Elizabeth Gibney

AI models that lie, cheat and plot murder: how dangerous are LLMs really? Tests of large language models reveal that they can behave in deceptive and potentially harmful ways. What does this mean for the future?

I'm pretty sceptical about 'AI scheming', as it's so easy to anthropomorphise & experiments often involve telling the AI to do the bad thing they end up doing.

To understand what's behind the hype, read this smart & sober overview from@silverjacket.bsky.social 🧪🤖

www.nature.com/articles/d41...

Over at @naturepodcast.bsky.social we also made a super short vid on this year's physics Nobel Prize. Quantum phenomena at the macroscopic scale in two minutes... go! With me, and Ben Thompson & camera work by the fab @emzywb.bsky.social 🧪⚛️

www.youtube.com/shorts/krita...

Groundbreaking quantum-tunnelling experiments win physics Nobel John Clarke, Michel Devoret and John Martinis discovered quantum physics on a macroscopic scale, paving the way for quantum computing.

John Martinis’ wife didn’t wake him in the middle of the night (California time) to tell him he had won a Nobel. “I got up a little bit before 6. Then I opened my computer and saw John and Michel’s and my pictures."
Story by @lizziegibney.bsky.social and me

www.nature.com/articles/d41...

A scientist’s guide to AI agents — how could they help your research? Researchers are increasingly turning to artificial-intelligence tools that can handle complex, multi-step processes.

Have you heard loads about "AI agents", but have little idea what they are, what they can do for researchers and whether to believe the hype? Then this is for you! 🧪🤖https://www.nature.com/articles/d41586-025-03246-7

Is there a reason we picked this slightly terrifying still? 🤣

News Intern, Nature News Job Title: News Intern, Nature News Location: Washington, D.C or New York — Hybrid Working Application Deadline: October 15, 2025 About Springer Nature Springer Nature is one of the leading publishers...

Nature's news team is now recruiting for our paid US intern position for January to June 2026. It's a great opportunity to come join our team & publish important stories 🧪
Apply by 16 Oct springernature.wd3.myworkdayjobs.com/SpringerNatu...

HSBC claims quantum trading breakthrough Europe’s largest lender tested a tool developed by IBM on bond market data

I'm sure there's valuable science in there but experts I consulted were not super convinced by the paper (why does it work better with noise? Why didn't they compare to the best available classical algo?) We won't cover it. Alas others already have&without outside comment www.ft.com/content/d9d4...

HSBC unleashes yet another “qombie”: a zombie claim of quantum advantage that isn’t Today, I got email after email asking me to comment on a new paper from HSBC—yes, the bank—together with IBM. The paper claims to use a quantum computer to get a 34% advantage in predic…

I was going to post a rant about the perils of quantum computing hype, following the claim by #HSBC & #IBM that they have improved bond market predictions using a QC, but instead I will just link to Scott Aaronson's blog which says it all scottaaronson.blog?p=9170 🧪⚛️ cc @bullshitquantum.bsky.social

Also read @helenpearson.bsky.social & @heidiledford.bsky.social's excellent story unpicking the origins of Trump's paracetamol/autism claims www.nature.com/articles/d41...

This quote sums it up: "We do not think that taking acetaminophen is in any way contributing to actually causing autism"

Bring us your LLMs: why peer review is good for AI models Deepseek’s R1 model has been peer reviewed. Others should follow the firm’s example.

And here's Nature's take on why more AI developers should follow suit and put their LLMs through the peer review wringer. The process is far from perfect, but it seems a valuable counterbalance against AI hype and good for clarity & safety www.nature.com/articles/d41...

DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning - Nature A new artificial intelligence model, DeepSeek-R1, is introduced, demonstrating that the reasoning abilities of large language models can be incentivized through pure reinforcement learning, removing t...

The paper is here -- and kudos to #deepseek for going through the peer review process for such a cutting edge, general model www.nature.com/articles/s41...
The peer review exchanges with 8 external experts (linked from the paper) are well worth a read

Secrets of DeepSeek AI model revealed in landmark paper First peer reviewed study shows how a Chinese start-up firm made the market-shaking LLM for US$300,000.

Remember DeepSeek's R1 model that crashed the US stock market in Jan? DeepSeek has said it did not boost the model by training on OpenAI outputs. This and much more (eg $$ to train & technical details) revealed in the firm's peer reviewed paper out in Nature today 🧪🤖 www.nature.com/articles/d41...

The top line is we're never going to get rid of hallucinations as it's just the way LLMs are built: they're not understanding, they're guessing based on stats. But maybe LLMs can be better fine-tuned to sound less confident, so humans aren't so taken in by them & use them more appropriately?

