Last month, I attended the SciPy Conference in Minneapolis, MN at the University of Minnesota McNamara Alumni Center. The annual conference celebrates open source scientific computing in Python. This is an especially insightful conference for Humane Intelligence to attend, as we prepare for the release of our new Python-based AI red teaming application under an open source software license.
The week started with tutorials, where community members teach each other how to use open source scientific computing software or about new methodologies. The main conference is where developers showcase what they’ve been building and is a chance to connect with their users. SciPy closed with developer sprints, where the community gets together and collaboratively builds open source scientific software in Python.
The main conference opened with Thomas Caswell’s talk, Stories in Code, in which he covered the different eras of the SciPy community including what today’s community is like with the prevalence of generative AI. On the next day, we heard from Amber Case, speaking to Calm Technology and the History of AI where she spoke about design that is meant to give important information to the user without being obtrusive. In the final keynote of the conference, Snakes in the Microwaves: How Python is Powering the Golden Age of SAR, Joseph Kennedy demonstrated how the SciPy stack is being used in synthetic aperture radar (SAR), which is a cutting edge technique to observe the Earth’s surface from space. This talk captured the spirit of the SciPy Conference, showcasing scientific discoveries as enabled by the tools that were built by this community.
There were many incredible talks outside of the keynotes, and we highlight 3 below:*
Automated Data Enrichment for Police Accountability: Where Agentic Judgment Earns Its Place – Hongsup Shin
The speaker built a LangGraph pipeline to create better datasets using the Texas Justice Initiative’s police shooting datasets and analyzed where agentic judgment can be used vs where determinism is required. This is a great example of using AI for social good and making choices in agentic systems to mitigate the risk of bad output in a high stakes field.
Just throw it away? Class imbalance lessons from molecular machine learning to meatballs – Jackie Valeri
The speaker compared downsampling strategies for imbalanced datasets across antibacterial molecule screening, sandwich taste ratings, and baseball at-bats. She demonstrated the efficacy of statistical techniques for the case where it is difficult or impossible to collect large datasets, something that is common across fields, and often occurs in the social sciences. Further, the methods that were used were not generative AI or LLM based, but rather more traditional machine learning techniques.
Derivations, Not Just Simulations: Teaching Applied Mathematics with Scientific Python – Michael Zargham
The speaker, one of Humane Intelligence’s advisors, presented a six-stage teaching pattern built on SymPy in marimo notebooks making advanced mathematics much more accessible to students as demonstrated through the classic Homicidal Chauffeur pursuit-evasion game. This talk was in the education track and was a great demonstration of creating an interface that allowed people to interact with substantive content in a straightforward way. A page with talk slides and an interactive notebook is available here.
There were so many other interesting talks and tutorials at the SciPy Conference, and here are some major takeaways from the week in Minneapolis:
Generative AI was a prevalent topic throughout the conference, both in terms of its use in scientific computing as well as its impact on open source software development. Many talks discussed leveraging LLMs or creating agentic AI systems in order to get to scientific results with use cases ranging from OCR to police accountability to weather forecasting. However, even though we did see a lot of uses for these types of systems, given that SciPy is a community dedicated to open source software for scientific computing, there are many factors to consider before rushing to a generative AI-driven solution.
Generative AI is changing the way that people contribute to open source software. AI-assisted coding has enabled contributors to produce code at rates never seen before. Even the communication around code contribution has been impacted, with AI-backed entities submitting code to projects. As a result, OSS projects are actively developing new policies to handle contributions, and there is no one-size-fits-all approach.
Open source scientific computing libraries, especially those featured at the SciPy Conference, are being used as part of generative AI workflows. As such, there is a push for developers to maintain or create compatibility of these libraries with these systems. In some ways this shows the value of this community, but also adds strain to a group that has already put in a lot of volunteer work to make these libraries sustainable. Thus, something needs to shift in the funding and economics of generative AI.
Documentation and tutorials in this new era of software development will likely need to evolve because people are using the libraries differently, often through the lens of LLMs. Part of the issue for the SciPy community is that one of the stories that the software tells is of scientific discovery. Agentic systems or LLMs that use scientific computing software without discernment risk being decontextualized from the scientific aspect of what these tools were built for, and can lead to misuse, invalid results, or more. One thing that remains clear is that there is no one “correct” answer to address these issues.
One of the most engaging aspects of the SciPy Conference is the birds of a feather (BOF) sessions. These take the form of panel discussions or community conversations. Humane Intelligence and our close collaborator, Dynamical Systems Group, led a guided BoF community discussion around building generative AI systems the scientific way, which we’ll discuss in an upcoming blog post. Keep an eye on our “Insights” section for more related content.
*This year, I was honored to be one of the co-chairs of the Data Driven Discovery, Machine Learning, and AI track where I got to review and watch many of the submissions.