Science ramps up AI papers and foundation models

The annual Stanford Human Centred AI Index report is out! This massive report is a fantastic "state of the union" type report on where we're at with AI, tracking progress and major events from the past year.

My story in Nature tracks how science papers are increasingly mentioning AI (up 26% from last year), the new foundation models announced for science, and some skepticism about the utility of 'agents' for performing end-to-end science.

Read it here https://www.nature.com/articles/d41586-026-01199-z 

Score: 50/50 for science

A massive 7-year project aiming to examine the repeatability of social science has come to a close. The project, called SCORE, found (as previous studies have found before) that only half of the tested papers could be replicated successfully when new researchers tackled the same question with new data. It's a glass half-empty / half-full situation... the authors say this is just a sign of scientists being human and making concessional honest mistakes, being messy in failing to report exactly what they did and how they did it, and the strange-but-true fact that doing something slightly differently can legitimately yield entirely different results. 

The team also tested whether people or machines could predict if a paper would replicate (by checking anything from the reputation of the authors to the sample size and the statistical power of the finding). People scored 76-78% at their best; computers failed miserably. But newer AI-based systems are doing far better.

Read my news story in Nature https://www.nature.com/articles/d41586-026-00955-5

Along with a Q&A with Brian Nosek, one of the most famous names in replicability studies and an advocate of open science https://www.nature.com/articles/d41586-026-00972-4

A rabi, a duck and a biologist walk into a bar...

When an improv performer and a bunch of biologists decide to kill time at conferences by counting jokes, this is what you get: seriously funny science.

https://www.nature.com/articles/d41586-026-00854-9 

What does "any lawful use" of AI in war mean?

AI is being used everywhere these days -- including by the military in the current attack on Iran, along with ongoing conflict in the Ukraine and Gaza.

The US Department of War is now insisting that any contractual procurement of AI for the military call for "any lawful use" of AI, without constraints. That has kicked up a fuss with their old supplier, Anthropic, and possible new suppliers including OpenAI, xAI and Google. 

The concerns revolve around mass domestic surveillance and the possibility of future fully-autonomous lethal weapons -- such as drones programmed to identify and kill enemy combatants without human intervention. Many say the latter is currently illegal under international law (though that's disputable). 

Plenty of efforts are ongoing to try to come up with international agreement, but it's hard -- not least because it's tricky to define what "fully-autonomous lethal weapons" even are. Experts are meeting this week in Geneva to talk about possible inclusion under the Convention on Certain Conventional Weapons -- the document which, for example, currently bans the use of blinding lasers in war. It will be a long road to any such agreement, but it's good that the United Nations at least has it in its sights.

My story for Nature 

https://www.nature.com/articles/d41586-026-00710-w

 

 

 

The weighty problem of neutrino mass

Neutrinos -- the ghostly particles that flood our Universe -- are supposed to weigh nothing. But physicists know they do, in fact, weigh something. How much? And which one is heaviest, and which lightest? Now they're zeroing in on answers.

See my article in Knowable magazine. 

https://knowablemagazine.org/content/article/physical-world/2026/physicists-make-progress-weighty-problem-neutrino-mass

Fluorescent proteins get a quantum glow up

Fluorescent proteins, taken from jellyfish, are a standard workhorse of biology. They let us light up proteins and check on the conditions inside of cells. Now, quantum physicists have shown that they can harness the quantum properties of electrons in these proteins, turning them into exquisitely sensitive sensors of magnetic fields. This also means we can turn them off and on remotely, making them useful for new imaging and therapeutic techniques. Read all about it here!

https://www.nature.com/articles/d41586-026-00662-1 

Don't you know that you're toxic?

Some researchers complain that the peer review comments on their work are vague, with broad, simple comments such as “not novel”, or even unprofessional, with comments such as “these authors don’t know what they’re talking about”. A new study checks in on whether an AI coach can help, suggesting ways to make an already-written peer review more constructive -- and polite. 

https://www.nature.com/articles/d41586-026-00536-6

Do AI translators work well enough for scientific papers?

As of 11 Feb, the popular arXiv preprint server asks that all papers be in English, or be accompanied by a full English translation. They say automated translations are fine, so long as they are faithful to the work.

So: are AI translators up for this task?

Find out in my story in Nature: https://www.nature.com/articles/d41586-026-00229-0 

arXiv cuts down on AI slop

The popular arXiv preprint server, which holds nearly 3 million manuscripts, mainly in computing science, physics, and math, has put up a new hurdle for first-time submitters, in an attempt to cut down on 'AI slop'.

