SMS scnews item created by Catherine Meister at Thu 6 Nov 2025 0908
Type: Seminar
Modified: Thu 6 Nov 2025 0911
Distribution: World
Expiry: 25 Nov 2025
Calendar1: 6 Nov 2025 1000
CalLoc1: SMRI Seminar Room (A12 Room 301)
CalTitle1: Cole Wyeth, Introduction to Solomonoff Induction
Calendar2: 11 Nov 2025 1100
CalTitle2: Cole Wyeth, Introduction to Solomonoff Induction (continued)
Calendar3: 13 Nov 2025 1000
CalTitle3: Vinayak Pathak, TBA
Calendar4: 18 Nov 2025 1100
CalTitle4: Susan Wei, Bayesian predictive inference and transformers
Calendar5: 20 Nov 2025 1000
CalTitle5: Vinayak Pathak lecture series
Calendar6: 25 Nov 2025
CalTitle6: Structural duality
Calendar7: 13 Nov 2025 1500
CalTitle7: Cole Wyeth, Introduction to Solomonoff Induction (continued)
Calendar8: 20 Nov 2025 1500
CalTitle8: Will Troiani, Smooth Solomonoff induction
Auth: cmeister@159-196-153-133.9fc499.syd.nbn.aussiebb.net (cmei0631) in SMS-SAML

AI Safety Focus Period at SMRI: Wyeth, Pathak, Wei, Troiani

AI Safety Focus Period at SMRI



Over the next 6 weeks we will have several talks around the subject of AI Safety at SMRI as part of the focus period:

If you would like to be kept up to date, please email Geordie (g.williamson@sydney.edu.au) to be added to the zulip channel where discussions and outings are being planned.

Alternatively, here is an overview of the talks over the next few weeks:

(All talks are in the SMRI seminar room):

Thursday 6th November, 10am: Cole Wyeth, Introduction to Solomonoff Induction

Tuesday 11th November, 11am: Cole Wyeth, continued.

Thursday 13th November, 10am Vinayak Pathak, TBA.
3pm, Cole Wyeth, continued.

Tuesday 18th November, 11am: Susan Wei, Bayesian predictive inference and transformers

Thursday 20th November, 10am: Vinayak Pathak lecture series.
3pm, Will Troiani, Smooth Solomonoff induction

Tuesday 25th November, 11am: Structural duality

Some titles and abstracts below.

Title: Introduction to Solomonoff Induction
Abstract: Solomonoff induction (SI) has been proposed as the ideal sequence prediction method. Though direct approximation of SI is intractable, it arguably provides a conceptual "limit" for increasingly powerful (and general) sequence predictors such as autoregressive foundation models. In one sentence, SI is a Bayesian mixture of all generative probabilistic programs. I will formalize this definition, then prove some basic (computability) properties and performance guarantees. Time permitting, I will also try to illustrate the conceptual usefulness of SI and its relationship to modern machine learning.

Title: Structural duality
Abstract: We describe a theorem which relates structure in the data (concentration along a submanifold) to structure in the geometry of the loss landscape. The suggestive picture is that patterns in the data cast shadows that are "valleys" in the loss (hat tip to Olah "every valley in the loss landscape is a shadow cast by some circuit" from https://transformer-circuits.pub/2023/interpretability-dreams/index.html)

Actions:
ball Calendar (ICS file) download, for import into your favourite calendar application
ball UNCLUTTER for printing
ball AUTHENTICATE to mark the scnews item as read
School members may try to .