Loading today’s stories
Science1 distinct publisher3 min readPublished
The method is a brute search over combinations of equipment already sitting in the lab rather than anything like a chatbot, which moves the scarce human skill from wiring the setup to writing down exactly what the setup should maximise.
The Scientist · Science desk
science
Vienna's plan to put a trapped-ion quantum computer inside an electron microscope1 distinct publisher
science
A pole-on magnetar hands vacuum birefringence its first astrophysical candidate1 distinct publisher
science
A phage kinase with no target list: EMBL finds one enzyme that breaks several bacterial defences1 distinct publisher
science
HIPAA Covers Less Than You Think, And "Anonymized" Is Not A Legal Shield1 distinct publisher
Compiled by The ScientistSomething wrong?How this is made
An optimizer maximises what you wrote down, inside the simulator you handed it. Mario Krenn locates the real difficulty in the specification rather than the search: defining the objective as accurately as possible stays a human task [16]. The search itself he describes as an enormous optimization problem over an overwhelmingly large space of experiments buildable from available components [8], which is a different machine from a language model trained to produce what is statistically likely [9].
Simulation is what closes the loop. Computers can now model a wide range of physical situations in manageable time [14], and Krenn's stated ambition is a universal physics simulator, on the argument that a few fundamental equations already take you a very long way [15]. That also predicts where the method has landed so far. Beyond the original quantum setup, the reported applications are fusion reactors, particle detectors and gravitational-wave detector sensitivity [12], three domains [1] whose forward physics is characterised well enough that a simulated score means something.
What this account leaves out is the magnitude of the improvement. There is no effect size, no count of AI-proposed setups that were subsequently built and measured, and no denominator for searches that came back with nothing usable [2]. That last gap matters more than it looks: a configuration that scores higher inside a simulator has beaten a model of the apparatus, and the apparatus is where alignment error, drift and noise live.
The founding story is worth reading as an anecdote, which is what it is. Krenn's group in Vienna had already failed to find a configuration that could demonstrate the quantum effects they were after [4]. He wrote the search in a few hours, left the machine running overnight [6], and found a proposal the next morning that satisfied the criteria [7]. It is a genuine result and a fair advertisement, but it is also not a matched comparison in which human designers were given the same objective and the same clock.
Philipp Haslinger of TU Wien offers the more useful signal for anyone deciding where to spend attention. In electron microscopy, he says, systematic work with entanglement and quantum-mechanical microscopy concepts is only beginning and human intuition is often very limited [10], so a search can surface microscope designs a person would probably never have proposed [11]. Read that as a claim about relative advantage rather than absolute power: the method pays most where the human prior is thinnest, and least where a century of accumulated intuition already covers the space.
That leaves the review problem. Krenn says some generated proposals are quickly understood, and others can be shown by calculation to perform better without anyone being able to say in words why [13]. A group can check such a design and build it. Arguing with it is harder, and that is a change in the relationship between an experimentalist and the instrument on the bench.
Ranked by verification strength, evidence, and original report placement.
Krenn: "It is an enormous optimization problem. There is an overwhelmingly large space of possible experiments that can be built from the available components. The computer has to search this space systematically in order to find the best possible solution."
The approach has little in common with large language models, which are trained on enormous amounts of data and then generate solutions that are statistically likely.
The approach is possible because modern computers can simulate a wide range of physical situations within a manageable amount of time.
An international research team presented the current state of AI-designed experiments in the journal Nature.
Mario Krenn is a professor of machine learning in science at the University of Tuebingen.
As a student in Vienna working on the setup for a quantum experiment, neither Krenn nor other members of his research group could find an experimental configuration capable of demonstrating the desired quantum effects.
Distinct publishers with included, body-backed reporting in this cluster.
phys.org
1 article · September 3, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One account, one DOI, two insiders
The underlying work is a peer-reviewed Nature survey with a printed DOI, which is the strongest thing going for it. Everything a reader actually receives, though, arrives through one phys.org write-up built from quotation: Krenn on his own method, Haslinger on his hopes for electron microscopy. Nothing in it can be checked without going to the paper.
Asserted use, nothing counted
Real use is claimed four times over — Krenn's overnight quantum-optics setup, then fusion reactors, particle detectors and gravitational-wave sensitivity — and not once with a facility, a date or a measured improvement attached. One first-hand success plus a three-item list is a foothold, not a track record.
A clear yes with nothing behind it
phys.org answers its own question with 'a clear yes' and reaches for a universal physics simulator, while the only figures in the whole piece are a few hours of coding and one night of runtime. The overstatement is in the register rather than the facts: the method plainly works somewhere, but 'more precise than human-designed experiments' is asserted, not shown, and the headline's hedged 'could outperform' sits oddly against the text's confidence.
Both voices are inside the method
The two people who speak are the chair who pioneered the technique and a centre head whose microscopy programme stands to gain from it. That is not a reason to doubt them, but it explains the absent material: no null results, no cost of running these searches, no physicist who tried a generated design and disliked it. A press account assembled from institutional communications rarely goes looking for that.
Clear text, one window
What phys.org says is unambiguous and internally consistent, and the mechanism it describes — describe components, simulate, search — is easy to follow and hard to misread. Confidence stops there: with one publisher and no numbers, we can be sure what is being claimed and much less sure how far it reaches.