[ad_1]
Nature, Published online: 02 December 2024; doi:10.1038/d41586-024-03940-y
How many researchers are using the AI tool? Nature gathers data and talks to members of the academic community.
[ad_2]
Source link
Tag: Computer science
-
<b>ChatGPT turns two: how the AI chatbot has changed scientists’ </b><b>lives</b>
-
DARE project puts Europe on the map for chip development
[ad_1]
As part of the DARE project, the Barcelona Supercomputing Center will coordinate the development of Europe’s hardware and software ecosystem with an initial investment of €240m.
The DARE (Digital Autonomy with RISC-V Europe) project for developing high-performance computing chips is considered “a large-scale strategic initiative that, in terms of impact and investment, could be as groundbreaking as CERN, Galileo, or Airbus.”
A statement issued by the Government of the Generalitat provides more details about this initiative, which was announced last Friday following the meeting between the President of the Government, Pedro Sánchez, and the President of the Generalitat, Salvador Illa.
DARE project develops critical high-performance supercomputers
Half of the investment in the DARE project, €120m, will be provided by the European Commission through EuroHPC JU (European Joint Undertaking for High-Performance Computing).
The Ministry of Science, Innovation and Universities will contribute €34.4m, while the rest of the funding will come from contributions from 45 European partners participating in the project.
The Barcelona Supercomputing Center – National Supercomputing Center (BSC-CNS), home to the MareNostrum 5 supercomputer, will lead the development of the hardware and software ecosystem for critical high-performance computing (HPC) and artificial intelligence applications over the next three years.
RISC-V is a type of open-source architecture for designing processors, so no licensing fees are required.
Its significance lies in its potential to pave the way for European technological sovereignty, which currently relies on American and Asian companies in the chip field.
Ensuring Europe’s independence in the digital ecosystem
The BSC has been working for 10 years on developing RISC-V chips and has already coordinated the EPI (European Processor Initiative) project.
The centre’s director, Mateo Valero, explained that DARE “will develop the hardware and software ecosystem needed to create European supercomputers like the future MareNostrum 6, autonomous cars, and an ethical European AI.”
Europe needs its own technology that ensures the security and autonomy of its digital infrastructures.
The Minister of Science, Innovation, and Universities, Diana Morant, stated: “The DARE project will position Spain on the European map for the development of chips, a key strategic sector to address complex issues in areas such as health, climate change, and energy.”
[ad_2]
Source link -
This billion-dollar firm plans to build giant quantum computers from light. Can it succeed?
[ad_1]
A prototype cryogenic cabinet in which PsiQuantum is testing quantum computing using light in silicon chips.Credit: PsiQuantum
During his years in academia, Jeremy O’Brien says he had “precisely zero ambition to go into business”. Yet today, the Australian physicist runs a private quantum-computing firm that venture-capital funders and governments have rushed to invest in — and which is making some of the field’s boldest promises.
In a whirlwind nine years since O’Brien founded PsiQuantum with three academic colleagues, the company has quietly raised more than US$1 billion and values itself at more than $3 billion — meaning that its coffers probably rival those for internal quantum-computing efforts at Google or IBM. In the past year alone, PsiQuantum, which has 350 staff and is based in Palo Alto, California, has scored major investments from governments in Australia and the United States, adding to previous private funding rounds. “They have received one of the biggest venture-capital investments in the quantum community,” says Doug Finke, a computer scientist in Orange County, California, who works at the business-analysis firm Global Quantum Intelligence.
All that investment is chasing an audacious goal: using light in silicon chips to create a giant programmable quantum computer that can outperform classical machines — and to do it soon. By the end of 2027, the firm’s researchers told Nature, PsiQuantum aims to be operating a photonic quantum computer that can run commercially useful problems and is ‘fault-tolerant’: that is, it makes computations possible by correcting for the errors that are inherent in these fragile systems. If they succeed, this would put the firm ahead of its major rivals and leapfrog researchers doing toy problems on small-scale quantum computers.
Google uncovers how quantum computers can beat today’s best supercomputers
Yet compared with its competitors, PsiQuantum has shown very little. Rather than building up gradually, as others have done, by debuting systems of tens or hundreds of quantum bits, PsiQuantum is aiming to jump to a machine that will require something in the order of one million qubits. (PsiQuantum’s researchers haven’t published a specific number.) To do that, it will need to overcome technical challenges it has not proved it can solve, says Chao-Yang Lu, a physicist working on photonic quantum computing at the University of Science and Technology of China in Shanghai. He is one of several scientists who worry the firm is promising things it will struggle to deliver.
“My impression is there’s a lot of scepticism about how much progress PsiQuantum has made,” says Shimon Kolkowitz, a quantum physicist at the University of California, Berkeley. He calls a bet on them “extremely high risk”.
PsiQuantum researchers say the firm has achieved more than it has publicly shown, and that funders have scrutinized its plans. O’Brien himself talks about challenges in the past tense and insists that there is little doubt of success. And some independent researchers see its plans as being at least plausible. “I think it’s an amazing gamble,” says Pascale Senellart, a quantum optical physicist at the French National Centre for Scientific Research in Palaiseau. “It’s really worth exploring.”
Flying qubits
PsiQuantum’s approach is radically different to that of some major rivals (see ‘Comparing quantum computers’ at the end of this article), because of its choice of qubit — the basic unit of quantum information. Whereas the binary digits (bits) of classical computers encode either a 1 or a 0, qubits can be put into a ‘superposition’ — existing in two states at once, a combination of both 1 and 0, with a chance of being measured as either. Calculations come from ‘entangling’ these qubits, meaning that their quantum states become intrinsically linked and interdependent. To prevent errors from destroying the calculations, a quantum computer will need around 10,000 physical qubits working together to make each useful ‘logical’ qubit, O’Brien says. With a few hundred of these, researchers hope that quantum computers will be able to perform complex calculations, such as modelling chemical processes at the quantum level, that would be much too difficult for a classical machine.
Many firms in the field make their qubits from atoms, ions or tiny rings of a superconducting material — in each case, a physical object that has some mass and is often fixed in place. But PsiQuantum is one of a handful of companies that uses massless particles of light, or photons — sometimes known as flying qubits.
The idea to use light as a qubit isn’t new. In the early 2000s, optical quantum computing was one of the first platforms to be explored experimentally. Some of PsiQuantum’s founders were involved in the field’s birth, says Senellart, who is a co-founder of Quandela, a firm based near Paris that makes photonic quantum computers.
Making a quantum computer with light is “on paper, quite easy”, she says. PsiQuantum creates qubits by using an optical device called a beam splitter to send a single photon simultaneously down two routes (known as waveguides) etched into silicon. Because photons have no charge or mass, they are largely unaffected by their surroundings. This means that, even at room temperature, photon-based qubits are insensitive to many types of noise that plague rival hardware. This ability to maintain quantum information and travel long distances at speed makes it easy to build big and fast systems. “That’s a huge asset,” says Senellart.
IBM releases first-ever 1,000-qubit quantum chip
But photons also come with hurdles. It is hard to generate single, near-identical photons on demand. They are readily absorbed and lost. And getting the flighty particles to interact is a challenge. Although light waves do interfere with each other, that kind of interaction alone is not sufficient for constructing multi-qubit gates, in which qubits entangle to form basic logic operations.
