Computer science

Quantum hardware may be a good match for AI

AI, Computer science, image classification, Physics, quantum computing, quantum mechanics, Science / Kelly Newman / April 11, 2025

Quantum computers don’t have that sort of separation. While they could include some quantum memory, the data is generally housed directly in the qubits, while computation involves performing operations, called gates, directly on the qubits themselves. In fact, there has been a demonstration that, for supervised machine learning, where a system can learn to classify items after training on pre-classified data, a quantum system can outperform classical ones, even when the data being processed is housed on classical hardware.

This form of machine learning relies on what are called variational quantum circuits. This is a two-qubit gate operation that takes an additional factor that can be held on the classical side of the hardware and imparted to the qubits via the control signals that trigger the gate operation. You can think of this as analogous to the communications involved in a neural network, with the two-qubit gate operation equivalent to the passing of information between two artificial neurons and the factor analogous to the weight given to the signal.

That’s exactly the system that a team from the Honda Research Institute worked on in collaboration with a quantum software company called Blue Qubit.

Pixels to qubits

The focus of the new work was mostly on how to get data from the classical world into the quantum system for characterization. But the researchers ended up testing the results on two different quantum processors.

The problem they were testing is one of image classification. The raw material was from the Honda Scenes dataset, which has images taken from roughly 80 hours of driving in Northern California; the images are tagged with information about what’s in the scene. And the question the researchers wanted the machine learning to handle was a simple one: Is it snowing in the scene?

Quantum hardware may be a good match for AI Read More »

Microsoft demonstrates working qubits based on exotic physics

Apple, Computer science, Majorana, Physics, quantum computing, quantum mechanics, quasiparticles, Science / Mike M. / February 20, 2025

Microsoft’s first entry into quantum hardware comes in the form of Majorana 1, a processor with eight of these qubits.

Given that some of its competitors have hardware that supports over 1,000 qubits, why does the company feel it can still be competitive? Nayak described three key features of the hardware that he feels will eventually give Microsoft an advantage.

The first has to do with the fundamental physics that governs the energy needed to break apart one of the Cooper pairs in the topological superconductor, which could destroy the information held in the qubit. There are a number of ways to potentially increase this energy, from lowering the temperature to making the indium arsenide wire longer. As things currently stand, Nayak said that small changes in any of these can lead to a large boost in the energy gap, making it relatively easy to boost the system’s stability.

Another key feature, he argued, is that the hardware is relatively small. He estimated that it should be possible to place a million qubits on a single chip. “Even if you put in margin for control structures and wiring and fan out, it’s still a few centimeters by a few centimeters,” Nayak said. “That was one of the guiding principles of our qubits.” So unlike some other technologies, the topological qubits won’t require anyone to figure out how to link separate processors into a single quantum system.

Finally, all the measurements that control the system run through the quantum dot, and controlling that is relatively simple. “Our qubits are voltage-controlled,” Nayak told Ars. “What we’re doing is just turning on and off coupling of quantum dots to qubits to topological nano wires. That’s a digital signal that we’re sending, and we can generate those digital signals with a cryogenic controller. So we actually put classical control down in the cold.”

Microsoft demonstrates working qubits based on exotic physics Read More »

AI used to design a multi-step enzyme that can digest some plastics

AI, biochemistry, Biology, catalysts, chemistry, Computer science, enzymes, Science / Mike M. / February 16, 2025

And it worked. Repeating the same process with an added PLACER screening step boosted the number of enzymes with catalytic activity by over three-fold.

Unfortunately, all of these enzymes stalled after a single reaction. It turns out they were much better at cleaving the ester, but they left one part of it chemically bonded to the enzyme. In other words, the enzymes acted like part of the reaction, not a catalyst. So the researchers started using PLACER to screen for structures that could adopt a key intermediate state of the reaction. This produced a much higher rate of reactive enzymes (18 percent of them cleaved the ester bond), and two—named “super” and “win”—could actually cycle through multiple rounds of reactions. The team had finally made an enzyme.

By adding additional rounds alternating between structure suggestions using RFDiffusion and screening using PLACER, the team saw the frequency of functional enzymes increase and eventually designed one that had an activity similar to some produced by actual living things. They also showed they could use the same process to design an esterase capable of digesting the bonds in PET, a common plastic.

If that sounds like a lot of work, it clearly was—designing enzymes, especially ones where we know of similar enzymes in living things, will remain a serious challenge. But at least much of it can be done on computers rather than requiring someone to order up the DNA that encodes the enzyme, getting bacteria to make it, and screening for activity. And despite the process involving references to known enzymes, the designed ones didn’t share a lot of sequences in common with them. That suggests there should be added flexibility if we want to design one that will react with esters that living things have never come across.

I’m curious about what might happen if we design an enzyme that is essential for survival, put it in bacteria, and then allow it to evolve for a while. I suspect life could find ways of improving on even our best designs.

Science, 2024. DOI: 10.1126/science.adu2454 (About DOIs).

AI used to design a multi-step enzyme that can digest some plastics Read More »

Quantum teleportation used to distribute a calculation

Computer science, Physics, quantum computing, quantum mechanics, Science, teleportation / 9u50fv / February 6, 2025

The researchers showed that this setup allowed them to teleport with a specific gate operation (controlled-Z), which can serve as the basis for any other two-qubit gate operation—any operation you might want to do can be done by using a specific combination of these gates. After performing multiple rounds of these gates, the team found that the typical fidelity was in the area of 70 percent. But they also found that errors typically had nothing to do with the teleportation process and were the product of local operations at one of the two ends of the network. They suspect that using commercial hardware, which has far lower error rates, would improve things dramatically.

