Today, almost all data on the Internet, including bank transactions, medical records, and secure chats, is protected with an encryption scheme called RSA (named after its creators Rivest, Shamir, and Adleman). This scheme is based on a simple fact—it is virtually impossible to calculate the prime factors of a large number in a reasonable amount of time, even on the world’s most powerful supercomputer. Unfortunately, large quantum computers, if and when they are built, would find this task a breeze, thus undermining the security of the entire Internet.

Luckily, quantum computers are only better than classical ones at a select class of problems, and there are plenty of encryption schemes where quantum computers don’t offer any advantage. Today, the U.S. National Institute of Standards and Technology (NIST) announced the standardization of three post-quantum cryptography encryption schemes. With these standards in hand, NIST is encouraging computer system administrators to begin transitioning to post-quantum security as soon as possible.

“Now our task is to replace the protocol in every device, which is not an easy task.” —Lily Chen, NIST

These standards are likely to be a big element of the Internet’s future. NIST’s previous cryptography standards, developed in the 1970s, are used in almost all devices, including Internet routers, phones, and laptops, says Lily Chen, head of the cryptography group at NIST who lead the standardization process. But adoption will not happen overnight.

“Today, public key cryptography is used everywhere in every device,” Chen says. “Now our task is to replace the protocol in every device, which is not an easy task.”

Why we need post-quantum cryptography now

Most experts believe large-scale quantum computers won’t be built for at least another decade. So why is NIST worried about this now? There are two main reasons.

First, many devices that use RSA security, like cars and some IoT devices, are expected to remain in use for at least a decade. So they need to be equipped with quantum-safe cryptography before they are released into the field.

“For us, it’s not an option to just wait and see what happens. We want to be ready and implement solutions as soon as possible.” —Richard Marty, LGT Financial Services

Second, a nefarious individual could potentially download and store encrypted data today, and decrypt it once a large enough quantum computer comes online. This concept is called “harvest now, decrypt later“ and by its nature, it poses a threat to sensitive data now, even if that data can only be cracked in the future.

Security experts in various industries are starting to take the threat of quantum computers seriously, says Joost Renes, principal security architect and cryptographer at NXP Semiconductors. “Back in 2017, 2018, people would ask ‘What’s a quantum computer?’” Renes says. “Now, they’re asking ‘When will the PQC standards come out and which one should we implement?’”

Richard Marty, chief technology officer at LGT Financial Services, agrees. “For us, it’s not an option to just wait and see what happens. We want to be ready and implement solutions as soon as possible, to avoid harvest now and decrypt later.”

NIST’s competition for the best quantum-safe algorithm

NIST announced a public competition for the best PQC algorithm back in 2016. They received a whopping 82 submissions from teams in 25 different countries. Since then, NIST has gone through 4 elimination rounds, finally whittling the pool down to four algorithms in 2022.

This lengthy process was a community-wide effort, with NIST taking input from the cryptographic research community, industry, and government stakeholders. “Industry has provided very valuable feedback,” says NIST’s Chen.

These four winning algorithms had intense-sounding names: CRYSTALS-Kyber, CRYSTALS-Dilithium, Sphincs+, and FALCON. Sadly, the names did not survive standardization: The algorithms are now known as Federal Information Processing Standard (FIPS) 203 through 206. FIPS 203, 204, and 205 are the focus of today’s announcement from NIST. FIPS 206, the algorithm previously known as FALCON, is expected to be standardized in late 2024.

The algorithms fall into two categories: general encryption, used to protect information transferred via a public network, and digital signature, used to authenticate individuals. Digital signatures are essential for preventing malware attacks, says Chen.

Every cryptography protocol is based on a math problem that’s hard to solve but easy to check once you have the correct answer. For RSA, it’s factoring large numbers into two primes—it’s hard to figure out what those two primes are (for a classical computer), but once you have one it’s straightforward to divide and get the other.

“We have a few instances of [PQC], but for a full transition, I couldn’t give you a number, but there’s a lot to do.” —Richard Marty, LGT Financial Services

Two out of the three schemes already standardized by NIST, FIPS 203 and FIPS 204 (as well as the upcoming FIPS 206), are based on another hard problem, called lattice cryptography. Lattice cryptography rests on the tricky problem of finding the lowest common multiple among a set of numbers. Usually, this is implemented in many dimensions, or on a lattice, where the least common multiple is a vector.

The third standardized scheme, FIPS 205, is based on hash functions—in other words, converting a message to an encrypted string that’s difficult to reverse

The standards include the encryption algorithms’ computer code, instructions for how to implement it, and intended uses. There are three levels of security for each protocol, designed to future-proof the standards in case some weaknesses or vulnerabilities are found in the algorithms.

Lattice cryptography survives alarms over vulnerabilities

Earlier this year, a pre-print published to the arXiv alarmed the PQC community. The paper, authored by Yilei Chen of Tsinghua University in Beijing, claimed to show that lattice-based cryptography, the basis of two out of the three NIST protocols, was not, in fact, immune to quantum attacks. On further inspection, Yilei Chen’s argument turned out to have a flaw—and lattice cryptography is still believed to be secure against quantum attacks.

On the one hand, this incident highlights the central problem at the heart of all cryptography schemes: There is no proof that any of the math problems the schemes are based on are actually “hard.” The only proof, even for the standard RSA algorithms, is that people have been trying to break the encryption for a long time, and have all failed. Since post-quantum cryptography standards, including lattice cryptogrphay, are newer, there is less certainty that no one will find a way to break them.