To generate photons, the firm pumps laser light through silicon. These sources are probabilistic: they produce pairs of photons perhaps once in every 20 attempts. Having a pair is necessary because the spare, or ‘herald’, photon provides a heads-up that allows the computer to use the surviving photon.
With this strategy, each chip needs many such sources, as well as super-efficient waveguides and optical components that can handle photons without losing them.
To perform logic operations, PsiQuantum first builds up clusters of entangled photon qubits by bringing photons together so that their light waves interfere, then making measurements on some of them in ways that entangle the remaining qubits. Calculations then occur by performing a succession of such measurements on pairs of photons from different clusters. Those measurements destroy the pairs but entangle their clusters, a technique known as fusion-based quantum computing (see ‘Quantum computers with ‘flying qubits’’). Senellart says that PsiQuantum has an impressive team working on developing the theory behind such a computer. “They are coming up with a lot of smart schemes,” she says.
PsiQuantum’s approach continually generates and destroys photons, with each qubit needing to exist only for long enough to be entangled or measured with another, not for the duration of the calculation. And although each operation involves an element of chance and photons will get lost, failures are detectable as part of the measurement, says Mercedes Gimeno-Segovia, a physicist at PsiQuantum who is working on the computer’s architecture.
“It’s an incredibly forgiving way of doing quantum computing,” says Andrew White, an optical physicist at the University of Queensland in Brisbane, Australia, and a former academic colleague of O’Brien. “You can take very high error rates and still have it scale.”
PsiQuantum is not the only company pursuing quantum computing using light. Most firms plan to eventually scale up their systems by entangling devices using photonic qubits, so mastering them now makes sense, says Senellart. None is close to PsiQuantum’s billion-dollar backing. The company Xanadu, in Toronto, Canada, uses as its qubits ‘squeezed’ states, which encode information into the electromagnetic field of multiple photons. It has created a machine of more than 200 qubits and has raised a few hundred million dollars in funding. Quandela, which has raised $67.5 million, sells small machines that use on-demand sources of single photons, to boost efficiency and reduce a computer’s size. And ORCA Computing, based in London, is developing a way to store single photons as a short-term quantum memory; it has raised at least $15 million.
PsiQuantum’s pitch
PsiQuantum’s physicists say they have the advantage of building on the foundations of two existing mega industries: semiconductor manufacturing for computers, and photonics, which creates the fibre optics used in telecommunications cables. The firm’s chips — made in partnership with US semiconductor giant Global Foundries — combine conventional technology such as beam splitters with components such as single-photon detectors and sources, which are rarely used outside laboratories. These are all etched onto silicon wafers, where they manipulate telecoms-frequency photons.
“In a way, we’re taking existing technologies and making them behave quantum mechanically, as opposed to inventing completely new technology, then trying to figure out how to scale it,” says Mark Thompson, one of PsiQuantum’s co-founders and its chief technologist.
For its system to work, the company needs millions of precision electronic components that must operate at unprecedented efficiency and can’t simply be taken off the shelf. Detectors, for example, need to be chilled to around 4 kelvin using liquid helium. The biggest challenge, says O’Brien, has been creating optical switches to divert photons into calculations. For this, the company built its own facility to grow high-purity wafers of barium titanate, a material that can steer light efficiently. The firm spent a vast amount of money and made a big bet, says O’Brien: this leap of faith is one of the things he is proudest of.
Outsiders must take O’Brien and his colleagues largely at their word that they have solved such challenges. The firm, which once had a reputation for being secretive, is starting to open up. In April, it published a preprint outlining its hardware platform, which White calls jaw-dropping (K. Alexander et al. Preprint at arXiv https://doi.org/gtzxqj; 2024). It shows efficiencies, such as in getting light off a chip into a fibre, that are “well ahead of what the best university labs could do”, he says.
A PsiQuantum silicon photonic wafer being tested at the firm’s facility in Palo Alto, California.Credit: PsiQuantum
Some of the demonstrations in the paper are “extremely impressive”, says Senellart. But others are missing, such as the rate of photons produced per second, she adds. What’s more, some of the numbers are not as good as they need to be, says Graeme Smith, a quantum physicist at the University of Waterloo in Canada. For example, he says, the likelihood that a heralded particle is detected is reported as 26%, when it needs to be more than 50%. “It is not encouraging that, after many years, they are still struggling with good single photon sources, since it is the most basic building block of their proposed architecture,” he says. Any photonic quantum computer will need to reduce photon losses to levels never previously demonstrated at scale, says Stephanie Simmons, a quantum physicist at Simon Fraser University in Burnaby, Canada.
In response, Thompson says the team’s papers are “just the tip of the iceberg” of what PsiQuantum has achieved. Commercial secrecy might be one reason that the team is not publishing its best work. “To publish in academic journals is very time-consuming and quite a distraction,” he says.
But PsiQuantum’s approach also feels like a gamble, says Senellart, because it has opted not to publicly show small demonstrations, often called noisy intermediate-scale quantum (NISQ) devices. “Personally, I would sleep better” with the incremental approach, she says. If the firm can manage without, “it’s just amazing”.
Missing out this stage means no working prototypes are available, and has irked academics who like to see concrete milestones. With no NISQ computer, the firm has been less active than some at recruiting end users or software partners, says Finke, adding to its air of mystery. White says: “Publicly, they haven’t demonstrated in their papers anything other than entanglement between a couple of photons.”
O’Brien says this is because proving a computer works with, say, 100 qubits, reveals nothing about whether it will scale to one million. Academic demonstrations tend to use shortcuts that wouldn’t fly at scale, adds Pete Shadbolt, one of the firm’s co-founders and its chief scientific officer. And the dearth of successful applications for small, noisy systems justifies the company’s strategy, says O’Brien. “We always knew that a useful quantum computer is going to be a big machine.”
Indeed, the US physicist who coined the term NISQ, John Preskill, said last year that even he thought that no useful applications had emerged from the NISQ era.
The company says it is building internal prototypes of increasing scale and complexity, but doesn’t market them as quantum computers. The goal is to find where the engineering needs to be improved, rather than running algorithms to solve small-scale problems, says Thompson. In the United Kingdom, the firm has constructed prototype cabinet-sized devices, about 2 metres tall, which include cryogenic cooling equipment and many of the necessary computing components; larger ones will come online in the United States in 2025, says Shadbolt. Ultimately, PsiQuantum’s computer would involve in the order of 100 such devices, says O’Brien — around the size of a warehouse.
Raising cash
In response to questions about PsiQuantum’s progress and plans, spokesperson Alex Mack said that they have been thoroughly scrutinized by funders. And the company has a lot of success to show in that arena. In the past year, it has raised more than $1 billion in loans, equity and grants from governments in Australia and the United States, in exchange for building its first two quantum computers there: the first in Brisbane by 2027 and the next in Chicago, Illinois. In convincing funders, O’Brien’s combination of technical savvy and communication skills is an asset, says White. “You talk to his detractors, and you talk to his fans alike, and the thing they all land on is that he is unstoppable.” Working among investors from California’s Silicon Valley — where $1 billion isn’t seen as a huge amount — also helps, says Senellart.
Quantum computers: what are they good for?