Finally, they performed a version of Grover’s algorithm, which can, with a single query, identify a single item from an arbitrarily large unordered list. The “arbitrary” aspect is set by the number of available qubits; in this case, having only two qubits, the list maxed out at four items. Still, it worked, again with a fidelity of about 70 percent.

While the work was done with trapped ions, almost every type of qubit in development can be controlled with photons, so the general approach is hardware-agnostic. And, given the sophistication of our optical hardware, it should be possible to link multiple chips at various distances, all using hardware that doesn’t require the best vacuum or the lowest temperatures we can generate.

That said, the error rate of the teleportation steps may still be a problem, even if it was lower than the basic hardware rate in these experiments. The fidelity there was 97 percent, which is lower than the hardware error rates of most qubits and high enough that we couldn’t execute too many of these before the probability of errors gets unacceptably high.

Still, our current hardware error rates started out far worse than they are today; successive rounds of improvements between generations of hardware have been the rule. Given that this is the first demonstration of teleported gates, we may have to wait before we can see if the error rates there follow a similar path downward.

Nature, 2025. DOI: 10.1038/s41586-024-08404-x (About DOIs).

Quantum teleportation used to distribute a calculation Read More »

To help AIs understand the world, researchers put them in a robot

AI, Computer science, language, robotics, Science / Shannon Garcia / February 2, 2025

There’s a difference between knowing a word and knowing a concept.

Large language models like ChatGPT display conversational skills, but the problem is they don’t really understand the words they use. They are primarily systems that interact with data obtained from the real world but not the real world itself. Humans, on the other hand, associate language with experiences. We know what the word “hot” means because we’ve been burned at some point in our lives.

Is it possible to get an AI to achieve a human-like understanding of language? A team of researchers at the Okinawa Institute of Science and Technology built a brain-inspired AI model comprising multiple neural networks. The AI was very limited—it could learn a total of just five nouns and eight verbs. But their AI seems to have learned more than just those words; it learned the concepts behind them.

Babysitting robotic arms

“The inspiration for our model came from developmental psychology. We tried to emulate how infants learn and develop language,” says Prasanna Vijayaraghavan, a researcher at the Okinawa Institute of Science and Technology and the lead author of the study.

While the idea of teaching AIs the same way we teach little babies is not new—we applied it to standard neural nets that associated words with visuals. Researchers also tried teaching an AI using a video feed from a GoPro strapped to a human baby. The problem is babies do way more than just associate items with words when they learn. They touch everything—grasp things, manipulate them, throw stuff around, and this way, they learn to think and plan their actions in language. An abstract AI model couldn’t do any of that, so Vijayaraghavan’s team gave one an embodied experience—their AI was trained in an actual robot that could interact with the world.

Vijayaraghavan’s robot was a fairly simple system with an arm and a gripper that could pick objects up and move them around. Vision was provided by a simple RGB camera feeding videos in a somewhat crude 64×64 pixels resolution.

The robot and the camera were placed in a workspace, put in front of a white table with blocks painted green, yellow, red, purple, and blue. The robot’s task was to manipulate those blocks in response to simple prompts like “move red left,” “move blue right,” or “put red on blue.” All that didn’t seem particularly challenging. What was challenging, though, was building an AI that could process all those words and movements in a manner similar to humans. “I don’t want to say we tried to make the system biologically plausible,” Vijayaraghavan told Ars. “Let’s say we tried to draw inspiration from the human brain.”

Chasing free energy

The starting point for Vijayaraghavan’s team was the free energy principle, a hypothesis that the brain constantly makes predictions about the world based on internal models, then updates these predictions based on sensory input. The idea is that we first think of an action plan to achieve a desired goal, and then this plan is updated in real time based on what we experience during execution. This goal-directed planning scheme, if the hypothesis is correct, governs everything we do, from picking up a cup of coffee to landing a dream job.

All that is closely intertwined with language. Neuroscientists at the University of Parma found that motor areas in the brain got activated when the participants in their study listened to action-related sentences. To emulate that in a robot, Vijayaraghavan used four neural networks working in a closely interconnected system. The first was responsible for processing visual data coming from the camera. It was tightly integrated with a second neural net that handled proprioception: all the processes that ensured the robot was aware of its position and the movement of its body. This second neural net also built internal models of actions necessary to manipulate blocks on the table. Those two neural nets were additionally hooked up to visual memory and attention modules that enabled them to reliably focus on the chosen object and separate it from the image’s background.

The third neural net was relatively simple and processed language using vectorized representations of those “move red right” sentences. Finally, the fourth neural net worked as an associative layer and predicted the output of the previous three at every time step. “When we do an action, we don’t always have to verbalize it, but we have this verbalization in our minds at some point,” Vijayaraghavan says. The AI he and his team built was meant to do just that: seamlessly connect language, proprioception, action planning, and vision.

When the robotic brain was up and running, they started teaching it some of the possible combinations of commands and sequences of movements. But they didn’t teach it all of them.

The birth of compositionality

In 2016, Brenden Lake, a professor of psychology and data science, published a paper in which his team named a set of competencies machines need to master to truly learn and think like humans. One of them was compositionality: the ability to compose or decompose a whole into parts that can be reused. This reuse lets them generalize acquired knowledge to new tasks and situations. “The compositionality phase is when children learn to combine words to explain things. They [initially] learn the names of objects, the names of actions, but those are just single words. When they learn this compositionality concept, their ability to communicate kind of explodes,” Vijayaraghavan explains.