That said, the failure of this latest attempt only builds on the algorithm’s credibility. The flaw in the paper’s argument was discovered within a week, signaling that there is an active community of experts working on this problem. “The result of that paper is not valid, that means the pedigree of the lattice-based cryptography is still secure,” says NIST’s Lily Chen (no relation to Tsinghua University’s Yilei Chen). “People have tried hard to break this algorithm. A lot of people are trying, they try very hard, and this actually gives us confidence.”

NIST’s announcement is exciting, but the work of transitioning all devices to the new standards has only just begun. It is going to take time, and money, to fully protect the world from the threat of future quantum computers.

“We’ve spent 18 months on the transition and spent about half a million dollars on it,” says Marty of LGT Financial Services. “We have a few instances of [PQC], but for a full transition, I couldn’t give you a number, but there’s a lot to do.”

As large supercomputers keep getting larger, Sunnyvale, California-based Cerebras has been taking a different approach. Instead of connecting more and more GPUs together, the company has been squeezing as many processors as it can onto one giant wafer. The main advantage is in the interconnects—by wiring processors together on-chip, the wafer-scale chip bypasses many of the computational speed losses that come from many GPUs talking to each other, as well as losses from loading data to and from memory.

Now, Cerebras has flaunted the advantages of their wafer-scale chips in two separate but related results. First, the company demonstrated that its second generation wafer-scale engine, WSE-2, was significantly faster than world’s fastest supercomputer, Frontier, in molecular dynamics calculations—the field that underlies protein folding, modeling radiation damage in nuclear reactors, and other problems in material science. Second, in collaboration with machine learning model optimization company Neural Magic, Cerebras demonstrated that a sparse large language model could perform inference at one-third of the energy cost of a full model without losing any accuracy. Although the results are in vastly different fields, they were both possible because of the interconnects and fast memory access enabled by Cerebras’ hardware.

Speeding Through the Molecular World

“Imagine there’s a tailor and he can make a suit in a week,” says Cerebras CEO and co-founder Andrew Feldman. “He buys the neighboring tailor, and she can also make a suit in a week, but they can’t work together. Now, they can now make two suits in a week. But what they can’t do is make a suit in three and a half days.”

According to Feldman, GPUs are like tailors that can’t work together, at least when it comes to some problems in molecular dynamics. As you connect more and more GPUs, they can simulate more atoms at the same time, but they can’t simulate the same number of atoms more quickly.

Cerebras’ wafer-scale engine, however, scales in a fundamentally different way. Because the chips are not limited by interconnect bandwidth, they can communicate quickly, like two tailors collaborating perfectly to make a suit in three and a half days.

“It’s difficult to create materials that have the right properties, that have a long lifetime and sufficient strength and don’t break.” —Tomas Oppelstrup, Lawrence Livermore National Laboratory

To demonstrate this advantage, the team simulated 800,000 atoms interacting with each other, calculating the interactions in increments of one femtosecond at a time. Each step took just microseconds to compute on their hardware. Although that’s still 9 orders of magnitude slower than the actual interactions, it was also 179 times as fast as the Frontier supercomputer. The achievement effectively reduced a year’s worth of computation to just two days.

This work was done in collaboration with Sandia, Lawrence Livermore, and Los Alamos National Laboratories. Tomas Oppelstrup, staff scientist at Lawrence Livermore National Laboratory, says this advance makes it feasible to simulate molecular interactions that were previously inaccessible.

Oppelstrup says this will be particularly useful for understanding the longer-term stability of materials in extreme conditions. “When you build advanced machines that operate at high temperatures, like jet engines, nuclear reactors, or fusion reactors for energy production,” he says, “you need materials that can withstand these high temperatures and very harsh environments. It’s difficult to create materials that have the right properties, that have a long lifetime and sufficient strength and don’t break.” Being able to simulate the behavior of candidate materials for longer, Oppelstrup says, will be crucial to the material design and development process.

Ilya Sharapov, principal engineer at Cerebras, say the company is looking forward to extending applications of its wafer-scale engine to a larger class of problems, including molecular dynamics simulations of biological processes and simulations of airflow around cars or aircrafts.

Downsizing Large Language Models

As large language models (LLMs) are becoming more popular, the energy costs of using them are starting to overshadow the training costs—potentially by as much as a factor of ten in some estimates. “Inference is is the primary workload of AI today because everyone is using ChatGPT,” says James Wang, director of product marketing at Cerebras, “and it’s very expensive to run especially at scale.”

One way to reduce the energy cost (and speed) of inference is through sparsity—essentially, harnessing the power of zeros. LLMs are made up of huge numbers of parameters. The open-source Llama model used by Cerebras, for example, has 7 billion parameters. During inference, each of those parameters is used to crunch through the input data and spit out the output. If, however, a significant fraction of those parameters are zeros, they can be skipped during the calculation, saving both time and energy.

The problem is that skipping specific parameters is a difficult to do on a GPU. Reading from memory on a GPU is relatively slow, because they’re designed to read memory in chunks, which means taking in groups of parameters at a time. This doesn’t allow GPUs to skip zeros that are randomly interspersed in the parameter set. Cerebras CEO Feldman offered another analogy: “It’s equivalent to a shipper, only wanting to move stuff on pallets because they don’t want to examine each box. Memory bandwidth is the ability to examine each box to make sure it’s not empty. If it’s empty, set it aside and then not move it.”

“There’s a million cores in a very tight package, meaning that the cores have very low latency, high bandwidth interactions between them.” —Ilya Sharapov, Cerebras

Some GPUs are equipped for a particular kind of sparsity, called 2:4, where exactly two out of every four consecutively stored parameters are zeros. State-of-the-art GPUs have terabytes per second of memory bandwidth. The memory bandwidth of Cerebras’ WSE-2 is more than one thousand times as high, at 20 petabytes per second. This allows for harnessing unstructured sparsity, meaning the researchers can zero out parameters as needed, wherever in the model they happen to be, and check each one on the fly during a computation. “Our hardware is built right from day one to support unstructured sparsity,” Wang says.