However, PsiQuantum’s large public investments have raised eyebrows. In Australia, the National Audit Office is considering reviewing PsiQuantum’s US$620-million package from federal and state governments because of allegations that funders ran a system that lacked transparency and fair competition. Documents released under freedom of information requests made by Australian shadow science minister Paul Fletcher showed that PsiQuantum met at least twice with the Australian government to discuss an unsolicited bid. This occurred months before a handful of other firms, who were ultimately unsuccessful, were invited to pitch for funding.
This August, Fletcher told Sky News that the process was “essentially a backward engineered sham” to cover up that the government had already made the decision. (Australia’s federal government didn’t reply to Nature’s request for comment. Queensland’s state government — which changed political hands after an October election — has criticized the previous state administration for the deal, and says it is investigating.) Some researchers in Australia have also expressed concern about a country investing so much into a single company, although the government has funded other quantum-computing firms.
Mack confirms that PsiQuantum originally made an unsolicited pitch to the Australian government. “A lengthy diligence process followed,” he says, adding that Australia’s chief scientist led the process. He adds that much of the firm’s government funding is contingent on “successful execution” of milestones, such as its prototypes.
As one sign of support for PsiQuantum, it has withstood the scrutiny of around 50 specialists at the US Defense Advanced Research Projects Agency (DARPA). In 2023, DARPA selected PsiQuantum and Microsoft to advance in a programme investigating whether an underexplored approach to quantum computing can achieve “utility-scale operation”. The agency examined chips and analysed PsiQuantum’s plans, says Shadbolt. Under DARPA’s close inspection, “all the dirty laundry has to come out”, adds Simmons, who has experience of DARPA initiatives through her firm Photonic, based in Vancouver, Canada.
A DARPA document on its programme states that the projects are plausible but not guaranteed. Like anyone working on an evolving technology, PsiQuantum might find that some tasks are impossible or that a competitor can do something better or cheaper, says Finke. For White, even a failure would be a win, because, he says, the advances the firm is making will boost the wider photonics industry.
A final question is whether a completed machine will do what PsiQuantum promises. Lu worries that the firm, and some other quantum-computing start-ups, are making bold claims that are fanning inflated expectations. “There is a growing concern,” he says. Quantum computers offer potential advantages over classical machines, but “nobody really knows if quantum computers will help make better battery cells or design new pharmaceuticals”, adds Smith.
True to form, O’Brien remains confident. With enough logical error-corrected qubits carrying out enough operations, it will be possible, he says, to answer questions of profound value that would otherwise be unanswerable. “I think the utility of a quantum computer is unequivocal.”
[ad_2]
Source link -
$160m funding to boost US semiconductor manufacturing
[ad_1]
The United States Department of Defense has announced a new $160m investment to bolster US semiconductor manufacturing.
This funding, part of the bipartisan CHIPS and Science Act, supports the Microelectronics Commons, a network of eight technology hubs dedicated to advancing the nation’s semiconductor production capabilities.
This latest allocation follows earlier CHIPS Act investments, totalling nearly $269m in September and nearly $240m the previous year, to strengthen the United States’ global semiconductor leadership.
Strengthening US semiconductor manufacturing: The role of the hubs
Of the latest $160m, $148m will be directed toward Microelectronics Commons Hubs to support infrastructure, operations, and workforce development.
This investment, distributed across the eight established hubs, targets regional semiconductor innovation and workforce readiness.
Each hub collaborates with local universities, research foundations, and industry leaders to address key challenges in the US semiconductor manufacturing industry.
Here’s a closer look at the funding distribution:
- Northeast Microelectronics Coalition (NEMC), led by Massachusetts Technology Collaborative, received $18.7m.
- Silicon Crossroads Microelectronics Commons (SCMC), led by the Applied Research Institute in Indiana, was awarded $16.6m.
- California Defense Ready Electronics and Microdevices Superhub (CA DREAMS), led by USC, obtained $27m.
- Commercial Leap Ahead for Wide Bandgap Semiconductors (CLAWS) in North Carolina, led by North Carolina State University, received $23.7m.
- Southwest Advanced Prototyping (SWAP) in Arizona, led by Arizona State University, received $18.7m.
- Midwest Microelectronics Consortium (MEMC) in Ohio was granted $12.3m.
- Northeast Regional Defense Technology (NORDTECH) in New York received $10.6m.
- California-Pacific-Northwest AI Hardware Hub (NWAI), based at Stanford University, was allocated $15.3m.
An additional $10m will fund a Cross-Hub Enablement Solution (CHES) to facilitate shared access to Electronic Design Automation (EDA) tools and cloud computing resources across all hubs.
National collaboration in focus at the 2024 symposium
The Microelectronics Commons network convened its annual meeting in Washington, DC, from October 28–30, gathering over 2,000 participants to discuss advancements in semiconductor manufacturing and technology.
Deputy Secretary of Defense Kathleen Hicks emphasized the unity and strategic importance of semiconductor production, stating, “Chips bring America together.”
During the symposium, each hub presented progress updates, highlighting key projects, workforce training initiatives, and the importance of “lab-to-fab” pathways—efforts to expedite research to manufacturing.
This cross-industry collaboration is crucial to ensuring a robust semiconductor supply chain that supports national security.
Sustaining US leadership in semiconductor innovation
As the US semiconductor manufacturing sector continues to expand through the CHIPS and Science Act, industry stakeholders are optimistic about the potential to secure America’s position as a global leader in advanced microelectronics.
This ongoing investment in the Microelectronics Commons strengthens the path toward a resilient and self-sufficient semiconductor industry.
[ad_2]
Source link -
How AI is reshaping science and society
[ad_1]
ChatGPT and the Future of AI: The Deep Language Revolution Terrence J. Sejnowski The MIT Press (2024)
Some of the surprise winners of this year’s Nobel prizes were the developers of AlphaFold, an artificial intelligence (AI) model that can accurately predict the 3D structure of more than 200 million proteins. AlphaFold is powered by artificial neural networks that can glean patterns from how amino acids interact in known proteins and then use that information to model the behaviour of unknown proteins. Chatbots such as ChatGPT rely on similar technology to ‘learn’ and mimic human language.
The revolution set in motion by this technology is compellingly explored in Terrence Sejnowski’s ChatGPT and the Future of AI — a sequel to his 2018 book The Deep Learning Revolution. Sejnowski, a computational neurobiologist, takes readers on a whirlwind tour of the evolution of AI — from simple computational models of a single neuron built in the 1950s to breakthroughs in deep learning over the past two decades that have resulted in the birth of large language models (LLMs), which can generate human-like responses to questions.
ChatGPT one year on: who is using it, how and why?
Sejnowski’s book is a must-read for those seeking to understand the world we live in — a world in which machines transform the fabric of human cognition. Put simply, a neural network is a web of interconnected nodes, or neurons, that can process data and learn from them by adjusting the strength of the connections between the nodes. When the strength of these connections is altered to achieve a desired output during a training phase, the model learns to detect underlying patterns.
The fundamental inspiration behind neural-network architectures is the human brain. Sejnowski suggests that untangling the mystery of why these simulated models of the brain work so well — especially when they are trained on large amounts of data — could prove to be as seminal as the discovery of DNA. By rigorously interrogating the “otherworldly behaviour” of LLMs, new conceptual frameworks might emerge, he predicts.
