How to keep up with AI safety

I read a lot of AI safety news because we make a free monthly roundup on it. It's a field with research, engineering and policy stories arriving at the same time. I don't work in AI safety myself. This is what I've learned from making the roundup.

The short answer: keep a short list of sources, check it at a regular time, and record what kind of result each story contains. Most of this guide is about why that takes care in AI safety, and the ways people usually go about it.

Why is AI safety hard to keep up with?

AI safety news comes from several places. There are papers and preprints, lab publications, announcements about deployed systems, incident reports and work from government research bodies. They also cover very different questions, from what a reasoning trace tells you to how an agent might be contained.

Look at this month's roundup. Apollo Research proposed persistent outside evaluators for scheming safety cases. A Wayne State team introduced a measure for mismatches between reasoning and final answers. A Toronto and McGill group found circuit metrics could rank one candidate below an equally sized alternative. NVIDIA announced a watchdog designed to sit outside an agent's software environment. OpenAI disclosed six agent-misalignment incidents.

The strength of those stories varies too. Apollo's proposal identifies a problem with current safety cases, but it doesn't show that embedded evaluation works. The NVIDIA announcement is a commercial design rather than an independent test. The OpenAI incidents are useful operational observations, while the report says they don't establish a general tendency in models. That distinction is easy to lose in a fast-moving feed.

How do people usually keep up?

Most people use some mix of these:

• Newsletters. Good for a weekly or monthly selection, with enough context to see why a paper matters. The catch is that somebody else chose the stories, and their priorities may not match your work.
• Journal alerts and preprint feeds. The quickest route to new research on interpretability, evaluations or oversight. They're noisy, and titles rarely tell you whether a result came from a toy model, a controlled experiment or a frontier system.
• Podcasts. Easy to fit into a commute, and useful for getting the shape of a month. They are less useful when you need to inspect a method, number or caveat later.
• Conferences and meetings. Good for hearing about work before it is published and for asking researchers what they think the result means. They happen occasionally, so they can't carry your whole routine.
• Colleagues and communities. Often where the important caveats surface, especially for papers that are being discussed before they settle. The stream depends on who happens to share something and what they have time to read.

None of these covers everything, so most people end up with a patchwork.

What tends to work

• Keep the list short. Three or four sources you trust are better than fifteen you feel guilty about. If one hasn't been useful for a month, drop it.
• Pick a time. Fifteen minutes on the same day each week works better than trying to catch up whenever you have a spare moment.
• Write one line when you save something. Say whether it is a proposal, a theoretical result, a controlled experiment or an operational report, and why it matters to your work.
• Go back to the source for anything you'll act on. A summary can tell you what changed. For the exact setup, result or limitation, read the original paper or announcement.

That's most of it. AI safety is broad enough that a routine can still leave out an important corner, so the aim is to make the gaps visible rather than pretend you've read everything.

One more option

This is why I built Iris. You describe your work in your own words, something like "whether chain-of-thought monitorability results apply to deployed reasoning models", and Iris reads across the field and writes you a briefing on what changed for you, with every source linked so you can check the details yourself.

It's one option, and a good routine with a few research feeds works well too. If you want to see what a briefing looks like first, our free weekly AI safety briefing is the general version, written for everyone in the field. We also make a free monthly podcast, Keeping up with AI Safety, which takes about five minutes and includes the written report and every source. You can also listen on Spotify. When you'd like one built around your own work, you can set up your own briefing.

If you work in AI safety and think I've missed something, I'd like to hear about it: joe@irisbriefings.com.