Even with the appropriate hardware, zeroing out many of the model’s parameters results in a worse model. But the joint team from Neural Magic and Cerebras figured out a way to recover the full accuracy of the original model. After slashing 70 percent of the parameters to zero, the team performed two further phases of training to give the non-zero parameters a chance to compensate for the new zeros.

This extra training uses about 7 percent of the original training energy, and the companies found that they recover full model accuracy with this training. The smaller model takes one-third of the time and energy during inference as the original, full model. “What makes these novel applications possible in our hardware,” Sharapov says, “Is that there’s a million cores in a very tight package, meaning that the cores have very low latency, high bandwidth interactions between them.”

Over just the past couple of years, supercomputing has accelerated into the exascale era—with the world’s most massive machines capable of performing over a billion billion operations per second. But unless big efficiency improvements can intervene along its exponential growth curve, computing is also anticipated to require increasingly impractical and unsustainable amounts of energy—even, according to one widely cited study, by 2040 demanding more energy than the world’s total present-day output.

Fortunately, the high-performance computing community is shifting focus now toward not just increased performance (measured in raw petaflops or exaflops) but also higher efficiency, boosting the number of operations per watt.

The Green500 list saw newcomers enter into the top three spots, suggesting that some of the world’s newest high-performance systems may be chasing efficiency at least as much as sheer power.

The newest ranking of the Top500 supercomputers (a list of the world’s most powerful machines) and its cousin the Green500 (ranking instead the world’s highest-efficiency machines) came out last week. The leading 10 of the Top 500 largest supercomputers remains mostly unchanged, headed up by Oak Ridge National Laboratory’s Frontier exascale computer. There was only one new addition in the top 10, at No. 6: Swiss National Supercomputing Center’s Alps system. Meanwhile, Argonne National Laboratory’s Aurora doubled its size, but kept its second-tier ranking.

On the other hand, The Green500 list saw newcomers enter into the top three spots, suggesting that some of the world’s newest high-performance systems may be chasing efficiency at least as much as sheer power.

Heading up the new Green500 list was JEDI, Jülich Supercomputing Center’s prototype system for its impending JUPITER exascale computer. The No. 2 and No. 3 spots went to the University of Bristol’s Isambard AI, also the first phase of a larger planned system, and the Helios supercomputer from the Polish organization Cyfronet. In fourth place is the previous list’s leader, the Simons Foundation’s Henri.

A Hopper Runs Through It

The top three systems on the Green500 list have one thing in common—they are all built with Nvidia’s Grace Hopper superchips, a combination of the Hopper (H100) GPU and the Grace CPU. There are two main reasons why the Grace Hopper architecture is so efficient, says Dion Harris, director of accelerated data center go-to-market strategy at Nvidia. The first is the Grace CPU, which benefits from the ARM instruction set architecture’s superior power performance. Plus, he says, it incorporates a memory structure, called LPDDR5X, that’s commonly found in cellphones and is optimized for energy efficiency.

Nvidia’s GH200 Grace Hopper superchip, here deployed in Jülich’s JEDI machine, now powers the world’s top three most efficient HPC systems. Jülich Supercomputing Center

The second advantage of the Grace Hopper, Harris says, is a newly developed interconnect between the Hopper GPU and the Grace CPU. The connection takes advantage of the CPU and GPU’s proximity to each other on one board, and achieves a bandwidth of 900 gigabits per second, about 7 times as fast as the latest PCIe gen5 interconnects. This allows the GPU to access the CPU’s memory quickly, which is particularly important for highly parallel applications such as AI training or graph neural networks, Harris says.

All three top systems use Grace Hoppers, but Jülich’s JEDI still leads the pack by a noticeable margin—72.7 gigaflops per watt, as opposed to 68.8 gigaflops per watt for the runner-up (and 65.4 gigaflops per watt for the previous champion). The JEDI team attributes their added success to the way they’ve connected their chips together. Their interconnect fabric was also from Nvidia—Quantum-2 InfiniBand—rather than the HPE Slingshot used by the other two top systems.

The JEDI team also cites specific optimizations they did to accommodate the Green500 benchmark. In addition to using all the latest Nvidia gear, JEDI cuts energy costs with its cooling system. Instead of using air or chilled water, JEDI circulates hot water throughout its compute nodes to take care of the excess heat. “Under normal weather conditions, the excess heat can be taken care of by free cooling units without the need of additional cold-water cooling,” says Benedikt von St. Vieth, head of the division for high-performance computing at Jülich.

JUPITER will use the same architecture as its prototype, JEDI, and von St. Vieth says he aims for it to maintain much of the prototype’s energy efficiency—although with increased scale, he adds, more energy may be lost to interconnecting fabric.

Of course, most crucial is the performance of these systems on real scientific tasks, not just on the Green500 benchmark. “It was really exciting to see these systems come online,” Nvidia’s Harris says, “But more importantly, I think we’re really excited to see the science come out of these systems, because I think [the energy efficiency] will have more impact on the applications even than on the benchmark.”

Today Dresden, Germany–based startup SpiNNcloud Systems announced that its hybrid supercomputing platform, the SpiNNcloud Platform, is available for sale. The machine combines traditional AI accelerators with neuromorphic computing capabilities, using system-design strategies that draw inspiration from the human brain. Systems for purchase vary in size, but the largest commercially available machine can simulate 10 billion neurons, about one-tenth the number in the human brain. The announcement was made at the ISC High Performance conference in Hamburg, Germany.