He sees parallels between the current debates over the meanings of ‘intelligence’ and ‘understanding’ and discussions a century ago about the nature of life. Back then, vitalists argued that life is a non-physical force — an essence infused in living things but not in, say, rocks — whereas mechanists thought that life could be fully explained by physical and chemical processes. Just as the discovery of the DNA double helix bridged conceptual gaps and transformed biology, Sejnowski predicts that advances in AI will bring forth revelations about intelligence.
Evolving understanding
The holy grail of AI, Sejnowski explains, is artificial general intelligence: a machine that can think, learn and solve problems across a wide range of tasks, much like a human can. The current generation of LLMs is far from that. Referred to pejoratively by some researchers as ‘stochastic parrots’, they mostly mimic human language without true comprehension.
Ready or not, AI is coming to science education — and students have opinions
The road ahead for AI is one of interdisciplinary collaboration, Sejnowski argues, in which insights from biology, neuroscience and computer science converge to guide AI development. Sejnowski imagines that insights about the “fundamental principles of intelligence” — such as adaptability, flexibility and the ability to make general inferences from limited information — will catalyse the next generation of machine intelligence.
The AI language revolution, which is how Sejnowski refers to the era of LLMs, is already reshaping many aspects of human life. As LLMs evolve, they will surpass their primary role as tools and start acting as collaborators in domains such as health care, education and law. That’s already beginning to happen, as shown by AlphaFold. The author liberally uses ChatGPT to provide summaries at the end of each chapter, and conversations with the chatbot are littered throughout the book. He even playfully acknowledges ChatGPT as a co-author.
The power of LLMs also lies in how users interact with them. Sejnowski flags the increasingly important skill of prompt engineering, which stresses that subtle changes in how you instruct a model can shape its output. The author offers helpful hacks: ask for multiple responses, not just one; be specific, so that the model can converge on the best answer quickly; shape your dialogue as if you are talking with a real person.
Sejnowski proposes a “reverse Turing test”, in which the intelligence of the prompter is assessed on the basis of the quality of their interactions with the AI. Such proficiency tests might become common as AI use spreads.
Can AI have common sense? Finding out will be key to achieving machine intelligence
The next generation of LLMs must undergo a developmental process akin to the childhood learning phase in humans, he surmises, learning from real-world interactions as well as data. Long-term memory and goal-oriented behaviour must be integrated so that systems can adapt and plan effectively. The addition of robotics and sensorimotor systems would allow AI tools to perceive and interact with their environment, nudging current models towards artificial general autonomy.
Although Sejnowski is optimistic about the technology, he acknowledges that there are many unsolved challenges that will need to be addressed to ensure the long-term sustainability of AI. It is likely to disrupt industries and the job market. And, the increasing necessity for heavy computational processing power has brought to the forefront the need for more-efficient, greener AI models.
Although it is some way off, it is also important, Sejnowski writes, to seriously examine the possibility of AI exceeding human intelligence, because preparing now will allow us to anticipate and mitigate potential threats. A careful and well-regulated strategy to manage ethical and existential risks — such as of AI-induced errors and political weaponization of LLMs — is essential to ensure that AI benefits society.
ChatGPT and the Future of AI sets out how the next wave of AI-driven breakthroughs could alter the landscape of knowledge itself. Sejnowski’s exploration serves as both a guide and a warning as we navigate the promises and perils of AI’s rapid advancement.
Competing Interests
The author declares no competing interests.
[ad_2]
Source link -
EU invests €133m into Dutch photonic chips
[ad_1]
The European Union is set to invest €133m in production facilities for photonic chips in the Netherlands.
The PIXEurope consortium, which consists of parties from 11 countries, including the Netherlands, has been selected for contract negotiations to develop a European pilot plant for photonic chips.
The funding is part of a total of €380m and falls under the Chips Joint Undertaking.
It comes one month after Italy-based start-up Ephos raised $8.5m to create glass-based photonic chips at scale.
The importance of photonic chips in a low-carbon future
Photonic chips use light rather than electrons to perform calculations, with advantages in speed and power consumption.
This makes them ideal for use in areas such as data centres and motoring.
According to Dirk Beljaarts, the Dutch economy minister, “photonics is a technology of strategic importance” for the Netherlands.
He said: “We aim to gain a strong European competitive position in this area. From knowledge, innovation, supply to final production, this is necessary for the jobs and income of the future, for solving social challenges and our national security.”
Promoting research and development in the semiconductor industry
The investment forms part of a total amount of €380m to set up pilot photonic chip production plants throughout Europe, under the Chips Joint Undertaking, which centres around a European public-private partnership to promote research and development in the semiconductor industry.
Europe has been making a concerted effort over the past few years to be a leader in the semiconductor space.
In 2023, the region adopted the EU Chips Act, which aims to increase the EU’s share of global chip production from 10pc to at least 20pc by the end of the decade.
Since the Chips Act was enacted, the EU signed a deal with India to build robust supply chains and foster innovation together.
The €43bn act is designed to invest in industry players and research labs, including the Tyndall National Institute in Cork.
[ad_2]
Source link -
Chips Competence Centres to strengthen EU semiconductor sector
[ad_1]
Following a recent call, the European Commission announced the selection of 27 Chips Competence Centres across 24 Member States and Norway to enhance technical expertise and skills development in the semiconductor sector.
The first wave of Chips Competence Centres, co-financed by the EU and national contributions, represents an investment of over €170m.
A second call will be launched in 2025, allowing all 27 Member States, plus Iceland and Norway.
Each of the proposals will now proceed to individual contract negotiations.
Overview of EU chips initiatives
The European Chips Act is a key strategy in redefining global technology markets. It casts an ambitious vision to bolster the EU’s market share in chip production from a modest 10% to a substantial 20%.
It’s designed to boost large-scale semiconductor manufacturers and aid SMEs by effectively lowering entry costs into this highly competitive market.
The Act aims to build more semiconductor plants across Europe through attractive financial incentives.
Strengthening Europe’s semiconductor sector
The Chips Competence Centres are set to play a fundamental role in strengthening Europe’s semiconductor innovation ecosystem.
They will provide businesses—especially SMEs and start-ups—with essential resources to develop semiconductor solutions, including support, training, and access to large infrastructure facilities established under the Chips Act, such as pilot lines and the design platform.
Reflecting and reinforcing regional and national strengths, each centre will specialise in one or more key technology areas.
This targeted approach secures long-term investment in localised expertise, ensuring a sustained focus on innovation and growth.
Each Chips Competence Centre will meet specific regional needs
Complementing this effort, a European network connecting the Chips Competence Centres will be established.
This network will serve as a resource hub, allowing centres to support Europe’s semiconductor industry better while addressing the unique needs of each regional ecosystem.
In early 2025, a second call will be launched to extend coverage, inviting all 27 Member States, including Iceland and Norway, to establish their Competence Centres.
For additional context, further details on the Chips Joint Undertaking (Chips JU) and the Chips for Europe Initiative under the Chips Act are available.
[ad_2]
Source link -
What Trump’s election victory could mean for AI, climate and more
[ad_1]
From repealing climate policies to overturning guidance on the safe development of artificial intelligence (AI), Republican Donald Trump made plenty of promises during his presidential campaign that could affect scientists and science policy. But fulfilling all of his pledges won’t be easy.
The US is the world’s science superpower — but for how long?