Until now, someone wanting to submit to arXiv for the first time only needed an email address affiliated with a reputable academic or research institution, such as a university. But a rule instituted on 21 January now requires first-time posters to be endorsed by an established arXiv author in their own field. People who have previously posted in the same disciplinary section of arXiv do not need an endorsement.

The move is an attempt to clamp down on a rising tide of fraudulent submissions, says University of Amsterdam astronomer Ralph Wijers, chair of the arXiv editorial council. A large fraction, he says, are generated with artificial intelligence (AI). The new rule is “mostly to try and discourage very junior, unskilled people from trying to get something started by sending some rubbish to arXiv,” he says. 

Read my story, in Science  

https://www.science.org/content/article/arxiv-preprint-server-clamps-down-ai-slop  

Nature Briefing: AI and Robotics

I have just started as editor of Nature Briefing: AI & Robotics, taking over with thanks from Josh Axelrod.

Want short, fun bits of AI news in your inbox every other week? Sign up!  (You may need a Nature account)

Donald Trump in Davos

Trump's speech is much longer than Carney's. He mentions his "enemies", does a lot of bragging about the US economy, tells some lies about green energy (and does some literal tilting at windmills, hilariously), says the usual BS about how the US "gave Greenland back" and now Denmark is soooo ungrateful, and some other oddities. But here's my favourite excerpt. Remember... he is in Davos. Which is in Switzerland. And Switzerland's president, by the way, is Guy Bernard Parmelin.

Here we go... Brace yourself.

Trump: "we put a 30 per cent tariff on Switzerland and all hell broke loose. They were calling, I mean like you wouldn’t believe. And I know so many people from Switzerland. Incredible place, incredible brilliant place. But I then realized that they’re only good because of us.

... And the I guess prime minister — I don’t think president — I think prime minister called. A woman. And she was very repetitive.

She said, ‘no, no, no, you cannot do that, 30 per cent. You cannot do that. We are a small, small country.’ I said, ‘yeah, but you have a big, big deficit. You may be small, but you have a big deficit than big countries.’

She said, ‘no, no, no, please, you cannot do it.’ Kept saying the same thing over and over. ‘We are a small country,’ she said, but you’re a big country in terms of… and she just rubbed me the wrong way, I’ll be honest with you.

And I said, ‘all right, thank you, ma’am. Appreciate it.’

‘Do not do this.’

‘Thank you very much.’ And I made it 39 per cent.

And then all hell really broke out. And I was paid visits by everybody. Rolex came to see me.  They all came to see me. But I realized and I reduced it because I don’t want to hurt people, I don’t want to hurt them.

... And as she said, "it’s a small place" and I realized with that, I don’t know, I was so, because she was so aggressive. And I realized in that conversation that the United States is keeping the whole world afloat."

 

!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 

 

Treatments for ADHD: beyond stimulants

If your child has ADHD or you're curious about its treatments, check out my piece for Nature.

https://www.nature.com/articles/d41586-026-00095-w

All brains are different. ADHD isn't a "problem", it's a variant in brain wiring. We can harness its advantages, or alter our environment to reduce the disadvantages. Along the way, we can also treat the symptoms that are making life hard. This feature is about that. 

Main messages:

- Stimulants like Ritilin and Adderall work really, really well: they are some of the most efficacious drugs on the market for ANY condition.

- These drugs improve all kinds of measures, not just the core symptoms of inattention, hyperactivity and impulsivity, but also self esteem, reduced car crashes, less criminal behaviour, etc. But they don't magically make your kid smarter (they don't necessarily do better on tests) and they might wear off after a few years of use.

- stimulants come with issues for some people (addictive qualities, misuse, anxiety) 

- There are other drugs available, notably alpha agonists, which a lot of people haven't heard of but can work just as well or better for some people.

- other innovative drugs are in the pipeline but none look particularly promising. Again, stimulants work REALLY well so they're hard to beat.

- what about non-drug options? You can stimulate or exercise the brain to build up the bits that need work in ADHD patients.

- exercise, meditation, talking therapy - none have been shown to work very well (but they probably help with other issues, like self esteem etc!)

- neural feedback (where you get a visual representation of what's going on in your brain, to help you direct blood flow etc to certain bits) also doesn't seem to work well, even though it sounds like it should work.

- neural stimulation (where you literally give tiny electric shocks to part of your brain from a headband) might work! It's early days so hard to say, and there's different types, but some of it might work.

One of the researchers I spoke to has ADHD himself and has never chosen medication... instead he drinks a lot of coffee, and lets his interest in his work, and his hyperfixation, help him to achieve great things. That story didn't make it into the piece. But it just goes to show... ADHD brains for the win.