Eric J Ma's Website

And just like that, SciPy 2024 is a wrap!

Because I missed last year's conference, attending it this year in person was a real treat. I met old friends, made new connections, and got caught up with the Python scientific computing and data science tooling ecosystem in line with my goals.

Tutorials

This year, I was not chosen as a tutorial presenter. That was all good, as I delivered two talks. (There will be separate posts going out on them.) As such, I decided to attend tutorials and learn what I could.

Tutorial 1: Quarto

The first tutorial I attended was the Quarto tutorial "Unlocking Dynamic Reproducible Documents: A Quarto Tutorial for Scientific Communication", led by Mine Çetinkaya-Rundel. Though I had tried Quarto two years ago, I have not tried it since. As such, I thought it would be prudent to see if I could find a new trick or two from the tutorial, and I was not left disappointed.

Quarto has improved immensely over the past two years! Of the many improvements I saw, the two that stuck with me the most were special syntax used in creating reveal.js presentations: creating tabs within presentations and enabling incremental reveal of elements on slides. Two days before I was slated to speak, I was inspired enough to re-create my talks in Quarto markdown just for funsies.

Additionally, I learned that Quarto comes with the ability to publish pages to Confluence! This is definitely something worth testing when I'm back at work.

Tutorial 2: LLMs

The second tutorial I attended was an LLM tutorial titled "Pretraining and Finetuning LLMs from the Ground Up." In this tutorial, famed machine learner Sebastian Raschka led a tutorial that did a fairly deep dive into the fundamentals of training LLMs: tokenization, model architecture, formatting data for training, and actually training an LLM. Lightning.ai provided a workspace for executing code, but I opted to remotely SSH back into my home GPU server and follow the tutorial from there.

From my anecdotal interactions with others, the response seemed a mixed bag. This tutorial was too basic for those who were well-versed in LLMs. On the other hand, those whose skill level with language models was elementary but with prior machine-learning experience found this tutorial quite informative. This commentary shouldn't detract from the level of preparation that Sebastian put in; I think it's just a reflection of the hype cycle surrounding LLMs at the moment, where it's hard to target a particular skill level given the diverse groups who are interested in LLMs. My skill level was closer to the advanced side, but I haven't yet prepped data for training LLMs (or protein language models for that matter), so seeing that in action was informative.

Tutorial 3: Anywidget

The third tutorial I attended was the Anywidgets tutorial, "Bring your __repr__’s to life with anywidget", led by Anywidgets creators Trevor Manz and Nezar Abdennur. I believe this tutorial was the best taught of the ones I attended! Trevor and Nezar taught it in such a way that it gave us a very gentle on-ramp to Anywidget. It also included much commentary on how to map JavaScript idioms to Python idioms.

This was essentially a 3-notebook tutorial, paced at one notebook per hour (as I perceived it). The first notebook really boiled down the basics of an Anywidget: a Python class that enables a developer to add JavaScript to control the DOM associated with the next Jupyter notebook cell and link it back to Python objects. The second notebook added some complexity by creating widgets that interacted with others.

But the real fun was the third notebook: ipymario. The goal of that notebook was to create an Anywidget that sounded the Mario chime! If you can, picture yourself being in a tutorial room where the Mario chime randomly rings out from laptops across the room. That's how it was like, and it contributed immensely to the fun that we had.

Tutorial 4: Handling Unusually-Sized Data

The final tutorial was "Data of an Unusual Size (2024 edition): A practical guide to analysis and interactive visualization of massive datasets" by Pavithra Eswaramoorthy and Dharhas Pothina. This tutorial covered many concepts that I already knew, but there were nuggets of information I absolutely cherished: how to store data in the cloud. It turns out that Parquet files are an excellent choice for storing large tabular data in the cloud, as the tooling has caught up to allow us to read slices of data from Parquet files natively in the cloud, (a) omitting the need to set up databases for small and medium data, while (b) ensuring rapid read/write access. Tying this point with a recent tweet from Nikita Melkozerov (essentially how to use a $5 VPS + SQLite to rival hefty Postgres deployments), our choice of data storage system can make a lot of difference when accessing large data via the cloud.

Talks

My experience with talks this year was primarily from a speaker's viewpoint rather than an attendee. I'll summarize what I spoke about below, knowing that I'll be sharing a deeper dive into the two talks later. I'll also highlight what I found to be the best part of the talks program — the Lightning Talks.

Talk 1: How to Foster an Open Source Culture within Your Data Science Team

The first talk I delivered was titled "How to Foster an Open Source Culture within Your Data Science Team." To foster that kind of culture, I had a two-fold thesis, one internally-facing and one externally-facing.

Facing internally, we, as data science team leads, can import into our organization ways of working from the open source world, most crucially introducing a do-ocracy and modelling the behaviour of "anyone can make a pull request." A do-ocracy is an organization where the do-ers are the ones who get to influence the direction of the organization by just rolling up their sleeves, organizing people, and making stuff happen. This stands in contrast to a culture of top-down management or worse, a culture of whining and complaining. Adopting a do-ocracy ensures that our team has a culture compatible with the open-source world.

Externally facing, I built upon the distinction between community-driven and company-backed open-source software and dissected the multiple motivations behind companies that release such software. In particular, I focused on companies like those I have worked at -- Moderna, Novartis, and other companies that do not have a commercialization thesis surrounding open source but use it instead as a recruitment and marketing tool. I then dissected the multiple stakeholders we may need to interact with and convince before releasing OSS software, such as colleagues in legal, management, leadership, and technical teams -- and, more importantly, their motivations.