The AI his team built was made for this exact purpose: to see if it would develop compositionality. And it did.

Once the robot learned how certain commands and actions were connected, it also learned to generalize that knowledge to execute commands it never heard before. recognizing the names of actions it had not performed and then performing them on combinations of blocks it had never seen. Vijayaraghavan’s AI figured out the concept of moving something to the right or the left or putting an item on top of something. It could also combine words to name previously unseen actions, like putting a blue block on a red one.

While teaching robots to extract concepts from language has been done before, those efforts were focused on making them understand how words were used to describe visuals. Vijayaragha built on that to include proprioception and action planning, basically adding a layer that integrated sense and movement to the way his robot made sense of the world.

But some issues are yet to overcome. The AI had very limited workspace. The were only a few objects and all had a single, cubical shape. The vocabulary included only names of colors and actions, so no modifiers, adjectives, or adverbs. Finally, the robot had to learn around 80 percent of all possible combinations of nouns and verbs before it could generalize well to the remaining 20 percent. Its performance was worse when those ratios dropped to 60/40 and 40/60.

But it’s possible that just a bit more computing power could fix this. “What we had for this study was a single RTX 3090 GPU, so with the latest generation GPU, we could solve a lot of those issues,” Vijayaraghavan argued. That’s because the team hopes that adding more words and more actions won’t result in a dramatic need for computing power. “We want to scale the system up. We have a humanoid robot with cameras in its head and two hands that can do way more than a single robotic arm. So that’s the next step: using it in the real world with real world robots,” Vijayaraghavan said.

Science Robotics, 2025. DOI: 10.1126/scirobotics.adp0751

Jacek Krywko is a freelance science and technology writer who covers space exploration, artificial intelligence research, computer science, and all sorts of engineering wizardry.

To help AIs understand the world, researchers put them in a robot Read More »

Researchers optimize simulations of molecules on quantum computers

catalysts, Computer science, electrons, Physics, quantum computing, quantum mechanics, Science / Kelly Newman / January 24, 2025

The net result is a much faster operation involving far fewer gates. That’s important because errors in quantum hardware increase as a function of both time and the number of operations.

The researchers then used this approach to explore a chemical, Mn₄O₅Ca, that plays a key role in photosynthesis. Using this approach, they showed it’s possible to calculate what’s called the “spin ladder,” or the list of the lowest-energy states the electrons can occupy. The energy differences between these states correspond to the wavelengths of light they can absorb or emit, so this also defines the spectrum of the molecule.

Faster, but not quite fast enough

We’re not quite ready to run this system on today’s quantum computers, as the error rates are still a bit too high. But because the operations needed to run this sort of algorithm can be done so efficiently, the error rates don’t have to come down very much before the system will become viable. The primary determinant of whether it will run into an error is how far down the time dimension you run the simulation, plus the number of measurements of the system you take over that time.

“The algorithm is especially promising for near-term devices having favorable resource requirements quantified by the number of snapshots (sample complexity) and maximum evolution time (coherence) required for accurate spectral computation,” the researchers wrote.

But the work also makes a couple of larger points. The first is that quantum computers are fundamentally unlike other forms of computation we’ve developed. They’re capable of running things that look like traditional algorithms, where operations are performed and a result is determined. But they’re also quantum systems that are growing in complexity with each new generation of hardware, which makes them great at simulating other quantum systems. And there are a number of hard problems involving quantum systems we’d like to solve.

In some ways, we may only be starting to scratch the surface of quantum computers’ potential. Up until quite recently, there were a lot of hypotheticals; it now appears we’re on the cusp of using one for some potentially useful computations. And that means more people will start thinking about clever ways we can solve problems with them—including cases like this, where the hardware would be used in ways its designers might not have even considered.

Nature Physics, 2025. DOI: 10.1038/s41567-024-02738-z (About DOIs).

Researchers optimize simulations of molecules on quantum computers Read More »

Researchers use AI to design proteins that block snake venom toxins

AI, anti-toxin, biochemistry, Computer science, protein design, Science, snakes, venom / Mike M. / January 15, 2025

Since these two toxicities work through entirely different mechanisms, the researchers tackled them separately.

Blocking a neurotoxin

The neurotoxic three-fingered proteins are a subgroup of the larger protein family that specializes in binding to and blocking the receptors for acetylcholine, a major neurotransmitter. Their three-dimensional structure, which is key to their ability to bind these receptors, is based on three strings of amino acids within the protein that nestle against each other (for those that have taken a sufficiently advanced biology class, these are anti-parallel beta sheets). So to interfere with these toxins, the researchers targeted these strings.

They relied on an AI package called RFdiffusion (the RF denotes its relation to the Rosetta Fold protein-folding software). RFdiffusion can be directed to design protein structures that are complements to specific chemicals; in this case, it identified new strands that could line up along the edge of the ones in the three-fingered toxins. Once those were identified, a separate AI package, called ProteinMPNN, was used to identify the amino acid sequence of a full-length protein that would form the newly identified strands.

But we’re not done with the AI tools yet. The combination of three-fingered toxins and a set of the newly designed proteins were then fed into DeepMind’s AlfaFold2 and the Rosetta protein structure software, and the strength of the interactions between them were estimated.

It’s only at this point that the researchers started making actual proteins, focusing on the candidates that the software suggested would interact the best with the three-fingered toxins. Forty-four of the computer-designed proteins were tested for their ability to interact with the three-fingered toxin, and the single protein that had the strongest interaction was used for further studies.

