Blog

Typically, an AI system is learned offline. You collect your data, then train and evaluate your model. All offline. You then deploy and cross your fingers that the live data is similar to the offline data and your model is able to be useful.

August is a summer month. I will not be convinced otherwise.

Pre-LLMs, huge amounts of data were generated and left unread/unheard/unseen. Or only seen by a small handful of individuals.

Now, web content ingested by LLMs is “seen” by these models, and these models are used by huge numbers of people. So as a content creator, can you claim that you have that reach, if ChatGPT would not be able to garner that reach without your content? It’s an inverse tragedy of the commons.

Today I learned (TIL) a new thing!

What will be the biggest AI moat?

In. Execute. Out.

What will happen to language if you spend more time talking to an AI than you do other people?

Curriculum learning is an interesting area of study within machine learning. However, it is often applied in a somewhat hand-wavy fashion. What does it mean for a particular training example to be harder than another one?

Even something as seemingly definitive as time can wiggle a little from different perspectives. If you set yourself a goal of writing every day, do you have until midnight in your time zone to meet the goal for the day, or until the day is done across the world? Perhaps more importantly, does it matter?

Define helpful. In a way that can be replicated.