After the talk, discussions with others revealed a consensus that someone in leadership must be willing to sponsor open-source software. A do-er in the organization must step up and create the framework for clearly deciding when to open-source a piece of software. This is no easy work. It takes time to gain the clarity necessary to set company-wide policy, but once in place, it can provide an impactful framework that others can rely on to make best-faith judgment calls on what to do with their work.

Talk 2: LlamaBot: a Pythonic interface to Large Language Models

My second talk was on LlamaBot! This passion-driven pedagogical project originated from my desire to learn how to build LLM applications, and my approach of "build to learn" has helped me forge a clearer mental model of how LLM application components work. This is reflected in the library. This one being the final talk of the final session on the final day of the main conference, I decided to lighten things up a bit.

I showed a demo of the automatic git commit message writer with audience live participation. Things got funnier when I demoed the Poor Man's Paul Ivanov Bad Pun Generator to the audience!

(For those who are unaware, Paul Ivanov is exceedingly well-known in the community for his timely, sharp, witty, and bad puns in response to almost any situation in the conference. For example, when a speaker recovers from slides failing to show up on screen, he'll quip, "I'll let that slide!")

But demos aside, I also took the chance to describe some of the design decisions made in building the library. I shared how I needed to engage in several rebuilds of the library to gain a level of clarity that also allowed me to make better design choices after the fact. I had a lot of audience engagement for this talk in the form of follow-up questions, and the sprints afterward gave me more ideas about a tutorial to propose for next year's conference.

Lightning Talks

Lightning talks are, arguably, the crowd's favourite part of the SciPy conference. Lightning talks are a maximum of 5 minutes long, and speakers cycle through until the hour of lightning talks is up. Speakers signed up at the NumFOCUS booth, with the queue of speakers clearing out each day. There's a designated Pun Table, where a roster of PUN-dits deliver their PUN-ny puns on the mic for the crowd to groan. Most of the talks are light-hearted and fun! Many of them are on tools that people are developing, which is the real value of lightning talks. It ends up being the highest-density source of information on the zeitgeist of the SciPy community.

As an example, I learned that there are three ways to access GPUs on CI/CD:

• GitHub provides it natively (and hence easily),
• Ubicloud provides it cheaply, and
• Quansight provides a limited server freely (!)

What otherwise would have felt like a challenge to set up is much easier than I originally thought! (For example, I no longer need to designate my home GPU tower as a CI runner.)

As another example, through a hilarious demo of a real-time synthesized David Attenborough-style commentary of live video (of the speakers and then of us), I learned of a new vendor, ElevenLabs, capable of hyper-realistic text-to-speech, which was pretty amazing to witness.

Sprints

With the main conference over, it was time for the sprints. The sprints are 2-day events, where code maintainers come together with first-timers and everyone in between to hack on the scientific Python ecosystem software. I led a sprint on LlamaBot, focusing on notebooks that show how to use LlamaBot in situations that are useful to the sprinters.

But what was illuminating for me was how frustrating it could be to get set up in the first place. Because I had no API keys to give out, I had to set up the sprinters with Ollama, which clogged the internet at the hotel where the sprints were hosted! If we had access to the conference center WiFi instead, that would have massively alleviated the problem.

Nonetheless, all was good: by the time we departed for lunch, most of us had gotten Ollama downloaded and set up the conda environment for development. It just took more time than I initially expected. By the end of the day, we had three pull requests merged, all of them focused on documentation improvements, which made it much easier for others to get started with LlamaBot.

As a side bonus, I also had a chance to see a diversity of setups and their restrictions, ranging from Linux machines to locked down and fire-walled Windows PC laptops, as well as folks using VSCode, PyCharm, Jupyter Lab, Jupyter Classic, and raw terminal. It was, most certainly, a plurality of environments that I hadn't considered before.

Day 2 was slower, and I only sprinted in the morning, working with the Prefix devs to set up pixi in my LlamaBot environment. It was supremely convenient to ask them questions in person rather than wait for an asynchronous online communication cycle. By lunchtime, I had pixi-based testing environments up and running, which gave me more confidence in incorporating it into my daily workflow.

Activities

A feature of the SciPy conference is the Activities Committee. This committee suggests possible activities for attendees to join, especially if they are alone at the conference.

The Activities Committee did an amazing job organizing activities for conference participants. On Friday, the last day of the conference, I went to a rock climbing gym with many other attendees, and thanks to the guidance of others like Alex Chabot-Leclerc, I could pick up some of the fundamentals of rock climbing and use them. I found it to be a very mentally stimulating activity! It was like solving a vertical and lateral puzzle simultaneously. As I write, I'm still physically aching in my forearm from climbing, but it felt awesome.

Apart from that, I spent the afternoon of the second day of sprints exploring Tacoma. At noon, we went to an arboretum that was a good 30 minutes of walking away, so I got in many steps (and a lot of hill climbing -- Tacoma is filled with hills). After lunch, I knew I couldn't miss out on the Tacoma Glass Museum since that was what Tacoma is known for. The highlight there was the live glass blowing! That said, sitting there and watching the painstaking process of glass blowing in front of glowing hot furnaces was too much for my Eskimo-conditioned body to bear, so I left early to view the glass art gallery instead.

Food

No conference report can be considered complete without an accounting of local restaurants. Here's what I had.

I have fond memories of going to Chili Thai for meals three times throughout the conference week, each time ordering a different spicy Thai curry. If not for going out with other groups, I would have gone twice more —- it's that good!

Apart from that, I checked in twice at Thekoi for their XO Ramen, twice at La Isabella for their Mexican food, thrice at Bite at the Murano Hotel (two breakfasts and one dinner), and once at the Courtyard Marriott's Bistro. But my favourite remains Chilli Thai!