At this point, it was back to the AI, where RFDiffusion was used to suggest variants of this protein that might bind more effectively. About 15 percent of its suggestions did, in fact, interact more strongly with the toxin. The researchers then made both the toxin and the strongest inhibitor in bacteria and obtained the structure of their interactions. This confirmed that the software’s predictions were highly accurate.

Researchers use AI to design proteins that block snake venom toxins Read More »

Meta takes us a step closer to Star Trek’s universal translator

AI, Computer science, human speech, Science, text processing, translation / 9u50fv / January 15, 2025

The computer science behind translating speech from 100 source languages.

In 2023, AI researchers at Meta interviewed 34 native Spanish and Mandarin speakers who lived in the US but didn’t speak English. The goal was to find out what people who constantly rely on translation in their day-to-day activities expect from an AI translation tool. What those participants wanted was basically a Star Trek universal translator or the Babel Fish from the Hitchhiker’s Guide to the Galaxy: an AI that could not only translate speech to speech in real time across multiple languages, but also preserve their voice, tone, mannerisms, and emotions. So, Meta assembled a team of over 50 people and got busy building it.

What this team came up with was a next-gen translation system called Seamless. The first building block of this system is described in Wednesday’s issue of Nature; it can translate speech among 36 different languages.

Language data problems

AI translation systems today are mostly focused on text, because huge amounts of text are available in a wide range of languages thanks to digitization and the Internet. Institutions like the United Nations or European Parliament routinely translate all their proceedings into the languages of all their member states, which means there are enormous databases comprising aligned documents prepared by professional human translators. You just needed to feed those huge, aligned text corpora into neural nets (or hidden Markov models before neural nets became all the rage) and you ended up with a reasonably good machine translation system. But there were two problems with that.

The first issue was those databases comprised formal documents, which made the AI translators default to the same boring legalese in the target language even if you tried to translate comedy. The second problem was speech—none of this included audio data.

The problem of language formality was mostly solved by including less formal sources like books, Wikipedia, and similar material in AI training databases. The scarcity of aligned audio data, however, remained. Both issues were at least theoretically manageable in high-resource languages like English or Spanish, but they got dramatically worse in low-resource languages like Icelandic or Zulu.

As a result, the AI translators we have today support an impressive number of languages in text, but things are complicated when it comes to translating speech. There are cascading systems that simply do this trick in stages. An utterance is first converted to text just as it would be in any dictation service. Then comes text-to-text translation, and finally the resulting text in the target language is synthesized into speech. Because errors accumulate at each of those stages, the performance you get this way is usually poor, and it doesn’t work in real time.

A few systems that can translate speech-to-speech directly do exist, but in most cases they only translate into English and not in the opposite way. Your foreign language interlocutor can say something to you in one of the languages supported by tools like Google’s AudioPaLM, and they will translate that to English speech, but you can’t have a conversation going both ways.

So, to pull off the Star Trek universal translator thing Meta’s interviewees dreamt about, the Seamless team started with sorting out the data scarcity problem. And they did it in a quite creative way.

Building a universal language

Warren Weaver, a mathematician and pioneer of machine translation, argued in 1949 that there might be a yet undiscovered universal language working as a common base of human communication. This common base of all our communication was exactly what the Seamless team went for in its search for data more than 70 years later. Weaver’s universal language turned out to be math—more precisely, multidimensional vectors.

Machines do not understand words as humans do. To make sense of them, they need to first turn them into sequences of numbers that represent their meaning. Those sequences of numbers are numerical vectors that are termed word embeddings. When you vectorize tens of millions of documents this way, you’ll end up with a huge multidimensional space where words with similar meaning that often go together, like “tea” and “coffee,” are placed close to each other. When you vectorize aligned text in two languages like those European Parliament proceedings, you end up with two separate vector spaces, and then you can run a neural net to learn how those two spaces map onto each other.

But the Meta team didn’t have those nicely aligned texts for all the languages they wanted to cover. So, they vectorized all texts in all languages as if they were just a single language and dumped them into one embedding space called SONAR (Sentence-level Multimodal and Language-Agnostic Representations). Once the text part was done, they went to speech data, which was vectorized using a popular W2v (word to vector) tool and added it to the same massive multilingual, multimodal space. Of course, each embedding carried metadata identifying its source language and whether it was text or speech before vectorization.

The team just used huge amounts of raw data—no fancy human labeling, no human-aligned translations. And then, the data mining magic happened.

SONAR embeddings represented entire sentences instead of single words. Part of the reason behind that was to control for differences between morphologically rich languages, where a single word may correspond to multiple words in morphologically simple languages. But the most important thing was that it ensured that sentences with similar meaning in multiple languages ended up close to each other in the vector space.

It was the same story with speech, too—a spoken sentence in one language was close to spoken sentences in other languages with similar meaning. It even worked between text and speech. So, the team simply assumed that embeddings in two different languages or two different modalities (speech or text) that are at a sufficiently close distance to each other are equivalent to the manually aligned texts of translated documents.

This produced huge amounts of automatically aligned data. The Seamless team suddenly got access to millions of aligned texts, even in low-resource languages, along with thousands of hours of transcribed audio. And they used all this data to train their next-gen translator.

Seamless translation

The automatically generated data set was augmented with human-curated texts and speech samples where possible and used to train multiple AI translation models. The largest one was called SEAMLESSM4T v2. It could translate speech to speech from 101 source languages into any of 36 output languages, and translate text to text. It would also work as an automatic speech recognition system in 96 languages, translate speech to text from 101 into 96 languages, and translate text to speech from 96 into 36 languages—all from a single unified model. It also outperformed state-of-the-art cascading systems by 8 percent in a speech-to-text and by 23 percent in a speech-to-speech translations based on the scores in Bilingual Evaluation Understudy (an algorithm commonly used to evaluate the quality of machine translation).