Trump, now the US president-elect for a second time, will have some advantages as he re-enters the White House in January. The first time he took office in 2017, his victory was a surprise, and many government watchers who spoke to Nature say that he didn’t have a solid plan. By contrast, the Trump administration that enters office next year will be better prepared, and Trump himself is likely to face fewer checks on his power now that he has consolidated control over the Republican establishment, says Matt Dallek, a political historian at George Washington University in Washington DC who studies the modern conservative movement.
But that still does not mean he will be able to do as he pleases, Dallek adds. “There’s a kind of revolutionary sweep to a lot of Trump’s promises that may collide with the messy reality of implementation.”
Here Nature talks to policy and other specialists about what might be in store on a range of science issues during a second Trump administration.
Artificial intelligence
Trump, who is industry-friendly, has promised to repeal US President Joe Biden’s executive order on AI, a guideline released last year for developing the technology safely and responsibly. Trump’s pledge echoes the Republican party’s platform, which says that the executive order “hinders AI Innovation”.
It will be possible for Trump to implement his plan as soon as he enters the White House given that executive orders can be revoked by a president at any time. But what will Trump put in its place?
‘We need to be ready for a new world’: scientists globally react to Trump election win
“The emphasis will shift away from the regulatory environment” and towards technology companies making their own voluntary decisions on safety, says Suresh Venkatasubramanian, director of the Center for Technological Responsibility, Reimagination, and Redesign at Brown University in Providence, Rhode Island. “I am personally sceptical that that will be enough” to address AI-associated risks to public safety, data-privacy concerns or the use of biased algorithms that disadvantage certain groups of people, Venkatasubramanian says.
Biden’s executive order emphasized making sure that AI models, which are trained on human-derived data, don’t output discriminatory results. That’s probably also not going to be a heavy priority for the new administration, Venkatasubramanian says. The Republican platform says that it will “support AI Development rooted in Free Speech and Human Flourishing”.
According to Roman Yampolskiy, a computer scientist and AI safety researcher at the University of Louisville in Kentucky, “it is a great idea to remove censorship and support free speech” in general. But, he says, “removing regulations around training of advanced AI systems is the worst possible thing we can do for the safety of the American people and the world”. Given the risks associated with developing superintelligent AI systems, which could potentially operate in unpredictable ways and cause harm to humans, Yampolskiy and other AI researchers have been arguing for a pause in AI development, which could only be achieved with stronger regulations.
Climate change
Many federal climate efforts are likely to stall or move in reverse under Trump, who has long denied the dangers of climate change while prioritizing the economic benefits of boosting domestic fossil-fuel production. Even so, policy specialists say that Trump is unlikely to stop the United States’s gradual shift towards clean energy.
The US election is monumental for science, say Nature readers — here’s why
For instance, it will not be easy to undo Biden’s signature climate achievement: the 2022 Inflation Reduction Act (IRA), which created a raft of federal investments now estimated at more than US$1 trillion in climate and clean energy that are scheduled to run until around 2032. Repealing that legislation would require an act of the US Congress. But even if Republicans end up in control of both congressional chambers, businesses and leaders in conservative US states that are already benefitting from IRA investments might not be eager to cut off the flow of federal money, says Joanna Lewis, who heads the science technology and international affairs programme at Georgetown University in Washington DC.
Trump could have a bigger — and more negative — impact on climate progress if he moves to weaken climate regulations put in place for things such as power plants and automobiles. Similarly, his promise to place new tariffs on goods from countries such as China and Mexico could actually increase the cost of clean-energy technologies, says David Victor, a political scientist at the University of California, San Diego.
The president-elect has also promised to once again pull the United States out of the 2015 Paris agreement, which commits member countries to limiting global warming to 1.5–2 °C above preindustrial levels. Trump’s administration had to wait until 2020 before formally leaving the agreement last time, and Biden moved to rejoin the agreement quickly after taking office several months later. But under the rules of the agreement, the leaving process would take only one year this time around.
Many climate observers say the absence of the United States — the world’s second-largest greenhouse-gas emitter — from the pact could reduce pressure on China and other nations to scale up their efforts to curb emissions just as time is running short. “This is a pivotal decade for climate action, and four more years of Trump could be disastrous in terms of mobilizing climate action,” Lewis says.
Health
In the weeks leading up to the US election, Trump teamed up with political figure Robert F. Kennedy Jr. on a platform promising to “make America healthy again” by tackling the root causes of chronic diseases, removing toxic substances from the environment and combatting corporate corruption. Trump has said that he will let Kennedy, who has questioned vaccine effectiveness, “go wild on” health, unnerving public-health and health-policy researchers.
It remains to be seen whether Trump will appoint Kennedy to a position such as director of US Health and Human Services (HHS) — or whether the US Senate would approve it — but it’s clear that Kennedy will have Trump’s ear on health issues.
Robert F. Kennedy Jr. ran for president in 2024 as an independent before withdrawing his bid and endorsing Trump.Credit: Rebecca Noble/Getty
Georges Benjamin, the executive director of the American Public Health Association in Washington DC, worries about Kennedy’s role in the new administration because he has long cast doubt on the vaccine-approval process, threatening to undermine confidence in jabs and cause a resurgence in illnesses such as measles. “People will get sick and die because of the confusion around vaccines, if [Kennedy and Trump] implement some of the things they verbalize,” he says.
Some of Kennedy’s goals, such as cracking down on ties to industry at regulatory agencies such as the US Food and Drug Administration, are noble, says Diana Zuckerman, president of the National Center for Health Research, a non-profit think tank in Washington DC. But those goals don’t jibe with what occurred during the first Trump administration, when Trump installed people with close industry ties to important health posts, such as former HHS director Alex Azar, so it’s hard to know what will happen, she says.
With Trump’s isolationalist approach and past comments he has made criticizing the World Health Organization, support for global health will also likely be “greatly scaled back” during Trump’s second term, says Ezekiel Emanuel, a bioethicist and long-time observer of the US biomedical funding landscape at the University of Pennsylvania in Philadelphia. The United States is “the key player” in funding of global-health initiatives, says Emanuel. This includes, for instance, a programme that aims to end the global AIDS epidemic. So it’s “hard to be optimistic” about the future, he adds.
Foreign science partnerships
During Trump’s first term, his administration barred people from a half-dozen countries that it said were “compromised by terrorism” from entering the United States and implemented an anti-espionage programme called the China Initiative that led to the arrests of scientists of Chinese heritage. The Biden administration overturned the travel ban and ended the China Initiative, but under Biden, federal officials have continued efforts to guard against foreign interference in US research.
US election debate: what Harris and Trump said about science
Specialists says it’s unclear whether the second Trump administration will revive the China Initiative, although the Republican-led US House of Representatives advanced legislation in September that would do so. But a reinstatement of the travel ban is likely, says Adam Cohen, a lawyer at Siskind Susser in Memphis, Tennessee, who focuses on academic immigration and who says the president has broad authority to institute such policies.
Like the first Trump administration, the new one will probably clamp down on granting visas to foreign researchers and students from some countries, says Jennifer Steele, an education policy researcher at American University in Washington DC. Policies that make it harder for international and US researchers to meet would also make it harder for new scientific collaborations to arise, says Caroline Wagner, a specialist in science, technology and international affairs at The Ohio State University in Columbus. That’s because such partnerships are fuelled by face-to-face contact. “Collaborations don’t begin with people just e-mailing each other across the miles,” she says.