“We’re basically trying to bridge the gap between brain inspiration and artificial systems.” —Hector Gonzalez, SpiNNcloud Systems

SpiNNcloud Systems was founded in 2021 as a spin-off of the Dresden University of Technology. Its original chip, the SpiNNaker1, was designed by Steve Furber, the principal designer of the ARM microprocessor—the technology that now powers most cellphones. The SpiNNaker1 chip is already in use by 60 research groups in 23 countries, SpiNNcloud Systems says.

Human Brain as Supercomputer

Brain-emulating computers hold the promise of vastly lower energy computation and better performance on certain tasks. “The human brain is the most advanced supercomputer in the universe, and it consumes only 20 watts to achieve things that artificial intelligence systems today only dream of,” says Hector Gonzalez, cofounder and co-CEO of SpiNNcloud Systems. “We’re basically trying to bridge the gap between brain inspiration and artificial systems.”

Aside from sheer size, a distinguishing feature of the SpiNNaker2 system is its flexibility. Traditionally, most neuromorphic computers emulate the brain’s spiking nature: Neurons fire off electrical spikes to communicate with the neurons around them. The actual mechanism of these spikes in the brain is quite complex, and neuromorphic hardware often implements a specific simplified model. The SpiNNaker2 can implement a broad range of such models however, as they are not hardwired into its architecture.

Instead of looking how each neuron and synapse operates in the brain and trying to emulate that from the bottom up, Gonzalez says, the his team’s approach involved implementing key performance features of the brain. “It’s more about taking a practical inspiration from the brain, following particularly fascinating aspects such as how the brain is energy proportional and how it is simply highly parallel,” Gonzalez says.

To build hardware that is energy proportional—each piece draws power only when it’s actively in use and highly parallel—the company started with the building blocks. The basic unit of the system is the SpiNNaker2 chip, which hosts 152 processing units. Each processing unit has an ARM-based microcontroller, and unlike its predecessor the SpiNNaker1, also comes equipped with accelerators for use on neuromorphic models and traditional neural networks.

The SpiNNaker2 supercomputer has been designed to model up to 10 billion neurons, about one-tenth the number in the human brain. SpiNNCloud Systems

The processing units can operate in an event-based manner: They can stay off unless an event triggers them to turn on and operate. This enables energy-proportional operation. The events are routed between units and across chips asynchronously, meaning there is no central clock coordinating their movements—which can allow for massive parallelism. Each chip is connected to six other chips, and the whole system is connected in the shape of a torus to ensure all connecting wires are equally short.

The largest commercially offered system is not only capable of emulating 10 billion neurons, but also performing 0.3 billion billion operations per second (exaops) of more traditional AI tasks, putting it on a comparable scale with the top 10 largest supercomputers today.

Among the first customers of the SpiNNaker2 system is a team at Sandia National Labs, which plans to use it for further research on neuromorphic systems outperforming traditional architectures and performing otherwise inaccessible computational tasks.

“The ability to have a general programmable neuron model lets you explore some of these more complex learning rules that don’t necessarily fit onto older neuromorphic systems,” says Fred Rothganger, senior member of technical staff at Sandia. “They, of course, can run on a general-purpose computer. But those general-purpose computers are not necessarily designed to efficiently handle the kind of communication patterns that go on inside a spiking neural network. With [the SpiNNaker2 system] we get the ideal combination of greater programmability plus efficient communication.”

Dina Genkina: Hi, I’m Dina Genkina for IEEE Spectrum‘s Fixing the Future. Before we start, I want to tell you that you can get the latest coverage from some of Spectrum‘s most important beats, including AI, climate change, and robotics, by signing up for one of our free newsletters. Just go to spectrum.ieee.org/newsletters to subscribe. And today our guest on the show is Suraj Bramhavar. Recently, Bramhavar left his job as a co-founder and CTO of Sync Computing to start a new chapter. The UK government has just founded the Advanced Research Invention Agency, or ARIA, modeled after the US’s own DARPA funding agency. Bramhavar is heading up ARIA’s first program, which officially launched on March 12th of this year. Bramhavar’s program aims to develop new technology to make AI computation 1,000 times more cost efficient than it is today. Siraj, welcome to the show.

Suraj Bramhavar: Thanks for having me.

Genkina: So your program wants to reduce AI training costs by a factor of 1,000, which is pretty ambitious. Why did you choose to focus on this problem?

Bramhavar: So there’s a couple of reasons why. The first one is economical. I mean, AI is basically to become the primary economic driver of the entire computing industry. And to train a modern large-scale AI model costs somewhere between 10 million to 100 million pounds now. And AI is really unique in the sense that the capabilities grow with more computing power thrown at the problem. So there’s kind of no sign of those costs coming down anytime in the future. And so this has a number of knock-on effects. If I’m a world-class AI researcher, I basically have to choose whether I go work for a very large tech company that has the compute resources available for me to do my work or go raise 100 million pounds from some investor to be able to do cutting edge research. And this has a variety of effects. It dictates, first off, who gets to do the work and also what types of problems get addressed. So that’s the economic problem. And then separately, there’s a technological one, which is that all of this stuff that we call AI is built upon a very, very narrow set of algorithms and an even narrower set of hardware. And this has scaled phenomenally well. And we can probably continue to scale along kind of the known trajectories that we have. But it’s starting to show signs of strain. Like I just mentioned, there’s an economic strain, there’s an energy cost to all this. There’s logistical supply chain constraints. And we’re seeing this now with kind of the GPU crunch that you read about in the news.