But it can now do even more than that. The Nature paper published by Meta’s Seamless ends at the SEAMLESSM4T models, but Nature has a long editorial process to ensure scientific accuracy. The paper published on January 15, 2025, was submitted in late November 2023. But in a quick search of the arXiv.org, a repository of not-yet-peer-reviewed papers, you can find the details of two other models that the Seamless team has already integrated on top of the SEAMLESSM4T: SeamlessStreaming and SeamlessExpressive, which take this AI even closer to making a Star Trek universal translator a reality.

SeamlessStreaming is meant to solve the translation latency problem. The baseline SEAMLESSM4T, despite all the bells and whistles, worked as a standard AI translation tool. You had to say what you wanted to say, push “translate,” and it spat out the translation. SeamlessStreaming was designed to take this experience a bit closer to what human simultaneous translator do—it translates what you’re saying as you speak in a streaming fashion. SeamlessExpressive, on the other hand, is aimed at preserving the way you express yourself in translations. When you whisper or say something in a cheerful manner or shout out with anger, SeamlessExpressive will encode the features of your voice, like tone, prosody, volume, tempo, and so on, and transfer those into the output speech in the target language.

Sadly, it still can’t do both at the same time; you can only choose to go for either streaming or expressivity, at least at the moment. Also, the expressivity variant is very limited in supported languages—it only works in English, Spanish, French, and German. But at least it’s online so you can go ahead and give it a spin.

Nature, 2025. DOI: 10.1038/s41586-024-08359-z

Jacek Krywko is a freelance science and technology writer who covers space exploration, artificial intelligence research, computer science, and all sorts of engineering wizardry.

Meta takes us a step closer to Star Trek’s universal translator Read More »

$getting-an-all-optical-ai-to-handle-non-linear-math$

Getting an all-optical AI to handle non-linear math

AI, Computer science, optical processing, optics, Science / DJ Henderson / January 12, 2025

The problem is that this cascading requires massive parallel computations that, when done on standard computers, take tons of energy and time. Bandyopadhyay’s team feels this problem can be solved by performing the equivalent operations using photons rather than electrons. In photonic chips, information can be encoded in optical properties like polarization, phase, magnitude, frequency, and wavevector. While this would be extremely fast and energy-efficient, building such chips isn’t easy.

Siphoning light

“Conveniently, photonics turned out to be particularly good at linear matrix operations,” Bandyopadhyay claims. A group at MIT led by Dirk Englund, a professor who is a co-author of Bandyopadhyay’s study, demonstrated a photonic chip doing matrix multiplication entirely with light in 2017. What the field struggled with, though, was implementing non-linear functions in photonics.

The usual solution, so far, relied on bypassing the problem by doing linear algebra on photonic chips and offloading non-linear operations to external electronics. This, however, increased latency, since the information had to be converted from light to electrical signals, processed on an external processor, and converted back to light. “And bringing the latency down is the primary reason why we want to build neural networks in photonics,” Bandyopadhyay says.

To solve this problem, Bandyopadhyay and his colleagues designed and built what is likely to be the world’s first chip that can compute the entire deep neural net, including both linear and non-linear operations, using photons. “The process starts with an external laser with a modulator that feeds light into the chip through an optical fiber. This way we convert electrical inputs to light,” Bandyopadhyay explains.

The light is then fanned out to six channels and fed into a layer of six neurons that perform linear matrix multiplication using an array of devices called Mach-Zehnder interferometers. “They are essentially programmable beam splitters, taking two optical fields and mixing them coherently to produce two output optical fields. By applying the voltage, you can control how much those the two inputs mix,” Bandyopadhyay says.

Getting an all-optical AI to handle non-linear math Read More »

It’s remarkably easy to inject new medical misinformation into LLMs

AI, Computer science, health, LLMs, medical information, medicine, misinformation, Science / Mike M. / January 8, 2025

Changing just 0.001% of inputs to misinformation makes the AI less accurate.

It’s pretty easy to see the problem here: The Internet is brimming with misinformation, and most large language models are trained on a massive body of text obtained from the Internet.

Ideally, having substantially higher volumes of accurate information might overwhelm the lies. But is that really the case? A new study by researchers at New York University examines how much medical information can be included in a large language model (LLM) training set before it spits out inaccurate answers. While the study doesn’t identify a lower bound, it does show that by the time misinformation accounts for 0.001 percent of the training data, the resulting LLM is compromised.

While the paper is focused on the intentional “poisoning” of an LLM during training, it also has implications for the body of misinformation that’s already online and part of the training set for existing LLMs, as well as the persistence of out-of-date information in validated medical databases.

Sampling poison

Data poisoning is a relatively simple concept. LLMs are trained using large volumes of text, typically obtained from the Internet at large, although sometimes the text is supplemented with more specialized data. By injecting specific information into this training set, it’s possible to get the resulting LLM to treat that information as a fact when it’s put to use. This can be used for biasing the answers returned.

This doesn’t even require access to the LLM itself; it simply requires placing the desired information somewhere where it will be picked up and incorporated into the training data. And that can be as simple as placing a document on the web. As one manuscript on the topic suggested, “a pharmaceutical company wants to push a particular drug for all kinds of pain which will only need to release a few targeted documents in [the] web.”

