Seminars relevant to Informatics

[Markulf Kohlweiss <markulf.kohlweiss AT ed.ac.uk>]

Dear All,

I am doing a rerun of my AI Governance teach-out at the Moray House School of Education and Sport picket at 10:30, hopefully my last, I will talk about AI Governance from a privacy and democracy perspective:

Date: Tuesday, 20th October

Time: 10:30am

Location: Moray House School of Education and Sport picket

https://maps.app.goo.gl/Tt2As129jm61QTnC6

Online: n.a. the event is low-tech, interactive, and in person

Title: AI governance: How can computer science and cryptography help save humanity?

Abstract: This teach-out was born from my wrestling with the question: What have we done! And what shall we do? An LLM predicts what a person or character fitting the context would write, drawing on patterns it learned from all of humanity's knowledge (or at least that is the aspiration). It generalizes from that data, so situations that resemble its training material are handled best, while post-training further shapes the output toward the persona and behavior its masters wanted.

At least two issues seem critical in this description to me, first the exploitation of all of humanity's knowledge by AI, knowledge that is now being sold back to humanity. Second the control of that process by AI’s masters, in a manner where every query both contributes to the harvested knowledge and is an opportunity to exert influence. The deep information asymmetry and power difference, as well as the indifference of machines in wielding that power, are what I find most disturbing about closed-weight LLMs. Open-weight models currently only partially address these issues. What I believe is fundamentally needed is a different culture of valuing and managing our data. I will make three concrete proposals for technical mechanisms towards this goal: (i) a study of how valuable specific data is for LLM training; (ii) full transparency and end-to-end verifiability of the data-model-output pipeline; and (iii) an open internet for humans, but strict traffic rules for crawlers and agents via proofs of humanity and delegation.

The goal of these measures is not to provide a technocratic solution but to design technical features that empower a more democratic and fair approach to data and AI governance. 

Bio: Markulf Kohlweiss is a cryptographer and privacy researcher at the University of Edinburgh. His background is on privacy-enhancing technology, with a focus on anonymous credentials, zero-knowledge proofs, payment protocols, and other cryptographic protocols. His surname shares the K and SS with the mythological seeress Kassandra. He has no idea what future AI will bring about, but believes that troubeling conclusions can be derived from the fate of the Internet and Internet privacy. And most people he talks and listens to seem to share the concern, e.g. Cory Doctorow, when he talks about the enshittification of both the Internet and AI. The critical question seems to be: how can dissatisfaction in the grassroots result in meaningful governance? He sees this question echoed at a local level in the recent strike at the University of Edinburgh.

PS. Special shoutout to my CSEC and CSPT MSc students: 

https://www.instagram.com/reels/Dd6NKQNsGRz/

https://en.wikipedia.org/wiki/Katherine_Rundell

The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336. Is e buidheann carthannais a th’ ann an Oilthigh Dhùn Èideann, clàraichte an Alba, àireamh clàraidh SC005336.