G’day, I’m Tobin
MIT PhD Candidate, 🇦🇺🇺🇸 Fulbrighter
tsouth (at) mit.edu
Security + AI + Society
Tobin is a PhD candidate at MIT on a Fulbright Future Scholarship advised by Sandy Pentland in the Human Dynamics group at the MIT Media Lab and Connection Science in the Institute for Data, Systems, and Society.
He works at the intersection of data, privacy, AI, and society. In particular, he is focused on building privacy, security, verifiability, and regulatory compliance into frontier AI from first principles. His work has spanned across fields, from decentralized private data sharing to verifiable evaluation attestations for large AI models. He has contributed to the recent work on regulation for generative AI and is keenly interested in working on practically deployable solutions to solve privacy and security challenges in a changing world of technology.
About me: I have a fellowship with the E14 VC fund, helping with AI investments, and previously founded two startups in Australia (both of which failed, great learning experience!). I have a non-executive role on the board of the MIT bigdata Living Lab in Adelaide and sit on the Graduate Student Council at MIT. In my undergraduate, I was Valedictorian of the graduating class of Mathematics, Computer Science, and Electrical Engineering.
Before all that, I grew up in the "Peanut capital of Australia", the agricultural town of Kingaroy, Queensland; but I consider Adelaide my real home. Don't hesitate to reach out if you have something interesting to talk about.
Work Highlights
Some highlights of select recent projects, research papers, and talks
Verifiable AI Model Evaluations
Making trustable claims about the performance and bias of AI models. Perfect for high-risk usecases and research reproducibility.
Transparency by Design
for LLMs
Understanding how retrieval-augmented generation (RAG) can provide auditability and updateability for LLMs.
Secure Community Transformers
A secure & privacy-preserving approach to data pooling and knowledge management for auditable information retrieval in LLMs.
Private Analysis
on Personal Data
Using private set intersections to find overlaps in private personal data without sharing it from arbitrary hashes.
zkTax: Programmable
financial attestations
Circom zkSNARK approach to programable cryptographic financial attestations of tax data signed by authorities.
Play with the tool
Plural Management
A management system for open-source teams that naturally distributed management and ownership over time.
You can find many more academic papers on my Google Scholar, coding projects on Github, and professional work on LinkedIn.
Recent Updates
- 2024, Feb — Co-authored grant on using privacy-preserving LLMs across supply chains.
- 2024, Jan — Presented at Davos on privacy, security, and LLMs.
- 2024, Jan — Contributed to the Singapore M3S program engaging with AI policy.
- 2023, Dec — Awards an MIT Generative AI grant to research end-to-end privacy for generative AI.
- 2023, Summer — Interned at Microsoft Research in the Cryptography and Plurality teams.
- 2023 — Part of the new Data Provenance initiative work for transparency in pretraining data.
- 2023 — I started a fellowship with the E14 VC Fund, helping with technical diligence.
Reach out if you want to talk AI, security, verifiability, or policy
If you’re looking for me elsewhere, my handle is usually
@tobinsouth