Decision Models & You May Not Be Stable But Your Algorithms Should Be
Details
REGISTER AT THE LUMA EVENT PAGE!!!
https://luma.com/r8mu0x5q
Join us for two talks and networking!
Title: Building with Decision Models
Abstract: Apps make thousands of small choices: which team gets a support ticket, whether an agent should run a tool call, which lead to follow up first. Ask an LLM and you get a paragraph back, and your code has to dig the answer out of it.
Decision models work differently. You send a situation and a list of options, and you get back a probability for each one in about a tenth of a second. TypeSafe released the first one, Jev, in September. Within four weeks OpenAI, Microsoft, Cloudflare, AWS and Perplexity had shipped their own.
This is a practical introduction for data scientists: what decision models are, how they differ from the LLMs and classifiers you already use, and where they fit. There’s a live race between an LLM and a decision model, and a look at what happens when they play video games, including an eight-model battle royale in my puzzle game Starlift.
You’ll leave knowing when to reach for one, when not to, and how to test one against your own data.
Bio: Banjo Obayomi is passionate about building with AI and has runs Geneartive AI DC to inspire builders. When not building the next big thing, Banjo likes to relax by playing video games especially JRPGs and exploring events happening around him
~~~~~~~~~~~
Title: You May Not Be Stable, But Your Algorithms Should Be
Abstract: Differential privacy was introduced in Dwork et al. 2006 as a means to preserve privacy of individual records within a dataset. Informally, when an analyst uses a differentially private algorithm, the output of that algorithm won't "look too different" whether or not your data was included in the dataset, and so little identifying information can be learned about you from the result of the analysis. It's since been adopted as the gold standard for privacy preserving analysis, but we now know it also guarantees generalization of learned models to unseen data! We'll introduce the formal definition of differential privacy, talk about some simple differentially private mechanisms, and see why algorithmic stability notions like differential privacy give us generalization guarantees
Bio: Jess Sorrell is an Assistant Professor in the Department of Computer Science at Johns Hopkins University. She holds a secondary appointment in the Department of Applied Math and Statistics, and is a member of the Data Science and AI (DSAI) Institute and the Mathematical Institute of Data Science (MINDS). She studies the theory of machine learning and cryptography, and particularly enjoys problems at their intersection
Agenda:
- 5:30 - Doors
- 5:30 - 6:15: Food & Socializing
- 6:30 - 6:45: Introductions
- 6:45 - 7:15: Building with Decision Models
- 7:15 - 7:45: You May Not Be Stable, But Your Algorithms Should Be
- 7:45 - 8:00: Wrap Up