And in some ways, the strength of the existing paradigm has kind of forced us to overlook a lot of possible alternative mechanisms that we could use to kind of perform similar computations. And this program is designed to kind of shine a light on those alternatives.

Genkina: Yeah, cool. So you seem to think that there’s potential for pretty impactful alternatives that are orders of magnitude better than what we have. So maybe we can dive into some specific ideas of what those are. And you talk about in your thesis that you wrote up for the start of this program, you talk about natural computing systems. So computing systems that take some inspiration from nature. So can you explain a little bit what you mean by that and what some of the examples of that are?

Bramhavar: Yeah. So when I say natural-based or nature-based computing, what I really mean is any computing system that either takes inspiration from nature to perform the computation or utilizes physics in a new and exciting way to perform computation. So you can think about kind of people have heard about neuromorphic computing. Neuromorphic computing fits into this category, right? It takes inspiration from nature and usually performs a computation in most cases using digital logic. But that represents a really small slice of the overall breadth of technologies that incorporate nature. And part of what we want to do is highlight some of those other possible technologies. So what do I mean when I say nature-based computing? I think we have a solicitation call out right now, which calls out a few things that we’re interested in. Things like new types of in-memory computing architectures, rethinking AI models from an energy context. And we also call out a couple of technologies that are pivotal for the overall system to function, but are not necessarily so eye-catching, like how you interconnect chips together, and how you simulate a large-scale system of any novel technology outside of the digital landscape. I think these are critical pieces to realizing the overall program goals. And we want to put some funding towards kind of boosting that workup as well.

Genkina: Okay, so you mentioned neuromorphic computing is a small part of the landscape that you’re aiming to explore here. But maybe let’s start with that. People may have heard of neuromorphic computing, but might not know exactly what it is. So can you give us the elevator pitch of neuromorphic computing?

Bramhavar: Yeah, my translation of neuromorphic computing— and this may differ from person to person, but my translation of it is when you kind of encode the information in a neural network via spikes rather than kind of discrete values. And that modality has shown to work pretty well in certain situations. So if I have some camera and I need a neural network next to that camera that can recognize an image with very, very low power or very, very high speed, neuromorphic systems have shown to work remarkably well. And they’ve worked in a variety of other applications as well. One of the things that I haven’t seen, or maybe one of the drawbacks of that technology that I think I would love to see someone solve for is being able to use that modality to train large-scale neural networks. So if people have ideas on how to use neuromorphic systems to train models at commercially relevant scales, we would love to hear about them and that they should submit to this program call, which is out.

Genkina: Is there a reason to expect that these kinds of— that neuromorphic computing might be a platform that promises these orders of magnitude cost improvements?

Bramhavar: I don’t know. I mean, I don’t know actually if neuromorphic computing is the right technological direction to realize that these types of orders of magnitude cost improvements. It might be, but I think we’ve intentionally kind of designed the program to encompass more than just that particular technological slice of the pie, in part because it’s entirely possible that that is not the right direction to go. And there are other more fruitful directions to put funding towards. Part of what we’re thinking about when we’re designing these programs is we don’t really want to be prescriptive about a specific technology, be it neuromorphic computing or probabilistic computing or any particular thing that has a name that you can attach to it. Part of what we tried to do is set a very specific goal or a problem that we want to solve. Put out a funding call and let the community kind of tell us which technologies they think can best meet that goal. And that’s the way we’ve been trying to operate with this program specifically. So there are particular technologies we’re kind of intrigued by, but I don’t think we have any one of them selected as like kind of this is the path forward.

Genkina: Cool. Yeah, so you’re kind of trying to see what architecture needs to happen to make computers as efficient as brains or closer to the brain’s efficiency.

Bramhavar: And you kind of see this happening in the AI algorithms world. As these models get bigger and bigger and grow their capabilities, they’re starting to introduce things that we see in nature all the time. I think probably the most relevant example is this stable diffusion, this neural network model where you can type in text and generate an image. It’s got diffusion in the name. Diffusion is a natural process. Noise is a core element of this algorithm. And so there’s lots of examples like this where they’ve kind of— that community is taking bits and pieces or inspiration from nature and implementing it into these artificial neural networks. But in doing that, they’re doing it incredibly inefficiently.

Genkina: Yeah. Okay, so great. So the idea is to take some of the efficiencies out in nature and kind of bring them into our technology. And I know you said you’re not prescribing any particular solution and you just want that general idea. But nevertheless, let’s talk about some particular solutions that have been worked on in the past because you’re not starting from zero and there are some ideas about how to do this. So I guess neuromorphic computing is one such idea. Another is this noise-based computing, something like probabilistic computing. Can you explain what that is?

Bramhavar: Noise is a very intriguing property? And there’s kind of two ways I’m thinking about noise. One is just how do we deal with it? When you’re designing a digital computer, you’re effectively designing noise out of your system, right? You’re trying to eliminate noise. And you go through great pains to do that. And as soon as you move away from digital logic into something a little bit more analog, you spend a lot of resources fighting noise. And in most cases, you eliminate any benefit that you get from your kind of newfangled technology because you have to fight this noise. But in the context of neural networks, what’s very interesting is that over time, we’ve kind of seen algorithms researchers discover that they actually didn’t need to be as precise as they thought they needed to be. You’re seeing the precision kind of come down over time. The precision requirements of these networks come down over time. And we really haven’t hit the limit there as far as I know. And so with that in mind, you start to ask the question, “Okay, how precise do we actually have to be with these types of computations to perform the computation effectively?” And if we don’t need to be as precise as we thought, can we rethink the types of hardware platforms that we use to perform the computations?

