The extent to which you can just do things now is staggering: I used AI to generate my own RTS, including the possibility to play against an agent by creating a machine readable gaming engine. I am still wrapping my head around how this will change industries.

Or: "Why are Chatbots trained to tell us what we want to hear so reluctant to say 'I don't know'?" It's a mystery. 🤔

Excited to share that our paper on "Approximate Bayesian Computation with Path Signatures" has been recognized with the outstanding paper award at @UncertaintyInAI! Grateful for my amazing collaborators @JoelNMDyer @pw_cannon

I always disliked the term "statistical science" for exactly that reason. Imo it makes much more sense to simply call it "statistics" and consider it part of philosophy.

An interesting thing is that most people who actually can bike also don't know explicitly about counter steering. It's tacit knowledge. youtu.be/9cNmUNHSBac

Btw, GPT 3 get's wrong.

But now? 😅

Not wearing a seatbelt?

Finally have my author copy of the first paper I published in my PhD. Having a physical representation of achievements like a paper publication make it more satisfying and I am a bit sad that mostly this isn't done anymore 😞

3) Because the error model increases the likelihood of out-of-distribution data, we don't see the failure modes of standard models that assume the simulator is well-specified: we obtain more conservative posteriors while simultaneously better identifying the true parameter.

2) by identifying which summary statistics, i.e. which aspect of a simulator is misspecified. Think, a Covid model could be able to accurately reflect deaths but not infections. Our approach can identify such problems and point researchers to shortcomings.

Assume you have a simulator and interpretable summary statistics. We use a spike-and-slab error model with a Bernoulli variable that adds heavy-tailed noise if the observations have low likelihood.