But there might be one bright spot on the collaboration front, at least for US-China partnerships. Denis Simon, a non-resident fellow at the Quincy Institute for Responsible Statecraft, a foreign policy think tank in Washington DC, thinks that a crucial pact governing US-China scientific cooperation that has been expired for the past year is likely to be signed by the Biden administration before Trump’s second inauguration in January. A renewal of the agreement, although it is will probably be more limited in scope owing to increased US-China tensions, would show that “both governments give their blessing” to collaborations, Simon says.
[ad_2]
Source link -
ChatGPT is transforming peer review — how can we use it responsibly?
[ad_1]
Since the artificial intelligence (AI) chatbot ChatGPT was released in late 2022, computer scientists have noticed a troubling trend: chatbots are increasingly used to peer review research papers that end up in the proceedings of major conferences.
There are several telltale signs. Reviews penned by AI tools stand out because of their formal tone and verbosity — traits commonly associated with the writing style of large language models (LLMs). For example, words such as commendable and meticulous are now ten times more common in peer reviews than they were before 2022. AI-generated reviews also tend to be superficial and generalized, often don’t mention specific sections of the submitted paper and lack references.
Ready or not, AI is coming to science education — and students have opinions
That’s what my colleagues and I at Stanford University in California found when we examined some 50,000 peer reviews for computer-science articles published in conference proceedings in 2023 and 2024. We estimate that 7–17% of the sentences in the reviews were written by LLMs on the basis of the writing style and the frequency at which certain words occur (W. Liang et al. Proc. 41st Int. Conf. Mach. Learn. 235, 29575–29620; 2024).
Lack of time might be one reason for using LLMs to write peer reviews. We found that the rate of LLM-generated text is higher in reviews that were submitted close to the deadline. This trend will only intensify. Already, editors struggle to secure timely reviews and reviewers are overwhelmed with requests.
Fortunately, AI systems can help to solve the problem that they have created. For that, LLM use must be restricted to specific tasks — to correct language and grammar, answer simple manuscript-related questions and identify relevant information, for instance. However, if used irresponsibly, LLMs risk undermining the integrity of the scientific process. It is therefore crucial and urgent that the scientific community establishes norms about how to use these models responsibly in the academic peer-review process.
Intellectual property and data privacy: the hidden risks of AI
First, it is essential to recognize that the current generation of LLMs cannot replace expert human reviewers. Despite their capabilities, LLMs cannot exhibit in-depth scientific reasoning. They also sometimes generate nonsensical responses, known as hallucinations. A common complaint from researchers who were given LLM-written reviews of their manuscripts was that the feedback lacked technical depth, particularly in terms of methodological critique (W. Liang et al. NEJM AI 1, AIoa2400196; 2024). LLMs can also easily overlook mistakes in a research paper.
Given those caveats, thoughtful design and guard rails are required when deploying LLMs. For reviewers, an AI chatbot assistant could provide feedback on how to make vague suggestions more actionable for authors before the peer review is submitted. It could also highlight sections of the paper, potentially missed by the reviewer, that already address questions raised in the review.
To assist editors, LLMs can retrieve and summarize related papers to help them contextualize the work and verify adherence to submission checklists (for instance, to ensure that statistics are properly reported). These are relatively low-risk LLM applications that could save reviewers and editors time if implemented well.
LLMs might, however, make mistakes even when performing low-risk information-retrieval and summarization tasks. Therefore, LLM outputs should be viewed as a starting point, not as the final answer. Users should still cross-check the LLM’s work.
ChatGPT one year on: who is using it, how and why?
Journals and conferences might be tempted to use AI algorithms to detect LLM use in peer reviews and papers, but their efficacy is limited. Although such detectors can highlight obvious instances of AI-generated text, they are prone to producing false positives — for example, by flagging text written by scientists whose first language is not English as AI-generated. Users can also avoid detection by strategically prompting the LLM. Detectors often struggle to distinguish reasonable uses of an LLM — to polish raw text, for instance — from inappropriate ones, such as using a chatbot to write the entire report.
Ultimately, the best way to prevent AI from dominating peer review might be to foster more human interactions during the process. Platforms such as OpenReview encourage reviewers and authors to have anonymized interactions, resolving questions through several rounds of discussion. OpenReview is now being used by several major computer-science conferences and journals.
The tidal wave of LLM use in academic writing and peer review cannot be stopped. To navigate this transformation, journals and conference venues should establish clear guidelines and put in place systems to enforce them. At the very least, journals should ask reviewers to transparently disclose whether and how they use LLMs during the review process. We also need innovative, interactive peer-review platforms adapted to the age of AI that can automatically constrain the use of LLMs to a limited set of tasks. In parallel, we need much more research on how AI can responsibly assist with certain peer-review tasks. Establishing community norms and resources will help to ensure that LLMs benefit reviewers, editors and authors without compromising the integrity of the scientific process.
Competing Interests
The author declares no competing interests.
[ad_2]
Source link -
Scalable watermarking for identifying large language model outputs
[ad_1]
Detailed SynthID-Text method
In this section, we provide a detailed description of SynthID-Text.
The LLM distribution
Most LLMs are autoregressive, providing the probability pLM(xt∣x<t) of the next token xt given the text so far x<t. Text is typically generated from the LLM using an autoregressive decoding method, which optionally modifies the LLM distribution pLM(⋅∣x<t) before sampling from it. Such modifications include top-k and top-p34 sampling, which truncate pLM(⋅∣x<t) to the k most likely tokens or the tokens covering the top-p probability mass; this can be combined with applying a temperature parameter τ (ref. 35). Although these modifications increase or decrease the amount of entropy in pLM(⋅∣x<t), SynthID-Text is compatible with any autoregressive decoding method that has non-zero entropy in the modified distribution. Thus, SynthID-Text is compatible with top-k sampling for all k ≥ 2, top-p sampling for all \(p\in \left(0,1\right]\), and all temperatures τ > 0.
SynthID-Text is applied after any such modifications have been made, so for the purposes of this paper we define the LLM distribution pLM(⋅∣x<t) to be the distribution after any such modifications.
Definition 1 (LLM distribution)
Given an autoregressive LLM, an autoregressive decoding method, and x<t = x1, …, xt−1, a sequence of tokens from the vocabulary V, the LLM distribution pLM(⋅∣x<t) is the probability distribution from which the decoding method samples the next token xt ∈ V.
Watermarking framework
We present SynthID-Text as comprising a random seed generator, a sampling algorithm and a scoring function; this is similar to the generative watermarking framework of ref. 21. Intuitively, the sampling algorithm samples text from the LLM in a way that is biased by random seeds provided on each step by the random seed generator; later we can identify the watermark by detecting this bias through the scoring function. We describe the random seed generator and sampling algorithm in this section and describe several scoring functions in Supplementary Information section A. See Supplementary Information section B for a detailed discussion of related generative watermarking approaches.