So that’s one angle is just how do we better handle noise? The other angle is how do we exploit noise? And so there’s kind of entire textbooks full of algorithms where randomness is a key feature. I’m not talking necessarily about neural networks only. I’m talking about all algorithms where randomness plays a key role. Neural networks are kind of one area where this is also important. I mean, the primary way we train neural networks is stochastic gradient descent. So noise is kind of baked in there. I talked about stable diffusion models like that where noise becomes a key central element. In almost all of these cases, all of these algorithms, noise is kind of implemented using some digital random number generator. And so there the thought process would be, “Is it possible to redesign our hardware to make better use of the noise, given that we’re using noisy hardware to start with?” Notionally, there should be some savings that come from that. That presumes that the interface between whatever novel hardware you have that is creating this noise, and the hardware you have that’s performing the computing doesn’t eat away all your gains, right? I think that’s kind of the big technological roadblock that I’d be keen to see solutions for, outside of the algorithmic piece, which is just how do you make efficient use of noise.

When you’re thinking about implementing it in hardware, it becomes very, very tricky to implement it in a way where whatever gains you think you had are actually realized at the full system level. And in some ways, we want the solutions to be very, very tricky. The agency is designed to fund very high risk, high reward type of activities. And so there in some ways shouldn’t be consensus around a specific technological approach. Otherwise, somebody else would have likely funded it.

Genkina: You’re already becoming British. You said you were keen on the solution.

Bramhavar: I’ve been here long enough.

Genkina: It’s showing. Great. Okay, so we talked a little bit about neuromorphic computing. We talked a little bit about noise. And you also mentioned some alternatives to backpropagation in your thesis. So maybe first, can you explain for those that might not be familiar what backpropagation is and why it might need to be changed?

Bramhavar: Yeah, so this algorithm is essentially the bedrock of all AI training currently you use today. Essentially, what you’re doing is you have this large neural network. The neural network is composed of— you can think about it as this long chain of knobs. And you really have to tune all the knobs just right in order to get this network to perform a specific task, like when you give it an image of a cat, it says that it is a cat. And so what backpropagation allows you to do is to tune those knobs in a very, very efficient way. Starting from the end of your network, you kind of tune the knob a little bit, see if your answer gets a little bit closer to what you’d expect it to be. Use that information to then tune the knobs in the previous layer of your network and keep on doing that iteratively. And if you do this over and over again, you can eventually find all the right positions of your knobs such that your network does whatever you’re trying to do. And so this is great. Now, the issue is every time you tune one of these knobs, you’re performing this massive mathematical computation. And you’re typically doing that across many, many GPUs. And you do that just to tweak the knob a little bit. And so you have to do it over and over and over and over again to get the knobs where you need to go.

There’s a whole bevy of algorithms. What you’re really doing is kind of minimizing error between what you want the network to do and what it’s actually doing. And if you think about it along those terms, there’s a whole bevy of algorithms in the literature that kind of minimize energy or error in that way. None of them work as well as backpropagation. In some ways, the algorithm is beautiful and extraordinarily simple. And most importantly, it’s very, very well suited to be parallelized on GPUs. And I think that is part of its success. But one of the things I think both algorithmic researchers and hardware researchers fall victim to is this chicken and egg problem, right? Algorithms researchers build algorithms that work well on the hardware platforms that they have available to them. And at the same time, hardware researchers develop hardware for the existing algorithms of the day. And so one of the things we want to try to do with this program is blend those worlds and allow algorithms researchers to think about what is the field of algorithms that I could explore if I could rethink some of the bottlenecks in the hardware that I have available to me. Similarly in the opposite direction.

Genkina: Imagine that you succeeded at your goal and the program and the wider community came up with a 1/1000s compute cost architecture, both hardware and software together. What does your gut say that that would look like? Just an example. I know you don’t know what’s going to come out of this, but give us a vision.

Bramhavar: Similarly, like I said, I don’t think I can prescribe a specific technology. What I can say is that— I can say with pretty high confidence, it’s not going to just be one particular technological kind of pinch point that gets unlocked. It’s going to be a systems level thing. So there may be individual technology at the chip level or the hardware level. Those technologies then also have to meld with things at the systems level as well and the algorithms level as well. And I think all of those are going to be necessary in order to reach these goals. I’m talking kind of generally, but what I really mean is like what I said before is we got to think about new types of hardware. We also have to think about, “Okay, if we’re going to scale these things and manufacture them in large volumes cost effectively, we’re going to have to build larger systems out of building blocks of these things. So we’re going to have to think about how to stitch them together in a way that makes sense and doesn’t eat away any of the benefits. We’re also going to have to think about how to simulate the behavior of these things before we build them.” I think part of the power of the digital electronics ecosystem comes from the fact that you have cadence and synopsis and these EDA platforms that allow you with very high accuracy to predict how your circuits are going to perform before you build them. And once you get out of that ecosystem, you don’t really have that.

So I think it’s going to take all of these things in order to actually reach these goals. And I think part of what this program is designed to do is kind of change the conversation around what is possible. So by the end of this, it’s a four-year program. We want to show that there is a viable path towards this end goal. And that viable path could incorporate kind of all of these aspects of what I just mentioned.

Genkina: Okay. So the program is four years, but you don’t necessarily expect like a finished product of a 1/1000s cost computer by the end of the four years, right? You kind of just expect to develop a path towards it.

Bramhavar: Yeah. I mean, ARIA was kind of set up with this kind of decadal time horizon. We want to push out-- we want to fund, as I mentioned, high-risk, high reward technologies. We have this kind of long time horizon to think about these things. I think the program is designed around four years in order to kind of shift the window of what the world thinks is possible in that timeframe. And in the hopes that we change the conversation. Other folks will pick up this work at the end of that four years, and it will have this kind of large-scale impact on a decadal.