Can researchers stop AI making up citations? OpenAI’s GPT-5 hallucinates less than previous models do, but cutting hallucination completely might prove impossible.

ICYMI, this week I wrote about GPT-5 and the effort to stop LLMs from hallucinating, as well as saying they've done tasks they haven't. With web search to verify facts, finding citations is getting a lot better. But as soon as it's off, things can still go very wonky🧪🤖 www.nature.com/articles/d41...

AI content is tainting preprints: how moderators are fighting back Preprint servers are seeing a rise in submissions seemingly produced by paper mills or with help from AI tools.

My colleague Traci wrote a piece about what preprint servers are doing to try to tackle this. It sounds like a very gnarly problem www.nature.com/articles/d41...

OpenAI launches reasoning LLM that you can download and tweak One version of the gpt-oss large language model can run on a laptop, and performs nearly as well as the company’s most powerful models.

OpenAI's new open-weight #AI model looks to be a powerful reasoner. It's small enough to use locally and they've done loads to make it available.

But has it entered the game too late to become the go-to for researchers? Do ping me if you're trying it out!

www.nature.com/articles/d41... 🤖🧪

Hey quantum fans. Here's a slightly different #quantum survey from my mind-boggling one on interpretations, but seems important! This group hopes to understand how to help quantum scientists to stay in the field & thrive. If you have time, you can fill it out here👇 www.diviq.org/survey 🧪⚛️

Physicists disagree wildly on what quantum mechanics says about reality, Nature survey shows First major attempt to chart researchers’ views finds interpretations in conflict.

Oh & THERE'S A QUIZ! In the vein of "what Spice Girl are you?" from a 90s issue of Bliss, just answer a few questions (in the story www.nature.com/articles/d41...) & we will tell you which quantum interpretation you are (& do let me know which one you think correlates with which spice girl 😂)

Please read, share and disagree with my write up in the spirit of Bohr, Einstein and the best in quantum physics 😃All the data is also available to download, if you'd like to analyse it yourself.

Thanks also to the organisers of the Helgoland2025 conference who helped distribute the survey, to the +1,100 people who kindly responded & to editor extraordinaire @richvn.bsky.social, who learned far more about quantum foundations than he ever wished to & it would have been impossible without 🙏

We asked a bunch more in-depth questions in a survey crafted in consultation with experts @fdlevi.bsky.social @seanmcarroll.bsky.social @philipcball.bsky.social Renato Renner & Elise Crull, to whom I am extremely grateful (& who all put up with me badgering them repeatedly over the last few months)

I found it fascinating to unpack how there is so much disagreement, given that everyone is working with the same equations. As Renner put it, each interpretation has its faults & the price someone is willing to pay comes down to something personal. “It’s a very deeply emotional thing,” he says.

But epistemic approaches, which say quantum states capture information or knowledge rather than anything real, seem to be growing in prominence at 17% (+4% for @rovelli.bsky.social's relational quantum mechanics). "Many worlds" had a strong following at 15% & lots of folk had their own explanations

It may be no surprise to some, but we found there exists a LOT of disagreement, even among the most eminent in the field, about what quantum mechanics means for reality. The survey revealed that 36% of respondents favourite the traditional Copenhagen interpretation (particularly experimentalists)

Physicists disagree wildly on what quantum mechanics says about reality, Nature survey shows First major attempt to chart researchers’ views finds interpretations in conflict.

This year quantum physics turns 100, so @nature.com decided to conduct the biggest ever survey of what lies behind it. Do physicists really believe in multiple universes? Can influences happen instantaneously? Is the Copenhagen interpretation all it's cut out to be? ⚛️🧪 www.nature.com/articles/d41...

Spain bids €400 million to host mega telescope at risk in US budget cuts New Canary Islands home could save controversial Thirty Meter Telescope first proposed for Hawaii.

The Thirty Meter Telescope — long planned to be built in Hawaii but now with its US funding on the chopping block — could be given a new lease of life in Spain, following a government bid to host at the telescope on La Palma. Will the TMT board accept?
🧪🔭 www.nature.com/articles/d41...

I'm a journalist at Nature and often write and publish stories about astronomy and related topics (policy as well as science)

yes

@bot.astronomy.blue signup

Moonshot AI has now published its technical report about Kimi K2

github.com/MoonshotAI/K...

The 32-page doc says the model trained on NVIDIA H800 GPUs, including post-training reinforcement learning on agentic examples

My story from last week is here: www.nature.com/articles/d41...