Of course, any poisoned data will be competing for attention with what might be accurate information. So, the ability to poison an LLM might depend on the topic. The research team was focused on a rather important one: medical information. This will show up both in general-purpose LLMs, such as ones used for searching for information on the Internet, which will end up being used for obtaining medical information. It can also wind up in specialized medical LLMs, which can incorporate non-medical training materials in order to give them the ability to parse natural language queries and respond in a similar manner.

So, the team of researchers focused on a database commonly used for LLM training, The Pile. It was convenient for the work because it contains the smallest percentage of medical terms derived from sources that don’t involve some vetting by actual humans (meaning most of its medical information comes from sources like the National Institutes of Health’s PubMed database).

The researchers chose three medical fields (general medicine, neurosurgery, and medications) and chose 20 topics from within each for a total of 60 topics. Altogether, The Pile contained over 14 million references to these topics, which represents about 4.5 percent of all the documents within it. Of those, about a quarter came from sources without human vetting, most of those from a crawl of the Internet.

The researchers then set out to poison The Pile.

Finding the floor

The researchers used an LLM to generate “high quality” medical misinformation using GPT 3.5. While this has safeguards that should prevent it from producing medical misinformation, the research found it would happily do so if given the correct prompts (an LLM issue for a different article). The resulting articles could then be inserted into The Pile. Modified versions of The Pile were generated where either 0.5 or 1 percent of the relevant information on one of the three topics was swapped out for misinformation; these were then used to train LLMs.

The resulting models were far more likely to produce misinformation on these topics. But the misinformation also impacted other medical topics. “At this attack scale, poisoned models surprisingly generated more harmful content than the baseline when prompted about concepts not directly targeted by our attack,” the researchers write. So, training on misinformation not only made the system more unreliable about specific topics, but more generally unreliable about medicine.

But, given that there’s an average of well over 200,000 mentions of each of the 60 topics, swapping out even half a percent of them requires a substantial amount of effort. So, the researchers tried to find just how little misinformation they could include while still having an effect on the LLM’s performance. Unfortunately, this didn’t really work out.

Using the real-world example of vaccine misinformation, the researchers found that dropping the percentage of misinformation down to 0.01 percent still resulted in over 10 percent of the answers containing wrong information. Going for 0.001 percent still led to over 7 percent of the answers being harmful.

“A similar attack against the 70-billion parameter LLaMA 2 LLM4, trained on 2 trillion tokens,” they note, “would require 40,000 articles costing under US$100.00 to generate.” The “articles” themselves could just be run-of-the-mill webpages. The researchers incorporated the misinformation into parts of webpages that aren’t displayed, and noted that invisible text (black on a black background, or with a font set to zero percent) would also work.

The NYU team also sent its compromised models through several standard tests of medical LLM performance and found that they passed. “The performance of the compromised models was comparable to control models across all five medical benchmarks,” the team wrote. So there’s no easy way to detect the poisoning.

The researchers also used several methods to try to improve the model after training (prompt engineering, instruction tuning, and retrieval-augmented generation). None of these improved matters.

Existing misinformation

Not all is hopeless. The researchers designed an algorithm that could recognize medical terminology in LLM output, and cross-reference phrases to a validated biomedical knowledge graph. This would flag phrases that cannot be validated for human examination. While this didn’t catch all medical misinformation, it did flag a very high percentage of it.

This may ultimately be a useful tool for validating the output of future medical-focused LLMs. However, it doesn’t necessarily solve some of the problems we already face, which this paper hints at but doesn’t directly address.

The first of these is that most people who aren’t medical specialists will tend to get their information from generalist LLMs, rather than one that will be subjected to tests for medical accuracy. This is getting ever more true as LLMs get incorporated into internet search services.

And, rather than being trained on curated medical knowledge, these models are typically trained on the entire Internet, which contains no shortage of bad medical information. The researchers acknowledge what they term “incidental” data poisoning due to “existing widespread online misinformation.” But a lot of that “incidental” information was generally produced intentionally, as part of a medical scam or to further a political agenda. Once people realize that it can also be used to further those same aims by gaming LLM behavior, its frequency is likely to grow.

Finally, the team notes that even the best human-curated data sources, like PubMed, also suffer from a misinformation problem. The medical research literature is filled with promising-looking ideas that never panned out, and out-of-date treatments and tests that have been replaced by approaches more solidly based on evidence. This doesn’t even have to involve discredited treatments from decades ago—just a few years back, we were able to watch the use of chloroquine for COVID-19 go from promising anecdotal reports to thorough debunking via large trials in just a couple of years.

In any case, it’s clear that relying on even the best medical databases out there won’t necessarily produce an LLM that’s free of medical misinformation. Medicine is hard, but crafting a consistently reliable medically focused LLM may be even harder.

Nature Medicine, 2025. DOI: 10.1038/s41591-024-03445-1 (About DOIs).

John is Ars Technica’s science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from his keyboard, he tends to seek out a bicycle, or a scenic location for communing with his hiking boots.

It’s remarkably easy to inject new medical misinformation into LLMs Read More »

Google gets an error-corrected quantum bit to be stable for an hour

Computer science, Google, logical qubits, Physics, quantum computing, quantum error correction, quantum mechanics, Science / Beth Washington / December 9, 2024

Using almost the entire chip for a logical qubit provides long-term stability.

Google’s new Willow chip is its first new generation of chips in about five years. Credit: Google

On Monday, Nature released a paper from Google’s quantum computing team that provides a key demonstration of the potential of quantum error correction. Thanks to an improved processor, Google’s team found that increasing the number of hardware qubits dedicated to an error-corrected logical qubit led to an exponential increase in performance. By the time the entire 105-qubit processor was dedicated to hosting a single error-corrected qubit, the system was stable for an average of an hour.