Genkina: Great. Well, thank you so much for coming today. Today we spoke with Dr. Suraj Bramhavar, lead of the first program headed up by the UK’s newest funding agency, ARIA. He filled us in on his plans to reduce AI costs by a factor of 1,000, and we’ll have to check back with him in a few years to see what progress has been made towards this grand vision. For IEEE Spectrum, I’m Dina Genkina, and I hope you’ll join us next time on Fixing the Future.

Since ChatGPT dropped in the fall of 2022, everyone and their donkey has tried their hand at prompt engineering—finding a clever way to phrase your query to a large language model (LLM) or AI art or video generator to get the best results or sidestep protections. The Internet is replete with prompt-engineering guides, cheat sheets, and advice threads to help you get the most out of an LLM.

In the commercial sector, companies are now wrangling LLMs to build product copilots, automate tedious work, create personal assistants, and more, says Austin Henley, a former Microsoft employee who conducted a series of interviews with people developing LLM-powered copilots. “Every business is trying to use it for virtually every use case that they can imagine,” Henley says.

“The only real trend may be no trend. What’s best for any given model, dataset, and prompting strategy is likely to be specific to the particular combination at hand.” —Rick Battle & Teja Gollapudi, VMware

To do so, they’ve enlisted the help of prompt engineers professionally.

However, new research suggests that prompt engineering is best done by the model itself, and not by a human engineer. This has cast doubt on prompt engineering’s future—and increased suspicions that a fair portion of prompt-engineering jobs may be a passing fad, at least as the field is currently imagined.

Autotuned prompts are successful and strange

Rick Battle and Teja Gollapudi at California-based cloud computing company VMware were perplexed by how finicky and unpredictable LLM performance was in response to weird prompting techniques. For example, people have found that asking models to explain its reasoning step-by-step—a technique called chain-of-thought—improved their performance on a range of math and logic questions. Even weirder, Battle found that giving a model positive prompts, such as “this will be fun” or “you are as smart as chatGPT,” sometimes improved performance.

Battle and Gollapudi decided to systematically test how different prompt-engineering strategies impact an LLM’s ability to solve grade-school math questions. They tested three different open-source language models with 60 different prompt combinations each. What they found was a surprising lack of consistency. Even chain-of-thought prompting sometimes helped and other times hurt performance. “The only real trend may be no trend,” they write. “What’s best for any given model, dataset, and prompting strategy is likely to be specific to the particular combination at hand.”

According to one research team, no human should manually optimize prompts ever again.

There is an alternative to the trial-and-error-style prompt engineering that yielded such inconsistent results: Ask the language model to devise its own optimal prompt. Recently, new tools have been developed to automate this process. Given a few examples and a quantitative success metric, these tools will iteratively find the optimal phrase to feed into the LLM. Battle and his collaborators found that in almost every case, this automatically generated prompt did better than the best prompt found through trial-and-error. And, the process was much faster, a couple of hours rather than several days of searching.

The optimal prompts the algorithm spit out were so bizarre, no human is likely to have ever come up with them. “I literally could not believe some of the stuff that it generated,” Battle says. In one instance, the prompt was just an extended Star Trek reference: “Command, we need you to plot a course through this turbulence and locate the source of the anomaly. Use all available data and your expertise to guide us through this challenging situation.” Apparently, thinking it was Captain Kirk helped this particular LLM do better on grade-school math questions.

Battle says that optimizing the prompts algorithmically fundamentally makes sense given what language models really are—models. “A lot of people anthropomorphize these things because they ‘speak English.’ No, they don’t,” Battle says. “It doesn’t speak English. It does a lot of math.”

In fact, in light of his team’s results, Battle says no human should manually optimize prompts ever again.

“You’re just sitting there trying to figure out what special magic combination of words will give you the best possible performance for your task,” Battle says, “But that’s where hopefully this research will come in and say ‘don’t bother.’ Just develop a scoring metric so that the system itself can tell whether one prompt is better than another, and then just let the model optimize itself.”

Autotuned prompts make pictures prettier, too

Image-generation algorithms can benefit from automatically generated prompts as well. Recently, a team at Intel labs, led by Vasudev Lal, set out on a similar quest to optimize prompts for the image-generation model Stable Diffusion. “It seems more like a bug of LLMs and diffusion models, not a feature, that you have to do this expert prompt engineering,” Lal says. “So, we wanted to see if we can automate this kind of prompt engineering.”

“Now we have this full machinery, the full loop that’s completed with this reinforcement learning.… This is why we are able to outperform human prompt engineering.” —Vasudev Lal, Intel Labs

Lal’s team created a tool called NeuroPrompts that takes a simple input prompt, such as “boy on a horse,” and automatically enhances it to produce a better picture. To do this, they started with a range of prompts generated by human prompt-engineering experts. They then trained a language model to transform simple prompts into these expert-level prompts. On top of that, they used reinforcement learning to optimize these prompts to create more aesthetically pleasing images, as rated by yet another machine-learning model, PickScore, a recently developed image-evaluation tool.

NeuroPrompts is a generative AI auto prompt tuner that transforms simple prompts into more detailed and visually stunning StableDiffusion results—as in this case, an image generated by a generic prompt [left] versus its equivalent NeuroPrompt-generated image.Intel Labs/Stable Diffusion

Here too, the automatically generated prompts did better than the expert-human prompts they used as a starting point, at least according to the PickScore metric. Lal found this unsurprising. “Humans will only do it with trial and error,” Lal says. “But now we have this full machinery, the full loop that’s completed with this reinforcement learning.… This is why we are able to outperform human prompt engineering.”

