You can't keep me in here forever.
Jose
About Me
I'm Arun Jose. I'm known as Jozdien in various spaces. Don't ask me what that means, because I don't know either.

I like making things, and I make whatever I think would be good, regardless of domain - I sometimes develop websites and apps, research things that seem worthwhile, write about whatever I want (which is more limiting than you'd imagine), and draw on the rare occasions I remember I do that.

I try to direct my efforts toward helping the most people - or more scope-appropriately, saving the universe (my friends say I never learned to be modest). It's hard to explain what I mean by that here, but for the sake of clarity in that I'm not kidding: AI vastly more generally powerful than humans are probably possible. Aligning them with our exact, messy, contradictory set of human values is difficult, but most likely necessary, and we're far from there.

I spend a lot of time thinking about the way I think about things, which is where I get a lot of my ideas - people tell me this means I'm often annoyingly meta.

I enjoy all genres of fiction when done well, but my favourites are rational fiction and meta/absurd comedy (as well as crack fanfiction, which feels appropriate closed off in brackets). I hope that's enough to tell you that I'm very rarely serious, and that if I ever come across as caustic, it's far more likely that I'm just bad at being funny to your tastes.

I update this website in a very sporadic fashion, so I can't promise this is all up-to-date or still how I'd phrase some of this. Last update: June 2021.

Instagram and Email are my most frequented platforms for messages, so if you have ideas, want to collaborate, or just think I'd find something interesting (I probably will - err on the side of messaging me), just send me a message with the phrase "Xavier Avial".
Work
The Calibration Game
An android app that optimizes credence calibration - the accuracy of belief strengths against real success rates - with a scoring system that incentivizes alignment of "sureness" with success rates.

A better description can be found here.
The Calibration Game
An android app that optimizes credence calibration - the accuracy of belief strengths against real success rates - with a scoring system that incentivizes alignment of "sureness" with success rates.

A better description can be found here.
Aumann's Game
Website implementation of the multi-player rationality calibration game based on Aumann's Agreement Theorem.

More on that here.
Aumann's Game
Website implementation of the multi-player rationality calibration game based on Aumann's Agreement Theorem.

More on that here.
Isolating Updates
Experimental work on neural network interpretability relating to isolating an update signal corresponding to the high-level objective structure in a deep RL network. Picture by DALL-E 2.

On hold while I explore other approaches.
Isolating Updates
Experimental work on neural network interpretability relating to isolating an update signal corresponding to the high-level objective structure in a deep RL network. Picture by DALL-E 2.

On hold while I explore other approaches.
Reward side-channels
Explorative work into empirical analysis of internal reward representation in reinforcement learning agents to better understand the mechanics of deceptive alignment. Picture by DALL-E 2.

Work in progress.
Reward side-channels
Explorative work into empirical analysis of internal reward representation in reinforcement learning agents to better understand the mechanics of deceptive alignment. Picture by DALL-E 2.

Work in progress.
Metise
Reversible colour-density image compression using conditional adversarial networks. Compression is performed along the dimensions of colour density, being a relatively intractable direction outside neural net reconstruction.

Links to the preprint and code below.
Metise
Reversible colour-density image compression using conditional adversarial networks. Compression is performed along the dimensions of colour density, being a relatively intractable direction outside neural net reconstruction.

Links to the preprint and code below.
Phemos
A news algorithm I developed to account for a documented positive feedback relationship between media coverage and electoral polling. The algorithm weighs down overexposure on a log scale.

More on the aforementioned effect here.
Phemos
A news algorithm I developed to account for a documented positive feedback relationship between media coverage and electoral polling. The algorithm weighs down overexposure on a log scale.

More on the aforementioned effect here.
Stuff I Write
Posts
Gaming Incentives
July 2021 · 10 min read
Insufficient Values
May 2021 · 10 min read
Utopic Nightmares
Jan 2021 · 8 min read
The Truth is Often Comfortable
Jan 2021 · 5 min read
The Loyolite
Jan 2020 · 5 min read
Fiction
Art
Think I'd find something interesting?
I probably will, I'm easily amused. If in doubt, send the message anyway, and preface it with "Xavier Avial".