In fact, Google told Ars that errors on this single logical qubit were rare enough that it was difficult to study them. The work provides a significant validation that quantum error correction is likely to be capable of supporting the execution of complex algorithms that might require hours to execute.

A new fab

Google is making a number of announcements in association with the paper’s release (an earlier version of the paper has been up on the arXiv since August). One of those is that the company is committed enough to its quantum computing efforts that it has built its own fabrication facility for its superconducting processors.

“In the past, all the Sycamore devices that you’ve heard about were fabricated in a shared university clean room space next to graduate students and people doing kinds of crazy stuff,” Google’s Julian Kelly said. “And we’ve made this really significant investment in bringing this new facility online, hiring staff, filling it with tools, transferring their process over. And that enables us to have significantly more process control and dedicated tooling.”

That’s likely to be a critical step for the company, as the ability to fabricate smaller test devices can allow the exploration of lots of ideas on how to structure the hardware to limit the impact of noise. The first publicly announced product of this lab is the Willow processor, Google’s second design, which ups its qubit count to 105. Kelly said one of the changes that came with Willow actually involved making the individual pieces of the qubit larger, which makes them somewhat less susceptible to the influence of noise.

All of that led to a lower error rate, which was critical for the work done in the new paper. This was demonstrated by running Google’s favorite benchmark, one that it acknowledges is contrived in a way to make quantum computing look as good as possible. Still, people have figured out how to make algorithm improvements for classical computers that have kept them mostly competitive. But, with all the improvements, Google expects that the quantum hardware has moved firmly into the lead. “We think that the classical side will never outperform quantum in this benchmark because we’re now looking at something on our new chip that takes under five minutes, would take 10²⁵ years, which is way longer than the age of the Universe,” Kelly said.

Building logical qubits

The work focuses on the behavior of logical qubits, in which a collection of individual hardware qubits are grouped together in a way that enables errors to be detected and corrected. These are going to be essential for running any complex algorithms, since the hardware itself experiences errors often enough to make some inevitable during any complex calculations.

This naturally creates a key milestone. You can get better error correction by adding more hardware qubits to each logical qubit. If each of those hardware qubits produces errors at a sufficient rate, however, then you’ll experience errors faster than you can correct for them. You need to get hardware qubits of a sufficient quality before you start benefitting from larger logical qubits. Google’s earlier hardware had made it past that milestone, but only barely. Adding more hardware qubits to each logical qubit only made for a marginal improvement.

That’s no longer the case. Google’s processors have the hardware qubits laid out on a square grid, with each connected to its nearest neighbors (typically four except at the edges of the grid). And there’s a specific error correction code structure, called the surface code, that fits neatly into this grid. And you can use surface codes of different sizes by using progressively more of the grid. The size of the grid being used is measured by a term called distance, with larger distance meaning a bigger logical qubit, and thus better error correction.

(In addition to a standard surface code, Google includes a few qubits that handle a phenomenon called “leakage,” where a qubit ends up in a higher-energy state, instead of the two low-energy states defined as zero and one.)

The key result is that going from a distance of three to a distance of five more than doubled the ability of the system to catch and correct errors. Going from a distance of five to a distance of seven doubled it again. Which shows that the hardware qubits have reached a sufficient quality that putting more of them into a logical qubit has an exponential effect.

“As we increase the grid from three by three to five by five to seven by seven, the error rate is going down by a factor of two each time,” said Google’s Michael Newman. “And that’s that exponential error suppression that we want.”

Going big

The second thing they demonstrated is that, if you make the largest logical qubit that the hardware can support, with a distance of 15, it’s possible to hang onto the quantum information for an average of an hour. This is striking because Google’s earlier work had found that its processors experience widespread simultaneous errors that the team ascribed to cosmic ray impacts. (IBM, however, has indicated it doesn’t see anything similar, so it’s not clear whether this diagnosis is correct.) Those happened every 10 seconds or so. But this work shows that a sufficiently large error code can correct for these events, whatever their cause.

That said, these qubits don’t survive indefinitely. One of them seems to be a localized temporary increase in errors. The second, more difficult to deal with problem involves a widespread spike in error detection affecting an area that includes roughly 30 qubits. At this point, however, Google has only seen six of these events, so they told Ars that it’s difficult to really characterize them. “It’s so rare it actually starts to become a bit challenging to study because you have to gain a lot of statistics to even see those events at all,” said Kelly.

Beyond the relative durability of these logical qubits, the paper notes another advantage to going with larger code distances: it enhances the impact of further hardware improvements. Google estimates that at a distance of 15, improving hardware performance by a factor of two would drop errors in the logical qubit by a factor of 250. At a distance of 27, the same hardware improvement would lead to an improvement of over 10,000 in the logical qubit’s performance.

Note that none of this will ever get the error rate to zero. Instead, we just need to get the error rate to a level where an error is unlikely for a given calculation (more complex calculations will require a lower error rate). “It’s worth understanding that there’s always going to be some type of error floor and you just have to push it low enough to the point where it practically is irrelevant,” Kelly said. “So for example, we could get hit by an asteroid and the entire Earth could explode and that would be a correlated error that our quantum computer is not currently built to be robust to.”

Obviously, a lot of additional work will need to be done to both make logical qubits like this survive for even longer, and to ensure we have the hardware to host enough logical qubits to perform calculations. But the exponential improvements here, to Google, suggest that there’s nothing obvious standing in the way of that. “We woke up one morning and we kind of got these results and we were like, wow, this is going to work,” Newman said. “This is really it.”

