hi :) i'm chryseis
most recently i worked on an attempt at a [world model for biology] with Reactome & Google Open Source, and led research on RSI & agentic-systems safety at Founded in 2018, a major AI research institute in China whose researchers have gone on to found Kimi/Moonshot, Z.ai/GLM, etc., a nonprofit AI research lab in Beijing.
i've been broadly interested in the ‘science’1 of ‘learning’2. right now particularly the transition from [learning from observational/static data] to [learning from interventional, closed-loop experimentation & experience]. i recently dabbled in biology as i wondered if it'd be a first experimental frontier. see some past work.
1 ‘science’
i'm tempted to call it interpretability, but i mean science in a broader sense. over the past few years interp has become nearly synonymous with mech interp (like the molecular biology of a trained network?), which is just a branch of interp. i strongly believe the right units of explanation need not be the units of implementation. the science perhaps also needs thermodynamics & developmental biology beyond molecular biology, so i'm always excited to see more macroscopic & dev interp work!
2 ‘learning’
i think of learning as the transformation of information into reusable structures. it's very similar to compression, optimization, and generalization, each of which i've studied/researched on separately, but they're subtly different. maybe the closest is ‘compression’, but it has to be compression relative to a system with finite compute (Solomonoff and MDL in the limit; epiplexity, in Finzi et al.'s sense, when the limit is dropped).
there are two exciting consequences of this framework:
- information is then not a property of data but always of the data–learner pair. learnable structure is created by computation, and data has value relative to what a learner can extract with its compute.
- it explains why the thing that tore up the world was a large language model rather than a large anything-else! language is a combinatorial (discrete, recursively compositional) medium carrying much of humanity's accumulated modeling of the world in the form of compressed, reusable abstractions.
this page is alive. you can meddle!
click around to plant life. drag to make a mess.
pattern source ↗fun fact: (biological) life is Turing-complete