Random seed generator
To generate a piece of watermarked text x1, …, xT, we require a sequence of random seeds \({r}_{1},\ldots ,{r}_{T}\in {\mathcal{R}}\) (where \({\mathcal{R}}\) is the space of all random seeds) to bias the sampling from the LLM distribution on each step. The random seed generator is the process by which we generate these random seeds. One approach is to make the random seed generator a deterministic function fr that takes as input the sequence of tokens so far x<t = x1, …, xt−1 and a watermarking key k and outputs a random seed \({r}_{t}:={f}_{r}({x}_{ < t},k)\in {\mathcal{R}}\). Randomizing the key k should randomize the seed; that is, for all \({x}_{ < t},{{\mathbb{P}}}_{{k \sim }\text{Unif}({\mathcal{R}})}\,[\,{f}_{r}({x}_{ < t},k)]=\text{Unif}\,({\mathcal{R}})\).
There are several possible choices for fr (ref. 21); for our experiments, we use the sliding window fr(x<t, k) ≔ h(xt−H, …, xt−1, k), which is a hash function h of the last H tokens (for some context length H ≥ 1) and of the key k. This random seed generator is the same as that used by refs. 22,23. In this work, we also assume the watermarking key k and random seed rt exist in the same space of nsec-bit integers, where nsec is the security parameter.
Definition 2 (random seed space, random seed distribution)
Given a security parameter nsec, the random seed space \({\mathcal{R}}={\{0,1\}}^{{n}_{\text{sec}}}\) is the space of all nsec-bit integers. The random seed distribution is the uniform distribution over all such integers \(\,\text{Unif}\,({\mathcal{R}})\).
We also assume that the family of functions \({\{h(\cdot ,\ldots ,\cdot ,k)\}}_{k\in {\mathcal{R}}}\) is a pseudorandom function family, meaning that (1) h(xt−H, …, xt−1, k) is efficiently computable for any xt−H, …, xt−1 and k, and (2) the distribution of \({\{h(\cdot ,\ldots ,\cdot ,k)\}}_{{k \sim }\text{Unif}({\mathcal{R}})}\) is computationally indistinguishable from a function sampled uniformly randomly from the set of all functions from VH to \({\{0,1\}}^{{n}_{\text{sec}}}\).
g-values
As illustrated in Fig. 2, Tournament sampling requires g-values to decide which tokens win each match in the tournament. Intuitively, we want a function that takes a token x ∈ V, a random seed \(r\in {\mathcal{R}}\) and the layer number ℓ ∈ {1, …, m}, and outputs a g-value gℓ(x, r) that is a pseudorandom sample from some probability distribution fg (the g-value distribution).
For example, in Fig. 2, the g-value distribution is Bernoulli(0.5). Given the random seed r, gℓ(x, r) produces pseudorandom g-values of 0 or 1 for each token x in the vocabulary, for each layer ℓ = 1, 2, 3. In this paper, we primarily use the Bernoulli(0.5) g-value distribution, although we also explore Uniform[0, 1]. In general, any g-value distribution can be chosen, as a hyperparameter of the Tournament sampling method.
Definition 3 (g-value distribution)
The g-value distribution is a probability distribution of any real-valued random variable. We write Fg to denote the cumulative distribution function, and fg to denote the probability density function (if continuous) or probability mass function (if discrete).
Next, we need a way to produce a hash \(h(x,{\ell },r)\in {\mathcal{R}}\) of a token x ∈ V, an integer ℓ ∈ {1, …, m} and a random seed \(r\in {\mathcal{R}}\). Let’s assume we have a pseudorandom function family \({\{h(\cdot ,\cdot ,r)\}}_{r\in {\mathcal{R}}}\) similar to the one described in the ‘Random seed generator’ section, such that the distribution of \({\{h(\cdot ,\cdot ,r)\}}_{{r \sim }{\rm{Unif}}({\mathcal{R}})}\) is computationally indistinguishable from a function sampled uniformly randomly from the set of all functions from V × [m] to \({\{0,1\}}^{{n}_{\sec }}\).
Definition 4 (g-value)
Given a g–value distribution with cumulative density function. Fg, a random seed \(r\in {\mathcal{R}}\), and integer ℓ ∈ 1, …, m, the layer-ℓ g-value of a token x ∈ V is given by:
$${g}_{{\ell }}(x,r)\,:={F}_{g}^{-1}\,\left(\frac{h(x,{\ell },r)}{{2}^{{n}_{\text{sec}}}}\right),$$
where \({F}_{g}^{-1}\) is the generalized inverse distribution function of Fg, and h is a hash function as described above.
Intuitively, Definition 4 says that we take a hash h(x, ℓ, r) of x, ℓ and r, which gives us a uniformly distributed n-bit integer, and divide it by 2n to get a number in [0, 1]. For large n, this converges to a uniformly distributed number in [0, 1]. We then perform inverse transform sampling to turn this number into a sample from the g-value distribution given by Fg.
Tournament sampling algorithm
Definition 5 (watermarking sampling algorithm)
In a watermarking scheme, a sampling algorithm \({\mathcal{S}}:\Delta V\times {\mathcal{R}}\to V\) is an algorithm that takes as input a probability distribution p ∈ ΔV and a random seed \(r\in {\mathcal{R}}\) and returns a token \({\mathcal{S}}(p,r)\in V\). If \({\mathcal{S}}\) always returns the same token given the same p and r, it is deterministic. Otherwise, \({\mathcal{S}}\) is probabilistic.
We propose a new probabilistic sampling algorithm called Tournament sampling. We present the simplest, single-layer version of Tournament sampling in Algorithm 1. Instead of sampling directly from pLM(⋅∣x<t), we sample N tokens from pLM(⋅∣x<t), compute their g-values as described in the previous section and choose uniformly among those that have the maximal g-value.
Algorithm 2 presents the full multilayer version of Tournament sampling, which has an additional hyperparameter m, the number of layers. The process can be thought of as a knockout tournament with m stages, where each match is an instantiation of the single-layer algorithm; this continues until there is one winner. Importantly, each layer ℓ of the tournament uses different g-values gℓ(⋅, rt) to decide the winners. Figure 2 gives a concrete example for m = 3 layers, N = 2 samples and a Bernoulli(0.5) g-value distribution.
Algorithm 1
Sampling a token with single-layer Tournament sampling
Require: LLM distribution pLM(⋅∣x<t), random seed \({r}_{t}\in {\mathcal{R}}\), number of samples N ≥ 2, g function with g-value distribution fg (see Definition 4).
1: Draw Y = [y1, y2, …, yN] containing N independent samples from pLM(⋅∣x<t) (may contain repeats).
2: \({Y}^{* }:=\,[\,y\in Y:{g}_{1}(\,y,{r}_{t})=\mathop{\max }\limits_{{y}^{{\prime} }\in Y}{g}_{1}(\,{y}^{{\prime} },{r}_{t})]\) (may contain repeats).
3: Sample xt ~ Unif(Y*)
4: return xt
Algorithm 2
Sampling a token with multilayer Tournament sampling.
Require: LLM distribution pLM(⋅∣x<t), random seed \({r}_{t}\in {\mathcal{R}}\), number of samples N ≥ 2, g function with g-value distribution fg (see Definition 4), number of layers m ≥ 1.
1: Draw Nm independent samples \({y}_{0}^{0},{y}_{1}^{0},\ldots ,{y}_{{N}^{m}-1}^{0} \sim {p}_{{\rm{LM}}}(\cdot | {x}_{ < t})\) (may contain repeats).