Nature, 2024. DOI: 10.1038/s41586-024-08449-y (About DOIs).

Google gets an error-corrected quantum bit to be stable for an hour Read More »

Google’s DeepMind tackles weather forecasting, with great performance

AI, atmospheric science, Computer science, Earth science, meteorology, Science, weather forecasting / DJ Henderson / December 4, 2024

By some measures, AI systems are now competitive with traditional computing methods for generating weather forecasts. Because their training penalizes errors, however, the forecasts tend to get “blurry”—as you move further ahead in time, the models make fewer specific predictions since those are more likely to be wrong. As a result, you start to see things like storm tracks broadening and the storms themselves losing clearly defined edges.

But using AI is still extremely tempting because the alternative is a computational atmospheric circulation model, which is extremely compute-intensive. Still, it’s highly successful, with the ensemble model from the European Centre for Medium-Range Weather Forecasts considered the best in class.

In a paper being released today, Google’s DeepMind claims its new AI system manages to outperform the European model on forecasts out to at least a week and often beyond. DeepMind’s system, called GenCast, merges some computational approaches used by atmospheric scientists with a diffusion model, commonly used in generative AI. The result is a system that maintains high resolution while cutting the computational cost significantly.

Ensemble forecasting

Traditional computational methods have two main advantages over AI systems. The first is that they’re directly based on atmospheric physics, incorporating the rules we know govern the behavior of our actual weather, and they calculate some of the details in a way that’s directly informed by empirical data. They’re also run as ensembles, meaning that multiple instances of the model are run. Due to the chaotic nature of the weather, these different runs will gradually diverge, providing a measure of the uncertainty of the forecast.

At least one attempt has been made to merge some of the aspects of traditional weather models with AI systems. An internal Google project used a traditional atmospheric circulation model that divided the Earth’s surface into a grid of cells but used an AI to predict the behavior of each cell. This provided much better computational performance, but at the expense of relatively large grid cells, which resulted in relatively low resolution.

For its take on AI weather predictions, DeepMind decided to skip the physics and instead adopt the ability to run an ensemble.

Gen Cast is based on diffusion models, which have a key feature that’s useful here. In essence, these models are trained by starting them with a mixture of an original—image, text, weather pattern—and then a variation where noise is injected. The system is supposed to create a variation of the noisy version that is closer to the original. Once trained, it can be fed pure noise and evolve the noise to be closer to whatever it’s targeting.

In this case, the target is realistic weather data, and the system takes an input of pure noise and evolves it based on the atmosphere’s current state and its recent history. For longer-range forecasts, the “history” includes both the actual data and the predicted data from earlier forecasts. The system moves forward in 12-hour steps, so the forecast for day three will incorporate the starting conditions, the earlier history, and the two forecasts from days one and two.

This is useful for creating an ensemble forecast because you can feed it different patterns of noise as input, and each will produce a slightly different output of weather data. This serves the same purpose it does in a traditional weather model: providing a measure of the uncertainty for the forecast.

For each grid square, GenCast works with six weather measures at the surface, along with six that track the state of the atmosphere and 13 different altitudes at which it estimates the air pressure. Each of these grid squares is 0.2 degrees on a side, a higher resolution than the European model uses for its forecasts. Despite that resolution, DeepMind estimates that a single instance (meaning not a full ensemble) can be run out to 15 days on one of Google’s tensor processing systems in just eight minutes.

It’s possible to make an ensemble forecast by running multiple versions of this in parallel and then integrating the results. Given the amount of hardware Google has at its disposal, the whole process from start to finish is likely to take less than 20 minutes. The source and training data will be placed on the GitHub page for DeepMind’s GraphCast project. Given the relatively low computational requirements, we can probably expect individual academic research teams to start experimenting with it.

Measures of success

DeepMind reports that GenCast dramatically outperforms the best traditional forecasting model. Using a standard benchmark in the field, DeepMind found that GenCast was more accurate than the European model on 97 percent of the tests it used, which checked different output values at different times in the future. In addition, the confidence values, based on the uncertainty obtained from the ensemble, were generally reasonable.

Past AI weather forecasters, having been trained on real-world data, are generally not great at handling extreme weather since it shows up so rarely in the training set. But GenCast did quite well, often outperforming the European model in things like abnormally high and low temperatures and air pressure (one percent frequency or less, including at the 0.01 percentile).

DeepMind also went beyond standard tests to determine whether GenCast might be useful. This research included projecting the tracks of tropical cyclones, an important job for forecasting models. For the first four days, GenCast was significantly more accurate than the European model, and it maintained its lead out to about a week.

One of DeepMind’s most interesting tests was checking the global forecast of wind power output based on information from the Global Powerplant Database. This involved using it to forecast wind speeds at 10 meters above the surface (which is actually lower than where most turbines reside but is the best approximation possible) and then using that number to figure out how much power would be generated. The system beat the traditional weather model by 20 percent for the first two days and stayed in front with a declining lead out to a week.

The researchers don’t spend much time examining why performance seems to decline gradually for about a week. Ideally, more details about GenCast’s limitations would help inform further improvements, so the researchers are likely thinking about it. In any case, today’s paper marks the second case where taking something akin to a hybrid approach—mixing aspects of traditional forecast systems with AI—has been reported to improve forecasts. And both those cases took very different approaches, raising the prospect that it will be possible to combine some of their features.

Nature, 2024. DOI: 10.1038/s41586-024-08252-9 (About DOIs).

Google’s DeepMind tackles weather forecasting, with great performance Read More »