What I'm reading and watching to keep up with AI

Keeping up with AI is mostly a filtering problem. More gets released in a week than anyone can read, and most of what reaches you through social feeds is a summary of a summary. Here's what I do about it.

One channel I actually watch

The AI Search YouTube channel is the one I keep coming back to. Its roundups start with the biggest release of the week and keep moving, with no long warm-up. That format shaped how I built my own daily radar. I'll watch a video at the end of a day, and when something looks significant, I stop and go find the source.

I mention the channel because I'd rather credit where an idea came from than pretend I invented the format.

Primary sources over tweets

The rule is simple. If I'm going to repeat a claim, I open the thing itself: the model card, the repository, the lab's own post, the paper page. A tweet can tell me where to look. It can't tell me what's true.

This has saved me from a lot of noise. A "breakthrough" is often a benchmark on a narrower task than the headline suggests. A new model can be gated behind a waitlist. A license can quietly rule out the use I had in mind. Each of those takes two minutes to find in the source and a lot longer to unlearn after you've repeated it.

What's on my radar

A script pulls lab blog feeds, trending models and daily papers from Hugging Face, GitHub releases, a few subreddits, Hacker News items above 80 points and some YouTube channels. It drops what I've seen. Claude drafts a post only when at least three items clear a real-substance bar, and every link has to load. I read the result and decide what to keep.

What I'm looking for varies by the day, but there are a few standing interests:

Open-weight models that run on one machine. If a model I can host myself reads Arabic documents well, I can offer a client something that never leaves their network. That changes what I'd build.

Tooling for agents. How roles, memory and tool access are handled in the products I use daily, since that's the layer my companies sit on.

Gulf AI. Compute, local content rules and talent programs, because I work in that market. I follow named sources and dated announcements.

What I read on paper and screen

Lab blog posts and technical reports, mostly in the evening. A few newsletters from people who read the papers themselves. Company filings and regional business press when I need to understand a market, not a model.

I try not to read every reaction. After a launch, the first day is full of confident takes from people who haven't used the product. I wait a few days, then read the ones that include a test.

What I skip

Anything that lists "ten tools that will change everything." Anything with no link to the thing it describes. Predictions with a date far enough away that nobody will check.

Why it matters for the work

I make build-or-buy calls for companies that bid on real contracts. A wrong guess about what a model can do costs money. Reading the source costs two minutes.

If you have a source I should be reading, send it along. I'll check it, and I'll tell you if it changes my view.