Blog

Ensuring high floors for performance is arguably more important than slightly raising the performance ceiling (SOTA chasing). The worst case examples are the ones that get shared, so ensure that they aren’t that bad.

Google search made it easier to access relevant and useful information via a clean design and very simple interface.

Building on yesterday’s post, there can be a pull towards digitalization as access to compute, technology, prediction, etc. gets easier and cheaper. But don’t blindly follow that pull at the expense of the physical.

If AI replaces architects, who will the contractors yell at?

I’m in Washington, DC today to present our FAIR-Frame paper at SIGIR.

With prediction getting cheaper and cheaper, there is more RPG-ification of day-to-day life. Decisions being outsourced to AI means that decisions are being outsourced to dice. These models are estimating probabilities and acting upon them. In a way, we are too when we make our own decisions, but it’s good to keep this in mind explicitly.

Algorithmic generated content is noisy.

Does it make sense to route all requests through a single interface? Even just a little bit of front end logic can make life easier, even if on the surface it is a little bit of friction. Would you prefer a little friction for a better result or a seamless path to the wrong answer?

What need is there for irregularity in language if a language model is injesting the text? At some point, do we accomodate the models and adjust grammar to make English easier to learn?

When will I be able to type the following into a chat box and have it just handle it: