UN body’s recommendations driven by AI advances and proliferation of consumer-oriented neurotech devices
It is the latest move in a growing international effort to put guardrails around a burgeoning frontier – technologies that harness data from the brain and nervous system.
Unesco has adopted a set of global standards on the ethics of neurotechnology, a field that has been described as “a bit of a wild west”.
Faire tourner un modèle de langage européen sur votre machine sans avoir besoin d’un serveur surpuissant branché sur une centrale nucléaire, c’est maintenant possible, les amis ! Hé oui,
EuroLLM
vient de prouver qu’on pouvait faire tourner un modèle à 9 milliards de paramètres dans un peu moins de 6 GB de RAM sur un simple laptop.
Une seule commande
Ollama
, et c’est parti mon kiki !!!
Bien sûr, il est encore loin des gros modèles proprio comme GPT-5 mais c’est le enfin le premier LLM européen que VOUS pouvez faire tourner en local. C’est respectueux de votre vie privée, des droits d’auteurs et c’est gratuit !
EuroLLM-9B
, le modèle de base, a été entraîné sur 4 trillions de tokens avec le supercalculateur
MareNostrum 5
à Barcelone (400 GPUs Nvidia H100) et l’architecture utilise du Grouped Query Attention, RoPE, SwiGLU et RMSNorm, comme tout LLM moderne qui se respecte.
Mais il existe d’autres versions comme EuroLLM-1.7B pour smartphones et bientôt EuroLLM-22B pour plus de puissance, ainsi qu’une version vision-language (EuroVLM-9B) et un modèle Mixture-of-Experts (EuroMoE-2.6B).
Et surtout c’est sous licence Apache 2.0. Donc l’usage commercial est autorisé, vous pouvez le fine-tuner sur vos données, et les modifications sont libres, sans redevance à payer. Ce n’est pas la première fois qu’il y a des LLM européens mais ils étaient soit sous licence trop restrictives ou un peu trop lourd pour être utilisé localement par les gens normaux comme vous et moi.
ollama run hf.co/bartowski/EuroLLM-9B-Instruct-GGUF
Ollama télécharge le modèle en version GGUF (format optimisé CPU/GPU), et vous pouvez commencer à discuter. Il existe aussi une version pré-packagée alibayram/erurollm-9b-instruct (attention, erurollm avec un “u”), quantized en Q4_K_M pour réduire la taille à 5,6 GB.
Si vous préférez Python et Hugging Face Transformers :
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "utter-project/EuroLLM-9B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
inputs = tokenizer("Explique-moi ce qu'est un LLM en français simple", return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0]))
Une fois téléchargé, le modèle reste en cache local. Vous pouvez alors l’utiliser offline, sans connexion internet. Et pour les machines avec moins de RAM, la version 1.7B tourne même sur des Raspberry Pi :
ollama run cas/eurollm-1.7b-instruct-q8
Alors pourquoi c’est important ?
EuroLLM ouvre l’IA européenne à des cas d’usage impossibles avec des API cloud. Par exemple une administration publique ne peut pas envoyer ses documents dans le cloud d’OpenAI… ce serait tout livrer aux américains. Elle peut donc fine-tuner EuroLLM localement. Un journaliste en zone sensible sans connexion fiable peut aussi embarquer le modèle sur son ordi portable. Même un chercheur qui manipule des données médicales confidentielles peut le faire en toute confiance avec EuroLLM puisque tout reste sur sa machine.
C’est cool quand même que l’Europe nous file un modèle gratuit qu’on peut installer chez soi et utiliser sans limite. Après c’est vrai que EuroLLM ne bat pas GPT-4 (pas encore) mais il est suffisamment bon pour 80% des cas d’utilisation réels tels que des résumés, de la traduction, des questions-réponses simples, de la génération de code basique.
La roadmap prévoit d’ajouter de la vision et de la voix aux modèles. D’ailleurs, comme je vous le disais, EuroVLM-9B est déjà en preview sur
Hugging Face
, ce qui ouvre la voie à de l’OCR multilingue, de l’analyse de documents visuels, ou encore à la création d’assistants vocaux dans n’importe quelle langue…
Voilà j’ai trouvé ça cool à tester et un grand merci à
Letsar
pour le partage !
Faire tourner un modèle de langage européen sur votre machine sans avoir besoin d’un serveur surpuissant branché sur une centrale nucléaire, c’est maintenant possible, les amis ! Hé oui,
EuroLLM
vient de prouver qu’on pouvait faire tourner un modèle à 9 milliards de paramètres dans un peu moins de 6 GB de RAM sur un simple laptop.
Une seule commande
Ollama
, et c’est parti mon kiki !!!
Bien sûr, il est encore loin des gros modèles proprio comme GPT-5 mais c’est le enfin le premier LLM européen que VOUS pouvez faire tourner en local. C’est respectueux de votre vie privée, des droits d’auteurs et c’est gratuit !
EuroLLM-9B
, le modèle de base, a été entraîné sur 4 trillions de tokens avec le supercalculateur
MareNostrum 5
à Barcelone (400 GPUs Nvidia H100) et l’architecture utilise du Grouped Query Attention, RoPE, SwiGLU et RMSNorm, comme tout LLM moderne qui se respecte.
Mais il existe d’autres versions comme EuroLLM-1.7B pour smartphones et bientôt EuroLLM-22B pour plus de puissance, ainsi qu’une version vision-language (EuroVLM-9B) et un modèle Mixture-of-Experts (EuroMoE-2.6B).
Et surtout c’est sous licence Apache 2.0. Donc l’usage commercial est autorisé, vous pouvez le fine-tuner sur vos données, et les modifications sont libres, sans redevance à payer. Ce n’est pas la première fois qu’il y a des LLM européens mais ils étaient soit sous licence trop restrictives ou un peu trop lourd pour être utilisé localement par les gens normaux comme vous et moi.
ollama run hf.co/bartowski/EuroLLM-9B-Instruct-GGUF
Ollama télécharge le modèle en version GGUF (format optimisé CPU/GPU), et vous pouvez commencer à discuter. Il existe aussi une version pré-packagée alibayram/erurollm-9b-instruct (attention, erurollm avec un “u”), quantized en Q4_K_M pour réduire la taille à 5,6 GB.
Si vous préférez Python et Hugging Face Transformers :
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "utter-project/EuroLLM-9B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
inputs = tokenizer("Explique-moi ce qu'est un LLM en français simple", return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0]))
Une fois téléchargé, le modèle reste en cache local. Vous pouvez alors l’utiliser offline, sans connexion internet. Et pour les machines avec moins de RAM, la version 1.7B tourne même sur des Raspberry Pi :
ollama run cas/eurollm-1.7b-instruct-q8
Alors pourquoi c’est important ?
EuroLLM ouvre l’IA européenne à des cas d’usage impossibles avec des API cloud. Par exemple une administration publique ne peut pas envoyer ses documents dans le cloud d’OpenAI… ce serait tout livrer aux américains. Elle peut donc fine-tuner EuroLLM localement. Un journaliste en zone sensible sans connexion fiable peut aussi embarquer le modèle sur son ordi portable. Même un chercheur qui manipule des données médicales confidentielles peut le faire en toute confiance avec EuroLLM puisque tout reste sur sa machine.
C’est cool quand même que l’Europe nous file un modèle gratuit qu’on peut installer chez soi et utiliser sans limite. Après c’est vrai que EuroLLM ne bat pas GPT-4 (pas encore) mais il est suffisamment bon pour 80% des cas d’utilisation réels tels que des résumés, de la traduction, des questions-réponses simples, de la génération de code basique.
La roadmap prévoit d’ajouter de la vision et de la voix aux modèles. D’ailleurs, comme je vous le disais, EuroVLM-9B est déjà en preview sur
Hugging Face
, ce qui ouvre la voie à de l’OCR multilingue, de l’analyse de documents visuels, ou encore à la création d’assistants vocaux dans n’importe quelle langue…
Voilà j’ai trouvé ça cool à tester et un grand merci à
Letsar
pour le partage !
Widespread adoption of artificial intelligence has been accompanied by new ways to harass women online
Gaatha Sarvaiya would like to post on social media and share her work online. An Indian law graduate in her early 20s, she is in the earliest stages of her career and trying to build a public profile. The problem is, with AI-powered deepfakes on the rise, there is no longer any guarantee that the images she posts will not be distorted into something violating or grotesque.
“The thought immediately pops in that, ‘OK, maybe it’s not safe. Maybe people can take our pictures and just do stuff with them,’” says Sarvaiya, who lives in Mumbai.
US technology company’s engineers want to exploit solar power and the falling cost of rocket launches
Google is hatching plans to put artificial intelligence datacentres into space, with its first trial equipment sent into orbit in early 2027.
Its scientists and engineers believe tightly packed constellations of about 80 solar-powered satellites could be arranged in orbit about 400 miles above the Earth’s surface equipped with the powerful processors required to meet rising demand for AI.
Scientists say almost all have weaknesses in at least one area that can ‘undermine validity of resulting claims’
Experts have found weaknesses, some serious, in hundreds of tests used to check the safety and effectiveness of new artificial intelligence models being released into the world.
Computer scientists from the British government’s AI Security Institute, and experts at universities including Stanford, Berkeley and Oxford, examined more than 440 benchmarks that provide an important safety net.
Agreement to use AWS datacentres, and Nvidia chips inside them, part of $1.4tn spending spree on AI infrastructure
OpenAI has signed a $38bn (£29bn) deal to use Amazon infrastructure to operate its artificial intelligence products, as part of a more than $1tn spending spree on computing power.
The agreement with Amazon Web Services means OpenAI will be able to use AWS datacentres, and the Nvidia chips inside them, immediately.
Unless you’ve been living under a stone, you can’t have failed to notice that 2025 marks the first 100 years of quantum mechanics. A massive milestone, to say the least, about which much has been written in Physics World and elsewhere in what is the International Year of Quantum Science and Technology (IYQ). However, I’d like to focus on a specific piece of quantum technology, namely quantum computing.
I keep hearing about quantum computers, so people must be using them to do cool things, and surely they will soon be as commonplace as classical computers. But as a physicist-turned-engineer working in the aerospace sector, I struggle to get a clear picture of where things are really at. If I ask friends and colleagues when they expect to see quantum computers routinely used in everyday life, I get answers ranging from “in the next two years” to “maybe in my lifetime” or even “never”.
Before we go any further, it’s worth reminding ourselves that quantum computing relies on several key quantum properties, including superposition, which gives rise to the quantum bit, or qubit. The basic building block of a quantum computer – the qubit – exists as a combination of 0 and 1 states at the same time and is represented by a probabilistic wave function. Classical computers, in contrast, use binary digital bits that are either 0 or 1.
Also vital for quantum computers is the notion of entanglement, which is when two or more qubits are co-ordinated, allowing them to share their quantum information. In a highly correlated system, a quantum computer can explore many paths simultaneously. This “massive scale” parallel processing is how quantum may solve certain problems exponentially faster than a classical computer.
The other key phenomenon for quantum computers is quantum interference. The wave-like nature of qubits means that when different probability amplitudes are in phase, they combine constructively to increase the likelihood of the right solution. Conversely, destructive interference occurs when amplitudes are out of phase, making it more likely to get the wrong answer.
Quantum interference is important in quantum computing because it allows quantum algorithms to amplify the probability of correct answers and suppress incorrect ones, making calculations much faster. Along with superposition and entanglement, it means that quantum computers could process and store vast numbers of probabilities at once, outstripping even the best classical supercomputers.
Towards real devices
To me, it all sounds exciting, but what have quantum computers ever done for us so far? It’s clear that quantum computers are not ready to be deployed in the real world. Significant technological challenges need to be overcome before they become fully realisable. In any case, no-one is expecting quantum computers to displace classical computers “like for like”: they’ll both be used for different things.
Yet it seems that the very essence of quantum computing is also its Achilles heel. Superposition, entanglement and interference – the quantum properties that will make it so powerful – are also incredibly difficult to create and maintain. Qubits are also extremely sensitive to their surroundings. They easily lose their quantum state due to interactions with the environment, whether via stray particles, electromagnetic fields, or thermal fluctuations. Known as decoherence, it makes quantum computers prone to error.
That’s why quantum computers need specialized – and often cryogenically controlled – environments to maintain the quantum states necessary for accurate computation. Building a quantum system with lots of interconnected qubits is therefore a major, expensive engineering challenge, with complex hardware and extreme operating conditions. Developing “fault-tolerant” quantum hardware and robust error-correction techniques will be essential if we want reliable quantum computation.
As for the development of software and algorithms for quantum systems, there’s a long way to go, with a lack of mature tools and frameworks. Quantum algorithms require fundamentally different programming paradigms to those used for classical computers. Put simply, that’s why building reliable, real-world deployable quantum computers remains a grand challenge.
What does the future hold?
Despite the huge amount of work that still lies in store, quantum computers have already demonstrated some amazing potential. The US firm D-Wave, for example, claimed earlier this year to have carried out simulations of quantum magnetic phase transitions that wouldn’t be possible with the most powerful classical devices. If true, this was the first time a quantum computer had achieved “quantum advantage” for a practical physics problem (whether the problem was worth solving is another question).
There is also a lot of research and development going on around the world into solving the qubit stability problem. At some stage, there will likely be a breakthrough design for robust and reliable quantum computer architecture. There is probably a lot of technical advancement happening right now behind closed doors.
The first real-world applications of quantum computers will be akin to the giant classical supercomputers of the past. If you were around in the 1980s, you’ll remember Cray supercomputers: huge, inaccessible beasts owned by large corporations, government agencies and academic institutions to enable vast amounts of calculations to be performed (provided you had the money).
And, if I believe what I read, quantum computers will not replace classical computers, at least not initially, but work alongside them, as each has its own relative strengths. Quantum computers will be suited for specific and highly demanding computational tasks, such as drug discovery, materials science, financial modelling, complex optimization problems and increasingly large artificial intelligence and machine-learning models.
These are all things beyond the limits of classical computer resource. Classical computers will remain relevant for everyday tasks like web browsing, word processing and managing databases, and they will be essential for handling the data preparation, visualization and error correction required by quantum systems.
And there is one final point to mention, which is cyber security. Quantum computing poses a major threat to existing encryption methods, with potential to undermine widely used public-key cryptography. There are concerns that hackers nowadays are storing their stolen data in anticipation of future quantum decryption.
Having looked into the topic, I can now see why the timeline for quantum computing is so fuzzy and why I got so many different answers when I asked people when the technology would be mainstream. Quite simply, I still can’t predict how or when the tech stack will pan out. But as IYQ draws to a close, the future for quantum computers is bright.
Modular and scalable: the ICE-Q cryogenics platform delivers the performance and reliability needed for professional computing environments while also providing a flexible and extendable design. The standard configuration includes a cooling module, a payload with a large sample space, and a side-loading wiring module for scalable connectivity (Courtesy: ICEoxford)
At the centre of most quantum labs is a large cylindrical cryostat that keeps the delicate quantum hardware at ultralow temperatures. These cryogenic chambers have expanded to accommodate larger and more complex quantum systems, but the scientists and engineers at UK-based cryogenics specialist ICEoxford have taken a radical new approach to the challenge of scalability. They have split the traditional cryostat into a series of cube-shaped modules that slot into a standard 19-inch rack mount, creating an adaptable platform that can easily be deployed alongside conventional computing infrastructure.
“We wanted to create a robust, modular and scalable solution that enables different quantum technologies to be integrated into the cryostat,” says Greg Graf, the company’s engineering manager. “This approach offers much more flexibility, because it allows different modules to be used for different applications, while the system also delivers the efficiency and reliability that are needed for operational use.”
The standard configuration of the ICE-Q platform has three separate modules: a cryogenics unit that provides the cooling power, a large payload for housing the quantum chip or experiment, and a patent-pending wiring module that attaches to the side of the payload to provide the connections to the outside world. Up to four of these side-loading wiring modules can be bolted onto the payload at the same time, providing thousands of external connections while still fitting into a standard rack. For applications where space is not such an issue, the payload can be further extended to accommodate larger quantum assemblies and potentially tens of thousands of radio-frequency or fibre-optic connections.
The cube-shaped form factor provides much improved access to these external connections, whether for designing and configuring the system or for ongoing maintenance work. The outer shell of each module consists of panels that are easily removed, offering a simple mechanism for bolting modules together or stacking them on top of each other to provide a fully scalable solution that grows with the qubit count.
The flexible design also offers a more practical solution for servicing or upgrading an installed system, since individual modules can be simply swapped over as and when needed. “For quantum computers running in an operational environment it is really important to minimize the downtime,” says Emma Yeatman, senior design engineer at ICEoxford. “With this design we can easily remove one of the modules for servicing, and replace it with another one to keep the system running for longer. For critical infrastructure devices, it is possible to have built-in redundancy that ensures uninterrupted operation in the event of a failure.”
Other features have been integrated into the platform to make it simple to operate, including a new software system for controlling and monitoring the ultracold environment. “Most of our cryostats have been designed for researchers who really want to get involved and adapt the system to meet their needs,” adds Yeatman. “This platform offers more options for people who want an out-of-the-box solution and who don’t want to get hands on with the cryogenics.”
Such a bold design choice was enabled in part by a collaborative research project with Canadian company Photonic Inc, funded jointly by the UK and Canada, that was focused on developing an efficient and reliable cryogenics platform for practical quantum computing. That R&D funding helped to reduce the risk of developing an entirely new technology platform that addresses many of the challenges that ICEoxford and its customers had experienced with traditional cryostats. “Quantum technologies typically need a lot of wiring, and access had become a real issue,” says Yeatman. “We knew there was an opportunity to do better.”
However, converting a large cylindrical cryostat into a slimline and modular form factor demanded some clever engineering solutions. Perhaps the most obvious was creating a frame that allows the modules to be bolted together while still remaining leak tight. Traditional cryostats are welded together to ensure a leak-proof seal, but for greater flexibility the ICEoxford team developed an assembly technique based on mechanical bonding.
The side-loading wiring module also presented a design challenge. To squeeze more wires into the available space, the team developed a high-density connector for the coaxial cables to plug into. An additional cold-head was also integrated into the module to pre-cool the cables, reducing the overall heat load generated by such large numbers of connections entering the ultracold environment.
Flexible for the future: the outer shell of the modules is covered with removable panels that make it easy to extend or reconfigure the system (Courtesy: ICEoxford)
Meanwhile, the speed of the cooldown and the efficiency of operation have been optimized by designing a new type of heat exchanger that is fabricated using a 3D printing process. “When warm gas is returned into the system, a certain amount of cooling power is needed just to compress and liquefy that gas,” explains Kelly. “We designed the heat exchangers to exploit the returning cold gas much more efficiently, which enables us to pre-cool the warm gas and use less energy for the liquefaction.”
The initial prototype has been designed to operate at 1 K, which is ideal for the photonics-based quantum systems being developed by ICEoxford’s research partner. But the modular nature of the platform allows it to be adapted to diverse applications, with a second project now underway with the Rutherford Appleton Lab to develop a module that that will be used at the forefront of the global hunt for dark matter.
Already on the development roadmap are modules that can sustain temperatures as low as 10 mK – which is typically needed for superconducting quantum computing – and a 4 K option for trapped-ion systems. “We already have products for each of those applications, but our aim was to create a modular platform that can be extended and developed to address the changing needs of quantum developers,” says Kelly.
As these different options come onstream, the ICEoxford team believes that it will become easier and quicker to deliver high-performance cryogenic systems that are tailored to the needs of each customer. “It normally takes between six and twelve months to build a complex cryogenics system,” says Graf. “With this modular design we will be able to keep some of the components on the shelf, which would allow us to reduce the lead time by several months.”
More generally, the modular and scalable platform could be a game-changer for commercial organizations that want to exploit quantum computing in their day-to-day operations, as well as for researchers who are pushing the boundaries of cryogenics design with increasingly demanding specifications. “This system introduces new avenues for hardware development that were previously constrained by the existing cryogenics infrastructure,” says Kelly. “The ICE-Q platform directly addresses the need for colder base temperatures, larger sample spaces, higher cooling powers, and increased connectivity, and ensures our clients can continue their aggressive scaling efforts without being bottlenecked by their cooling environment.”
You can find out more about the ICE-Q platform by contacting the ICEoxford team at iceoxford.com, or via email at sales@iceoxford.com. They will also be presenting the platform at the UK’s National Quantum Technologies Showcase in London on 7 November, with a further launch at the American Physical Society meeting in March 2026.
From publishing falsehoods to pushing far-right ideology, Grokipedia gives chatroom comments equal status to research
The eminent British historian Sir Richard Evans produced three expert witness reports for the libel trial involving the Holocaust denier David Irving, studied for a doctorate under the supervision of Theodore Zeldin, succeeded David Cannadine as Regius professor of history at Cambridge (a post endowed by Henry VIII) and supervised theses on Bismarck’s social policy.
That was some of what you could learn from Grokipedia, the AI-powered encyclopedia launched last week by the world’s richest person, Elon Musk. The problem was, as Prof Evans discovered when he logged on to check his own entry, all these facts were false.
When it comes to building a fully functional “fault-tolerant” quantum computer, companies and government labs all over the world are rushing to be the first over the finish line. But a truly useful universal quantum computer capable of running complex algorithms would have to entangle millions of coherent qubits, which are extremely fragile. Because of environmental factors such as temperature, interference from other electronic systems in hardware, and even errors in measurement, today’s devices would fail under an avalanche of errors long before reaching that point.
So the problem of error correction is a key issue for the future of the market. It arises because errors in qubits can’t be corrected simply by keeping multiple copies, as they are in classical computers: quantum rules forbid the copying of qubit states while they are still entangled with others, and are thus unknown. To run quantum circuits with millions of gates, we therefore need new tricks to enable quantum error correction (QEC).
Protected states
The general principle of QEC is to spread the information over many qubits so that an error in any one of them doesn’t matter too much. “The essential idea of quantum error correction is that if we want to protect a quantum system from damage then we should encode it in a very highly entangled state,” says John Preskill, director of the Institute for Quantum Information and Matter at the California Institute of Technology in Pasadena.
There is no unique way of achieving that spreading, however. Different error-correcting codes can depend on the connectivity between qubits – whether, say, they are coupled only to their nearest neighbours or to all the others in the device – which tends to be determined by the physical platform being used. However error correction is done, it must be done fast. “The mechanisms for error correction need to be running at a speed that is commensurate with that of the gate operations,” saysMichael Cuthbert, founding director of the UK’s National Quantum Computing Centre (NQCC). “There’s no point in doing a gate operation in a nanosecond if it then takes 100 microseconds to do the error correction for the next gate operation.”
At the moment, dealing with errors is largely about compensation rather than correction: patching up the problems of errors in retrospect, for example by using algorithms that can throw out some results that are likely to be unreliable (an approach called “post-selection”). It’s also a matter of making better qubits that are less error-prone in the first place.
Qubits are so fragile that their quantum state is very susceptible to the local environment, and can easily be lost through the process of decoherence. Current quantum computers therefore have very high error rates – roughly one error in every few hundred operations. For quantum computers to be truly useful, this error rate will have to be reduced to the scale of one in a million; especially as larger more complex algorithms would require one in a billion or even trillion error rates. This requires real-time quantum error correction (QEC).
To protect the information stored in qubits, a multitude of unreliable physical qubits have to be combined in such a way that if one qubit fails and causes an error, the others can help protect the system. Essentially, by combining many physical qubits (shown above on the left), one can build a few “logical” qubits that are strongly resistant to noise.
According to Maria Maragkou, commercial vice-president of quantum error-correction company Riverlane, the goal of full QEC has ramifications for the design of the machines all the way from hardware to workflow planning. “The shift to support error correction has a profound effect on the way quantum processors themselves are built, the way we control and operate them, through a robust software stack on top of which the applications can be run,” she explains. The “stack” includes everything from programming languages to user interfaces and servers.
With genuinely fault-tolerant qubits, errors can be kept under control and prevented from proliferating during a computation. Such qubits might be made in principle by combining many physical qubits into a single “logical qubit” in which errors can be corrected (see figure 1). In practice, though, this creates a large overhead: huge numbers of physical qubits might be needed to make just a few fault-tolerant logical qubits. The question is then whether errors in all those physical qubits can be checked faster than they accumulate (see figure 2).
The illustration gives an overview of quantum error correction (QEC) in action within a quantum processing unit. UK-based company Riverlane is building its Deltaflow QEC stack that will correct millions of data errors in real time, allowing a quantum computer to go beyond the reach of any classical supercomputer.
Fault-tolerant quantum computing is the ultimate goal, says Jay Gambetta, director of IBM research at the company’s centre in Yorktown Heights, New York. He believes that to perform truly transformative quantum calculations, the system must go beyond demonstrating a few logical qubits – instead, you need arrays of at least a 100 of them, that can perform more than 100 million quantum operations (108 QuOps). “The number of operations is the most important thing,” he says.
It sounds like a tall order, but Gambetta is confident that IBM will achieve these figures by 2029. By building on what has been achieved so far with error correction and mitigation, he feels “more confident than I ever did before that we can achieve a fault-tolerant computer.” Jerry Chow, previous manager of the Experimental Quantum Computing group at IBM, shares that optimism. “We have a real blueprint for how we can build [such a machine] by 2029,” he says (see figure 3).
Others suspect the breakthrough threshold may be a little lower: Steve Brierley, chief executive of Riverlane, believes that the first error-corrected quantum computer, with around 10 000 physical qubits supporting 100 logical qubits and capable of a million QuOps (a megaQuOp), could come as soon as 2027. Following on, gigaQuOp machines (109 QuOps) should be available by 2030–32, and teraQuOps (1012 QuOp) by 2035–37.
Platform independent
Error mitigation and error correction are just two of the challenges for developers of quantum software. Fundamentally, to develop a truly quantum algorithm involves taking full advantage of the key quantum-mechanical properties such as superposition and entanglement. Often, the best way to do that depends on the hardware used to run the algorithm. But ultimately the goal will be to make software that is not platform-dependent and so doesn’t require the user to think about the physics involved.
“At the moment, a lot of the platforms require you to come right down into the quantum physics, which is a necessity to maximize performance,” says Richard Murray of photonic quantum-computing company Orca. Try to generalize an algorithm by abstracting away from the physics and you’ll usually lower the efficiency with which it runs. “But no user wants to talk about quantum physics when they’re trying to do machine learning or something,” Murray adds. He believes that ultimately it will be possible for quantum software developers to hide those details from users – but Brierly thinks this will require fault-tolerant machines.
“In due time everything below the logical circuit will be a black box to the app developers”, adds Maragkou over at Riverlane. “They will not need to know what kind of error correction is used, what type of qubits are used, and so on.” She stresses that creating truly efficient and useful machines depends on developing the requisite skills. “We need to scale up the workforce to develop better qubits, better error-correction codes and decoders, write the software that can elevate those machines and solve meaningful problems in a way that they can be adopted.” Such skills won’t come only from quantum physicists, she adds: “I would dare say it’s mostly not!”
Yet even now, working on quantum software doesn’t demand a deep expertise in quantum theory. “You can be someone working in quantum computing and solving problems without having a traditional physics training and knowing about the energy levels of the hydrogen atom and so on,” says Ashley Montanaro, who co-founded the quantum software company Phasecraft.
On the other hand, insights can flow in the other direction too: working on quantum algorithms can lead to new physics. “Quantum computing and quantum information are really pushing the boundaries of what we think of as quantum mechanics today,” says Montanaro, adding that QEC “has produced amazing physics breakthroughs.”
Early adopters?
Once we have true error correction, Cuthbert at the UK’s NQCC expects to see “a flow of high-value commercial uses” for quantum computers. What might those be?
In this arena of quantum chemistry and materials science, genuine quantum advantage – calculating something that is impossible using classical methods alone – is more or less here already, says Chow. Crucially, however, quantum methods needn’t be used for the entire simulation but can be added to classical ones to give them a boost for particular parts of the problem.
Joint effort In June 2025, IBM in the US and Japan’s national research laboratory RIKEN, unveiled the IBM Quantum System Two, the first to be used outside the US. It involved IBM’s 156-qubit IBM Heron quantum computing system (left) being paired with RIKEN’s supercomputer Fugaku (right) — one of the most powerful classical systems on Earth. The computers are linked through a high-speed network at the fundamental instruction level to form a proving ground for quantum-centric supercomputing. (Courtesy: IBM and RIKEN)
For example, last year researchers at IBM teamed up with scientists at several RIKEN institutes in Japan to calculate the minimum energy state for the iron sulphide cluster (4Fe-4S) at the heart of the bacterial nitrogenase enzyme that fixes nitrogen. This cluster is too big and complex to be accurately simulated using the classical approximations of quantum chemistry. The researchers used a combination of both quantum computing (with IBM’s 72-qubit Heron chip) and RIKEN’s Fugaku high performance computing (HPC). This idea of “improving classical methods by injecting quantum as a subroutine” is likely to be a more general strategy, says Gambetta. “The future of computing is going to be heterogeneous accelerators [of discovery] that include quantum.”
Likewise, Montanaro says that Phasecraft is developing “quantum-enhanced algorithms”, where a quantum computer is used, not to solve the whole problem, but just to help a classical computer in some way. “There are only certain problems where we know quantum computing is going to be useful,” he says. “I think we are going to see quantum computers working in tandem with classical computers in a hybrid approach. I don’t think we’ll ever see workloads that are entirely run using a quantum computer.” Among the first important problems that quantum machines will solve, according to Montanaro, are the simulation of new materials – to develop, for example, clean-energy technologies (see figure 4).
“For a physicist like me,” says Preskill, “what is really exciting about quantum computing is that we have good reason to believe that a quantum computer would be able to efficiently simulate any process that occurs in nature.”
3 Structural insights
(Courtesy: Phasecraft)
A promising application of quantum computers is simulating novel materials. Researchers from the quantum algorithms firm Phasecraft, for example, have already shown how a quantum computer could help simulate complex materials such as the polycrystalline compound LK-99, which was purported by some researchers in 2024 to be a room-temperature superconductor.
Using a classical/quantum hybrid workflow, together with the firm’s proprietary material simulation approach to encode and compile materials on quantum hardware, Phasecraft researchers were able to establish a classical model of the LK99 structure that allowed them to extract an approximate representation of the electrons within the material. The illustration above shows the green and blue electronic structure around red and grey atoms in LK-99.
Montanaro believes another likely near-term goal for useful quantum computing is solving optimization problems – both here and in quantum simulation, “we think genuine value can be delivered already in this NISQ era with hundreds of qubits.” (NISQ, a term coined by Preskill, refers to noisy intermediate-scale quantum computing, with relatively small numbers of rather noisy, error-prone qubits.)
One further potential benefit of quantum computing is that it tends to require less energy than classical high-performance computing, which is notoriously high. If the energy cost could be cut by even a few percent, it would be worth using quantum resources for that reason alone. “Quantum has real potential for an energy advantage,” says Chow. One study in 2020 showed that a particular quantum-mechanical calculation carried out on a HPC used many orders of magnitude more energy than when it was simulated on a quantum circuit. Such comparisons are not easy, however, in the absence of an agreed and well-defined metric for energy consumption.
Building the market
Right now, the quantum computing market is in a curious superposition of states itself – it has ample proof of principle, but today’s devices are still some way from being able to perform a computation relevant to a practical problem that could not be done with classical computers. Yet to get to that point, the field needs plenty of investment.
The fact that quantum computers, especially if used with HPC, are already unique scientific tools should establish their value in the immediate term, says Gambetta. “I think this is going to accelerate, and will keep the funding going.” It is why IBM is focusing on utility-scale systems of around 100 qubits or so and more than a thousand gate operations, he says, rather than simply trying to build ever bigger devices.
Montanaro sees a role for governments to boost the growth of the industry “where it’s not the right fit for the private sector”. One role of government is simply as a customer. For example, Phasecraft is working with the UK national grid to develop a quantum algorithm for optimizing the energy network. “Longer-term support for academic research is absolutely critical,” Montanaro adds. “It would be a mistake to think that everything is done in terms of the underpinning science, and governments should continue to support blue-skies research.”
The road ahead IBM’s current roadmap charts how the company plans on scaling up its devices to achieve a fault-tolerant device by 2029. Alongside hardware development, the firm will also focus on developing new algorithms and software for these devices. (Courtesy: IBM)
It’s not clear, though, whether there will be a big demand for quantum machines that every user will own and run. Before 2010, “there was an expectation that banks and government departments would all want their own machine – the market would look a bit like HPC,” Cuthbert says. But that demand depends in part on what commercial machines end up being like. “If it’s going to need a premises the size of a football field, with a power station next to it, that becomes the kind of infrastructure that you only want to build nationally.” Even for smaller machines, users are likely to try them first on the cloud before committing to installing one in-house.
According to Cuthbert , the real challenge in the supply-chain development is that many of today’s technologies were developed for the science community – where, say, achieving millikelvin cooling or using high-power lasers is routine. “How do you go from a specialist scientific clientele to something that starts to look like a washing machine factory, where you can make them to a certain level of performance,” while also being much cheaper, and easier to use?
But Cuthbert is optimistic about bridging this gap to get to commercially useful machines, encouraged in part by looking back at the classical computing industry of the 1970s. “The architects of those systems could not imagine what we would use our computation resources for today. So I don’t think we should be too discouraged that you can grow an industry when we don’t know what it’ll do in five years’ time.”
Montanaro too sees analogies with those early days of classical computing. “If you think what the computer industry looked like in the 1940s, it’s very different from even 20 years later. But there are some parallels. There are companies that are filling each of the different niches we saw previously, there are some that are specializing in quantum hardware development, there are some that are just doing software.” Cuthbert thinks that the quantum industry is likely to follow a similar pathway, “but more quickly and leading to greater market consolidation more rapidly.”
However, while the classical computing industry was revolutionized by the advent of personal computing in the 1970s and 80s, it seems very unlikely that we will have any need for quantum laptops. Rather, we might increasingly see apps and services appear that use cloud-based quantum resources for particular operations, merging so seamlessly with classical computing that we don’t even notice.
That, perhaps, would be the ultimate sign of success: that quantum computing becomes invisible, no big deal but just a part of how our answers are delivered.
In the first instalment of this two-part article, Philip Ball explores the latest developments in the quantum-computing industry
This episode of the Physics World Weekly podcast explores how quantum computing and artificial intelligence can be combined to help physicists search for rare interactions in data from an upgraded Large Hadron Collider.
My guest is Javier Toledo-Marín, and we spoke at the Perimeter Institute in Waterloo, Canada. As well as having an appointment at Perimeter, Toledo-Marín is also associated with the TRIUMF accelerator centre in Vancouver.
Pendant 30 ans, les experts en informatique quantique vous demandaient de les croire sur parole du genre “Mon ordi quantique est 13 000 fois plus rapides que ton PC Windows XP…”. Mais bon, ils sont rigolo car c’était impossible à vérifier ce genre de conneries… M’enfin ça c’était jusqu’à présent car
Google vient d’annoncer Quantum Echoes
, et on va enfin savoir grâce à ce truc, ce que l’informatique quantique a vraiment dans le ventre.
Depuis 2019 et
la fameuse “suprématie quantique” de Google
, on était en fait coincé dans un paradoxe de confiance assez drôle. Google nous disait “regardez, on a résolu un problème qui prendrait 10 milliards de milliards d’années à un supercalculateur”. Bon ok, j’veux bien les croire mais comment on vérifie ? Bah justement, on pouvait pas ! C’est un peu comme les promesses des gouvernements, ça n’engage que les gros teubés qui y croient ^^.
Heureusement grâce à Quantum Echoes, c’est la fin de cette ère du “Faites-nous confiance” car pour la première fois dans l’histoire de l’informatique quantique,
un algorithme peut être vérifié de manière reproductible
. Vous lancez le calcul sur la puce Willow de Google, vous obtenez un résultat. Vous relancez, vous obtenez le même. Votre pote avec un ordi quantique similaire lance le même truc, et il obtient le même résultat. Ça semble basique, mais pour le quantique, c’est incroyable !!
Willow, la puce quantique de Google
L’algorithme en question s’appelle OTOC (Out-Of-Time-Order Correlator), et il fonctionne comme un écho ultra-sophistiqué. Vous envoyez un signal dans le système quantique, vous perturbez un qubit, puis vous inversez précisément l’évolution du signal pour écouter l’écho qui revient. Cet écho quantique se fait également amplifier par interférence constructive, un phénomène où les ondes quantiques s’additionnent et deviennent plus fortes. Du coup, ça permet d’obtenir une mesure d’une précision hallucinante.
En partenariat avec l’Université de Californie à Berkeley, Google a testé ça sur deux molécules, une de 15 atomes et une autre de 28 atomes et les résultats obtenus sur leur ordinateur quantique correspondaient exactement à ceux de la RMN (Résonance Magnétique Nucléaire) traditionnelle. Sauf que Quantum Echoes
va 13 000 fois plus vite
qu’un supercalculateur classique pour ce type de calcul.
En gros, ce qui aurait pris 3 ans sur une machine classique prend 2 heures sur un Willow.
Cette vitesse, c’est impressionnant mais ce qui change la donne dans cette annonce, c’est cette notion de vérifiabilité ! Bref, c’est fini le bullshit, maintenant la structure de systèmes quantiques (des molécules aux aimants en passant par les trous noirs) sera vérifiable et comparable.
Et les applications concrètes sont déjà plutôt bien identifiées : Découverte de médicaments, pour comprendre comment les molécules se lient à leurs cibles, la science des matériaux, pour caractériser la structure moléculaire de nouveaux polymères ou les composants de batteries, la fusion nucléaire…etc tout ce qui nécessite de modéliser des phénomènes quantiques avec une précision extrême !
Google compare ça à un “quantum-scope”, capable de mesurer des phénomènes naturels auparavant inobservables un peu comme l’ont été le télescope et le microscope qui nous ont donné accès à de nouveaux mondes invisibles. Le Quantum Echoes nous donne un accès ce monde quantique sauf que cette fois, on pourra vérifier que la réalité est identique à celle annoncée par les scientifiques.
This year’s Q2B meeting took place at the end of last month in Paris at the Cité des Sciences et de l’Industrie, a science museum in the north-east of the city. The event brought together more than 500 attendees and 70 speakers – world-leading experts from industry, government institutions and academia. All major quantum technologies were highlighted: computing, AI, sensing, communications and security.
As was clear from the conference talks, quantum computers are undoubtedly advancing in leaps and bounds. One of their most important weak points, however, is that their fundamental building blocks (quantum bits, or qubits) are highly prone to errors. These errors are caused by interactions with the environment – also known as noise – and correcting them will require innovative software and hardware. Today’s machines are only capable of running on average a few hundred operations before an error occurs; but in the future, we will have to develop quantum computers capable of processing a million error-free quantum operations (known as a MegaQuOp) or even a trillion error-free operations (TeraQuOps).
QEC works by distributing one quantum bit of information – called a logical qubit – across several different physical qubits, such as superconducting circuits or trapped atoms. Each physical qubit is noisy, but they work together to preserve the quantum state of the logical qubit – at least for long enough to perform a calculation. It was Peter Shor who first discovered this method of formulating a quantum error correcting code by storing the information of one qubit onto a highly entangled state of nine qubits. A technique known as syndrome decoding is then used to diagnose which error was the likely source of corruption on an encoded state. The error can then be reversed by applying a corrective operation depending on the syndrome.
Computing advances A prototype quantum computer from NVIDIA that makes use of seven qubits. (Courtesy: Isabelle Dumé)
While error correction should become more effective as the number of physical qubits in a logical qubit increases, adding more physical qubits to a logical qubit also adds more noise. Much progress has been made in addressing this and other noise issues in recent years, however.
“We can say there’s a ‘fight’ when increasing the length of a code,” explains Hilaire. “Doing so allows us to correct more errors, but we also introduce more sources of errors. The goal is thus being able to correct more errors than we introduce. What I like with this picture is the clear idea of the concept of a fault-tolerant threshold below which fault-tolerant quantum computing becomes feasible.”
Developments in QEC theory
Speakers at the Q2B25 meeting shared a comprehensive overview of the most recent advancements in the field – and they are varied. First up, concatenated error correction codes. Prevalent in the early days of QEC, these fell by the wayside in favour of codes like surface code, but are making a return as recent work has shown. Concatenated codes can achieve constant encoding rates and a quantum computer operating on a linear, nearest-neighbour connectivity was recently put forward. Directional codes, the likes of which are being developed by Riverlane, are also being studied. These leverage native transmon qubit logic gates – for example, iSWAP gates – and could potentially outperform surface codes in some aspects.
The panellists then described bivariate bicycle codes, being developed by IBM, which offer better encoding rates than surface codes. While their decoding can be challenging for real-time applications, IBM’s “relay belief propagation” (relay BP) has made progress here by simplifying decoding strategies that previously involved combining BP with post-processing. The good thing is that this decoder is actually very general and works for all the “low-density parity check codes” — one of the most studied class of high performance QEC codes (these also include, for example, surface codes and directional codes).
There is also renewed interest in decoders that can be parallelized and operate locally within a system, they said. These have shown promise for codes like the 1D repetition code, which could revive the concept of self-correcting or autonomous quantum memory. Another possibility is the increased use of the graphical language ZX calculus as a tool for optimizing QEC circuits and understanding spacetime error structures.
