made available on Dropout, so I won’t get into the story, but it was such a fun idea! The energy of the people was electric so many beautiful nerds that were fully immersed in the story. Then, there was a Q&A afterwards, which was just as much fun.
Afterwards, I had a lovely walk back to the city down the canal. And spent some extra days in Glasgow. I have to be honest. I immediately missed living in Scotland, so I’ll definitely have to come back and visit more. Everyone kept saying how Glasgow is the “artsy” place, so I went on a big walking tour looking at almost 20 huge murals. I also created a Pokemon Go route so others can follow this fun free art show.
Then I immediately travelled to Berlin for PyCon Germany and PyData Berlin. It’s my first PyCon DE and first in-person Python event in quite a while. It was SO lovely! Saw some of my favourite people in the community. Learned about a lot of things, like Pixi and Uv, as well as, Dev containers. Unfortunately I missed all the cool tutorials about Pytest and Hypothesis and Metaclasses. (They’re first come, first served in the morning, and if you know me, you know mornings aren’t my forte.)
While I was away, my colleagues at ECMWF wrote a blog post and snuck in our announcement that we’ll be open-sourcing large parts of our codebase over the next year!
Thing I Like
I got a new fidget toy, which is a small rainbow slug, and it’s very, very satisfying.
Hot off the Press
On Socials
Mastodon really enjoyed PlotNeuralNet (coincidentally, I did, too, using it for my PhD). I wrote a small recap of my time at PyCon Germany and PyData Berlin on LinkedIn, which was widely appreciated. I also posted about our announcement at ECMWF that we’ll be open-sourcing systems for weather forecasting using AI.
Python Deadlines
I checked all the “looming deadlines” and it turns out all of them got 2-week extensions.
As for new conferences, I found PyCon Australia announced and CfPs for PyCon Hong Kong, PyCon Malaysia, PyCon Spain, and EuroScipy.
Machine Learning Insights
Last week I asked, What's the potential of deep learning in enhancing real-time response to natural disasters?, and here’s the gist of it:
Spicy question, isn’t it?
On the one hand, we have to ensure that we maintain a “single voice principle”, so basically in a case of a disaster there needs to be a trusted entity that rings the alarm. This entity can of course use deep learning systems to inform that choice to trigger an alarm, but it’s something to keep top of mind.
Some ways to inform these choices are:
- Early Warning and Prediction Systems: Deep learning models can analyze vast amounts of data from satellites, sensors, and historical records to predict natural disasters such as hurricanes or floods more accurately and with greater lead time. This capability enables authorities to issue warnings earlier, potentially saving lives and reducing property damage.
- Damage Assessment: After a disaster, deep learning algorithms can analyze images from satellites, drones, or on-the-ground cameras to quickly assess damage to infrastructure. This rapid assessment helps prioritize emergency responses and speeds up the recovery process. These models have to run “on the edge” to be accessible in affected situations.
- Communication Systems: During natural disasters, communication infrastructure can be severely compromised. Deep learning models can help in the restoration of communication networks by optimizing the deployment of temporary communication networks, ensuring that affected populations and emergency responders can communicate effectively. However, realistically there is a question if we can even run these models in a compromised situation.
Generally, we can try to use machine learning systems to streamline parts of existing systems, where data bandwidth or complexity is increasing or timeliness of systems can be improved. We are not strictly bound to deep learning here, as some of these systems can probably be implemented with fast on-device algorithms like XGBoost or SVMs. But in the end it’s important to use strict validation of the systems and use additional information like explainable AI to ensure trust and transparency, which is often a bit harder with deep learning.
Got this from a friend? Subscribe here!
Question of the Week
- Can you explain the concept of feature drift and how it affects machine learning models?
Post them on Mastodon and Tag me. I'd love to see what you come up with. Then I can include them in the next issue!
Tidbits from the Web
- Something is wrong with my brain laughing at this way too hard.
- Why you need a “WTF notebook” for your career (especially if you like to fix things)
- This one is for the Dropout Gamechanger fans
Jesper Dramsch is the creator of PythonDeadlin.es, ML.recipes, data-science-gui.de and the Latent Space Community.
I laid out my ethics including my stance on sponsorships, in case you're interested!