How to keep up with AI for science

I read a lot about AI for science, because we make a free monthly roundup on it. It covers drug discovery, protein design, materials, labs, research agents, and the policy around them. This is what I've learned about keeping up without letting it take over your week.

The short answer: pick a small number of sources you trust, check them at a regular time, and write down whether a result was tested in the real world. Most of this guide is about why that's harder in AI for science than it sounds, and the ways people usually go about it.

Why is AI for science hard to keep up with?

Mostly because it isn't one field. It includes drug discovery and protein design, materials and chemistry, weather and climate models, autonomous labs, research agents, benchmarks, datasets, and funding. The useful news can come from a research paper, a lab announcement, a company, a government programme, or a report about a new system.

Look at September's roundup. Stanford's virtual biotech analysed 50,000 clinical trials and produced a lung-cancer design that was later built and validated independently. DreamFold reported protein-folding pathways roughly 30,000 times faster than molecular dynamics. AdaptiveFlow searched a 69-billion-compound space and produced experimentally supported inhibitors. Chalmers described a lab that generated and tested yeast experiments. MIT showed a robot assembling and tuning laser experiments.

The results also need careful reading. MolecureAI reported advancement rates above 79% across several computational tests, but its results stayed computational. DreamFold's evidence supports faster simulation, while the report says experimental structural validation is still needed. A strong benchmark score, a useful design, and a validated scientific result are different things.

How do people usually keep up?

Most people use some mix of these:

• Newsletters. Good for a regular overview, because somebody else has selected the important stories. The catch is that their selection may focus on a different part of AI for science from yours.
• Journal alerts and preprint feeds. The closest you get to new research. They're also noisy, and they can make it hard to tell a useful result from an early claim that still needs testing.
• Podcasts. Easy to fit into a commute, and useful for hearing context across biology, materials and labs. Less useful when you need to find the exact method, number or caveat later.
• Conferences and meetings. A good way to hear what people are working on before the full results appear. They happen only occasionally, and presentations can be difficult to compare with published work.
• Colleagues and communities. Often how the most relevant papers and tools reach you. It works until the people you rely on move to another topic or get busy.

None of them 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 given you anything useful in a month, drop it.
• Pick a time. Fifteen minutes on the same day each week works better than catching up whenever you can. A monthly check can be enough for topics that move more slowly.
• Write one line when you save something. Say whether it was computational, experimentally tested, or validated in a more complete setting. That makes a saved paper much easier to use later.
• Go back to the source for anything you'll act on. A summary tells you what something is about. For the exact result, method or limitation, read the original.

That's most of it. It takes discipline, and the range of AI for science makes it easy to spend all your time reading about work that isn't close to yours.

One more option

This is why I built Iris. You describe your work in your own words, something like "AI-guided protein design and autonomous biology labs", and each month it reads across the field and writes you a briefing on what changed for you, with every source linked so you can check anything yourself. It can separate a computational result from one that reached an experiment, which is useful when the headline makes the claim sound bigger than the work supports.

It's one option, and a good routine with a couple of newsletters works well too. If you'd like to see what a briefing looks like first, our free monthly podcast, Keeping up with AI for Science is about five minutes on the month's biggest stories, with the written report and every source underneath. When you'd like one built around your own work, you can set up your own briefing.

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