Hardware-specific challenges
The panel stressed that to achieve robust and reliable quantum systems, we will need to move beyond so-called hero experiments. For example, the demand for real-time decoding at megahertz frequencies with microsecond latencies is an important and unprecedented challenge. Indeed, breaking down the decoding problem into smaller, manageable bits has proven difficult so far.
There are also issues with qubit platforms themselves that need to be addressed: trapped ions and neutral atoms allow for high fidelities and long coherence times, but they are roughly 1000 times slower than superconducting and photonic qubits and therefore require algorithmic or hardware speed-ups. And that is not all: solid-state qubits (such as superconducting and spin qubits) suffer from a “yield problem”, with dead qubits on manufactured chips. Improved fabrication methods will thus be crucial, said the panellists.
Collaboration between academia and industry
The discussions then moved towards the subject of collaboration between academia and industry. In the field of QEC, such collaboration is highly productive today, with joint PhD programmes and shared conferences like Q2B, for example. Large companies also now boast substantial R&D departments capable of funding high-risk, high-reward research, blurring the lines between fundamental and application-oriented research. Both sectors also use similar foundational mathematics and physics tools.
At the moment there’s an unprecedented degree of openness and cooperation in the field. This situation might change, however, as commercial competition heats up, noted the panellists. In the future, for example, researchers from both sectors might be less inclined to share experimental chip details.
Last, but certainly not least, the panellists stressed the urgent need for more PhDs trained in quantum mechanics to address the talent deficit in both academia and industry. So, if you were thinking of switching to another field, perhaps now could be the time to jump.
A new optically addressable quantum bit (qubit) encoded in a fluorescent protein could be used as a sensor that can be directly produced inside living cells. The device opens up a new era for fluorescence microscopy to monitor biological processes, say the researchers at the University of Chicago Pritzker School of Molecular Engineering who designed the novel qubit.
Quantum technologies use qubits to store and process information. Unlike classical bits, which can exist in only two states, qubits can exist in a superposition of both these states. This means that computers employing these qubits can simultaneously process multiple streams of information, allowing them to solve problems that would take classical computers years to process.
Qubits can be manipulated and measured with high precision, and in quantum sensing applications they act as nanoscale probes whose quantum state can be initialized, coherently controlled and read out. This allows them to detect minute changes in their environment with exquisite sensitivity.
Optically addressable qubit sensors – that is, those that are read out using light pulses from a laser or other light source – are able to measure nanoscale magnetic fields, electric fields and temperature. Such devices are now routinely employed by researchers working in the physical sciences. However, their use in the life sciences is lagging behind, with most applications still at the proof-of-concept stage.
Difficult to position inside living cells
Many of today’s quantum sensors are based on nitrogen-vacancy (NV) centres, which are crystallographic defects in diamond. These centres occur when two neighbouring carbon atoms in diamond are replaced by a nitrogen atom and an empty lattice site and they act like tiny quantum magnets with different spins. When excited with laser pulses, the fluorescent signal that they emit can be used to monitor slight changes in the magnetic properties of a nearby sample of material. This is because the intensity of the emitted NV centre signal changes with the local magnetic field.
“The problem is that such sensors are difficult to position at well-defined sites inside living cells,” explains Peter Maurer, who co-led this new study together with David Awschalom. “And the fact that they are typically ten times larger than most proteins further restricts their applicability,” he adds.
“So, rather than taking a conventional quantum sensor and trying to camouflage it to enter a biological system, we therefore wanted to explore the idea of using a biological system itself and developing it into a qubit,” says Awschalom.
Fluorescent proteins, which are just 3 nm in diameter, could come into their own here as they can be genetically encoded, allowing cells to produce these sensors directly at the desired location with atomic precision. Indeed, fluorescent proteins have become the “gold standard” in cell biology thanks to this unique ability, says Maurer. And decades of biochemistry research has allowed researchers to generate a vast library of such fluorescent proteins that can be tagged to thousands of different types of biological targets.
“We recognized that these proteins possess optical and spin properties that are strikingly similar to those of qubits formed by crystallographic defects in diamond – namely that they have a metastable triplet state,” explain Awschalom and Maurer. “Building on this insight, we combined techniques from fluorescence microscopy with methods of quantum control to encode and manipulate protein-based qubits.”
In their work, which is detailed in Nature, the researchers used a near-infrared laser pulse to optically address a yellow fluorescent protein known as EYFP and read out its triplet spin state with up to 20% “spin contrast” – measured using optically detected magnetic resonance (ODMR) spectroscopy.
To test the technique, the team genetically modified the protein so that it was expressed in human embryonic kidney cells and Escherichia coli (E. coli) cells. The measured OMDR signals exhibited a contrast of up to 8%. While this performance is not as good as that of NV quantum sensors, the fluorescent proteins open the door to magnetic resonance measurements directly inside living cells – something that NV centres cannot do, says Maurer. “They could thus transform medical and biochemical studies by probing protein folding, monitoring redox states or detecting drug binding at the molecular scale,” he tells Physics World.
“A new dimension for fluorescence microscopy”
Beyond sensing, the unique quantum resonance “signatures” offer a new dimension for fluorescence microscopy, paving the way for highly multiplexed imaging far beyond today’s colour palette, Awschalom adds. Looking further ahead, using arrays of such protein qubits could even allow researchers to explore many-body quantum effects within biologically assembled structures.
Maurer, Awschalom and colleagues say they are now busy trying to improve the stability and sensitivity of their protein-based qubits through protein engineering via “directed evolution” – similar to the way that fluorescent proteins were optimized for microscopy.
“Another goal is to achieve single-molecule detection, enabling readout of the quantum state of individual protein qubits inside cells,” they reveal. “We also aim to expand the palette of available qubits by exploring new fluorescent proteins with improved spin properties and to develop sensing protocols capable of detecting nuclear magnetic resonance signals from nearby biomolecules, potentially revealing structural changes and biochemical modifications at the nanoscale.”
Dans le cas de la réalité augmentée, les utilisateurs pourront télécharger en 5G les maps 3D traitées en temps réel depuis des infrastructure Edge Computing. D’autre part, les informations associées à ces Maps 3D proviendront d’une source de données plus éloignée : le cloud.
Connexion
