New blog post!

Another more technical one today: Enriched Closed Lenses, by julesh

cybercat.institute/2024/04/12/enr…

New blog post!

Modular Error Reporting with Dependent Lenses, by André Videla

cybercat.institute/2024/04/08/mod…

Another new paper!

Rendering String Diagrams Recursively, by Celia Rubio-Madrigal and me

arxiv.org/abs/2404.02679

🧵 follows:

New blog post: Value Chain Integrity, by 🅾️𝕝𝕚𝕧𝕖𝕣 🅱️𝕖𝕚𝕘𝕖 econpatterns

cybercat.institute/2024/04/04/val…

New paper!

Reinforcement Learning in Categorical Cybernetics, by Riu Rodriguez Sakamoto and me

arxiv.org/abs/2404.02688

🧵 follows…

New blog post! It’s a more technical one today: Colimits of Selection Functions, by julesh

cybercat.institute/2024/04/01/col…

Abstract for an upcoming paper, to be released on the arXiv next week: Reinforcement Learning in Categorical Cybernetics

A neat observation by Bartosz Milewski I didn’t think of before: just as an optic is a coupled coparametrised forward pass and parametrised backward pass (a fact that Bruno Gavranović used to great effect) - so a parametrised optic is coupled *bi*parametrised forward and backward passes

New blog post!

On Organization, by 🅾️𝕝𝕚𝕧𝕖𝕣 🅱️𝕖𝕚𝕘𝕖 econpatterns

“The make-or-buy decision boils down to the cost of holding control over the production process vs the cost of losing control over it”

cybercat.institute/2024/03/22/on-…

New blog post on using lenses and profunctors to implement a neural network. bartoszmilewski.com/2024/03/22/neu…

Maybe I didn’t made it clear that **we know how to do this**

Compositional game theory is deployed in real applications. Not yet differential games, but that’s just putting known pieces together, I think 2 person-years at most

You can make this happen! Fund us at CyberCat Institute

New blog post!

Learning with Invariant Preferences, by julesh

A pitch to approach AI safety using ideas from the recent paper “Categorical Deep Learning: An Algebraic Theory of Architectures” arxiv.org/abs/2402.15332 from Symbolica & Google DeepMind

cybercat.institute/2024/03/18/lea…

The second thing! I wrote a blog post that touches on this point (it doesn’t talk explicitly about lenses but they’re implicitly what the entire post is about)
cybercat.institute/2024/02/06/pas…

I started saying “shaped like the cortex” for the multicategory of lenses, as opposed to the monoidal category of lenses

For example, Zanzi Tangle, now at Monoidal Cafe has figured out that React programs are shaped like the cortex… her extremely hype blog post is coming soon at CyberCat Institute

Somehow 🅾️𝕝𝕚𝕧𝕖𝕣 🅱️𝕖𝕚𝕘𝕖 managed to not say the words “reinforcement learning” or “free energy”, but it’s *plainly* a pitch that both of those things should be central to modelling economies and markets. And that’s, like, the most CyberCat Institute thing I’ve heard in my life

New blog post!

The Attention-Seeking Rational Actor - the second post in 🅾️𝕝𝕚𝕧𝕖𝕣 🅱️𝕖𝕚𝕘𝕖’s econpatterns series

“The fundamental economic exchange is surprises for eyeballs”

cybercat.institute/2024/03/15/att…

The most important thing about this thesis is that so much of it isn't specific to deep learning. The Para(Optic) pattern that Bruno first identified appears in almost every branch of machine learning in different guises - a foundation for the cybernetics of the 21st century

In machine learning, Para(Optic) is *the* abstraction. In general software it’s not, but it still crops up quite often as a design pattern. For example, we think the computational model of React looks like this

Another thing we’re looking for funding to develop at CyberCat Institute

RL agents learn in an MDP. But free energy lets you do RL for a merely controlled Markov chain. And that’s a *big* deal, because “where do the rewards come from?” is the hardest problem in RL

I want CyberCat Institute to focus on this, although we’ll be playing catchup with VERSES AI Research