2: for 1 ≤ ℓ ≤ m do
3: for 0 ≤ j ≤ Nm−ℓ − 1 do
4: \(Y:=\,[\,{y}_{Nj}^{{\ell }-1},{y}_{Nj+1}^{{\ell }-1},\ldots ,{y}_{Nj+N-1}^{{\ell }-1}]\) (may contain repeats).
5: \({Y}^{* }:=\,[\,y\in Y:{g}_{{\ell }}(\,y,{r}_{t})=\mathop{\max }\limits_{{y}^{{\prime} }\in Y}{g}_{{\ell }}(\,{y}^{{\prime} },{r}_{t})]\) (may contain repeats).
6: Sample \({y}_{j}^{{\ell }} \sim \,{\rm{Unif}}\,({Y}^{* })\).
7: end for
8: end for
9: return \({x}_{t}:={y}_{0}^{m}\)
Repeated context masking
To generate a full response, we could simply apply Algorithm 2 on every decoding step, using the sliding-window random seed generator (‘Random seed generator’ section) to generate the random seed rt for each step. However, it is possible that the same window of context, and thus the same random seed might occur more than once (particularly if the sliding-window size H is small or the response is long). It has been shown that in this scenario, the watermark can introduce a repeated bias that affects the quality of the text, for example, causing repeating loops24,25. One way to avoid this problem is to apply repeated context masking27, which prevents the watermark from being applied on step t if the context window (xt−H, …, xt−1) has been used to watermark previously.
We present the method in Algorithm 3, which we call K-sequence repeated context masking. The integer parameter K ≥ 1 controls for how long context windows are held in the history. In the simplest case of K = 1, we only hold the context history for the duration of generating a single response. For larger integers K > 1, we check against a history of contexts used in the last K responses. In the extreme case, we could set K = ∞ and retain the context history indefinitely. In Supplementary Information section G.2, we show that applying K-sequence repeated context masking achieves K-sequence non-distortion, an important property for quality preservation. In Supplementary Information section G.3, we discuss the trade-offs of smaller and larger K. For most of our experiments we use K = 1.
Algorithm 3
Generating watermarked responses with sliding-window random seed generation and K-sequence repeated context masking.
Require: LLM pLM(⋅∣⋅), context window size H, pseudorandom hash function h, watermarking key \(k\in {\mathcal{R}}\), sampling algorithm \({\mathcal{S}}:\Delta V\times {\mathcal{R}}\to V\), integer K ≥ 1, stream of prompts (x1, x2, …).
1: for i ≥ 1 do
2: \({C}_{i}:=\varnothing \)
3: t ≔ n where n is the length of \({{\bf{x}}}^{i}={{\bf{x}}}_{1}^{i},\ldots ,{{\bf{x}}}_{n}^{i}\)
4: while \({{\bf{x}}}_{t}^{i}\ne {\mathtt{EOS}}\) do
5: t ≔ t + 1
6: if \(({{\bf{x}}}_{t-H}^{i},\ldots ,{{\bf{x}}}_{t-1}^{i})\in {C}_{i}\cup {C}_{i-1}\cup \cdots \cup {C}_{i-K+1}\) then
7: Sample \({{\bf{x}}}_{t}^{i} \sim {p}_{{\rm{LM}}}(\cdot | {{\bf{x}}}_{ < t}^{i})\)
8: else
9: \({r}_{t}:=h({{\bf{x}}}_{t-H}^{i},\ldots ,{{\bf{x}}}_{t-1}^{i},k)\)
10: Sample \({{\bf{x}}}_{t}^{i}:={\mathcal{S}}({p}_{{\rm{LM}}}(\cdot | {{\bf{x}}}_{ < t}^{i}),{r}_{t})\)
11: \({C}_{i}:={C}_{i}\cup \{({{\bf{x}}}_{t-H}^{i},\ldots ,{{\bf{x}}}_{t-1}^{i})\}\)
12: end if
13: end while
14: return Response \({{\bf{y}}}^{i}:={{\bf{x}}}_{n+1:t}^{i}\)
15: end for
Scoring functions
A scoring function takes a piece of text x1, …, xT along with the random seeds r1, …, rT and computes a score, which can then be compared with a threshold to classify the text as watermarked or unwatermarked. Here the random seeds rt = fr(x<t, k) are from the random seed generator (‘Random seed generator’ section). It is noted that a scoring function only requires access to the tokenized text, the watermarking key k and the random seed generator fr; no access to the LLM is required.
For SynthID-Text, we propose several scoring functions, which are in Supplementary Information section A. All the scores are computed from the g-values of the text. The simplest of these is the mean score, which is simply the mean of the g-values across all timesteps and layers. We also propose a weighted mean score, which re-weights the evidence of each tournament layer. We propose frequentist versions of these scores, which perform a hypothesis test on these means to produce a P value. Lastly, we propose a parameterized Bayesian scoring function, which achieves better performance by learning from data (watermarked and unwatermarked texts) to compute the posterior probability that a text is watermarked.
Experimental details
LLMs and LLM configurations
In our experiments, we use the IT variants of the Gemma 2B and 7B models28. We also use the v0.2 Mistral 7B-IT model29. To generate text, we use top-k sampling36. Following default settings, we use k = 100 for the IT models. We experiment with temperatures of 0.5, 0.7 and 1.0, as varying the temperature changes the entropy of the model, which affects watermark detectability.
Data
To prompt our models we use the ELI530 dataset, which consists of English questions that require explanatory multi-sentence answers. This simulates a more task-oriented setting. For experiments with non-distortionary watermarking, our ELI5 test set and the development set each contain sets of 10,000 disjoint prompts that are used to prompt the model to obtain watermarked responses. For experiments with distortionary watermarking, we use 1,500 prompts from ELI5 for the test set to prompt the watermarked model. For the unwatermarked samples used as negatives, we use two disjoint sets of human-written responses to 10,000 questions from the ELI5 for the development and test sets.
Text lengths
For some experiments, we evaluate texts of fixed length—for example, 200 tokens. To obtain text of length exactly 200 tokens, we select the subset of texts that are longer than 200 tokens and then truncate them to have exactly 200 tokens.
Detectability metric
To report detectability, we use the true-positive rate (TPR) for a fixed false-positive rate (FPR) of x%, measured empirically. We denote this metric as TPR @ FPR = x%. For example to compute TPR @ FPR = 1%, we take the scores (under some scoring function) of the unwatermarked texts and compute a threshold corresponding to the top-1% highest scores. Then we compute the true-positive rate by measuring the fraction of watermarked texts that score above this threshold. Although some scoring functions allow a precise theoretical guarantee on the false-positive rate—for example, the frequentist scoring functions (Supplementary Information section A.3) which provide a P value—in this work we take the empirical approach described above.
Random seed generator settings
For all watermarking experiments (including Tournament, Gumbel and Soft Red List sampling algorithms), we use the same sliding-window-based random seed generator described in the ‘Random seed generator’ section, with context window size H = 4. We apply one-sequence repeated context masking (‘Repeated context masking’ section).
SynthID-Text settings
Unless otherwise mentioned, for all SynthID-Text experiments, we use m = 30 tournament layers, a Bernoulli(0.5) g-value distribution fg (Algorithm 2) and the Bayesian scoring function (Supplementary Information section A.4).
[ad_2]
Source link