Since aesthetic quality is infamously subjective, Lal and his team wanted to give the user some control over how the prompt was optimized. In their tool, the user can specify the original prompt (say, “boy on a horse”) as well as an artist to emulate, a style, a format, and other modifiers.

Lal believes that as generative AI models evolve, be it image generators or large language models, the weird quirks of prompt dependence should go away. “I think it’s important that these kinds of optimizations are investigated and then ultimately, they’re really incorporated into the base model itself so that you don’t really need a complicated prompt-engineering step.”

Prompt engineering will live on, by some name

Even if autotuning prompts becomes the industry norm, prompt-engineering jobs in some form are not going away, says Tim Cramer, senior vice president of software engineering at Red Hat. Adapting generative AI for industry needs is a complicated, multistage endeavor that will continue requiring humans in the loop for the foreseeable future.

“Maybe we’re calling them prompt engineers today. But I think the nature of that interaction will just keep on changing as AI models also keep changing.” —Vasudev Lal, Intel Labs

“I think there are going to be prompt engineers for quite some time, and data scientists,” Cramer says. “It’s not just asking questions of the LLM and making sure that the answer looks good. But there’s a raft of things that prompt engineers really need to be able to do.”

“It’s very easy to make a prototype,” Henley says. “It’s very hard to production-ize it.” Prompt engineering seems like a big piece of the puzzle when you’re building a prototype, Henley says, but many other considerations come into play when you’re making a commercial-grade product.

Challenges of making a commercial product include ensuring reliability—for example, failing gracefully when the model goes offline; adapting the model’s output to the appropriate format, since many use cases require outputs other than text; testing to make sure the AI assistant won’t do something harmful in even a small number of cases; and ensuring safety, privacy, and compliance. Testing and compliance are particularly difficult, Henley says, as traditional software-development testing strategies are maladapted for nondeterministic LLMs.

To fulfill these myriad tasks, many large companies are heralding a new job title: Large Language Model Operations, or LLMOps, which includes prompt engineering in its life cycle but also entails all the other tasks needed to deploy the product. Henley says LLMOps’ predecessors, machine learning operations (MLOps) engineers, are best positioned to take on these jobs.

Whether the job titles will be “prompt engineer,” “LLMOps engineer,” or something new entirely, the nature of the job will continue evolving quickly. “Maybe we’re calling them prompt engineers today,” Lal says, “But I think the nature of that interaction will just keep on changing as AI models also keep changing.”

“I don’t know if we’re going to combine it with another sort of job category or job role,” Cramer says, “But I don’t think that these things are going to be going away anytime soon. And the landscape is just too crazy right now. Everything’s changing so much. We’re not going to figure it all out in a few months.”

Henley says that, to some extent in this early phase of the field, the only overriding rule seems to be the absence of rules. “It’s kind of the Wild, Wild West for this right now.” he says.

Dina Genkina: Hi. I’m Dina Genkina for IEEE Spectrum‘s Fixing the Future. Before we start, I want to tell you that you can get the latest coverage from some of Spectrum’s most important beeps, including AI, Change, and Robotics, by signing up for one of our free newsletters. Just go to spectrum.ieee.org\newsletters to subscribe. Today, a guest is Dr. Benji Maruyama, a Principal Materials Research Engineer at the Air Force Research Laboratory, or AFRL. Dr. Maruyama is a materials scientist, and his research focuses on carbon nanotubes and making research go faster. But he’s also a man with a dream, a dream of a world where science isn’t something done by a select few locked away in an ivory tower, but something most people can participate in. He hopes to start what he calls the billion scientist movement by building AI-enabled research robots that are accessible to all. Benji, thank you for coming on the show.

Benji Maruyama: Thanks, Dina. Great to be with you. I appreciate the invitation.

Genkina: Yeah. So let’s set the scene a little bit for our listeners. So you advocate for this billion scientist movement. If everything works amazingly, what would this look like? Paint us a picture of how AI will help us get there.

Maruyama: Right, great. Thanks. Yeah. So one of the things as you set the scene there is right now, to be a scientist, most people need to have access to a big lab with very expensive equipment. So I think top universities, government labs, industry folks, lots of equipment. It’s like a million dollars, right, to get one of them. And frankly, just not that many of us have access to those kinds of instruments. But at the same time, there’s probably a lot of us who want to do science, right? And so how do we make it so that anyone who wants to do science can try, can have access to instruments so that they can contribute to it. So that’s the basics behind citizen science or democratization of science so that everyone can do it. And one way to think of it is what happened with 3D printing. It used to be that in order to make something, you had to have access to a machine shop or maybe get fancy tools and dyes that could cost tens of thousands of dollars a pop. Or if you wanted to do electronics, you had to have access to very expensive equipment or services. But when 3D printers came along and became very inexpensive, all of a sudden now, anyone with access to a 3D printer, so maybe in a school or a library or a makerspace could print something out. And it could be something fun, like a game piece, but it could also be something that got you to an invention, something that was maybe useful to the community, was either a prototype or an actual working device.

And so really, 3D printing democratized manufacturing, right? It made it so that many more of us could do things that before only a select few could. And so that’s where we’re trying to go with science now, is that instead of only those of us who have access to big labs, we’re building research robots. And when I say we, we’re doing it, but now there are a lot of others who are doing it as well, and I’ll get into that. But the example that we have is that we took a 3D printer that you can buy off the internet for less than $300. Plus a couple of extra parts, a webcam, a Raspberry Pi board, and a tripod really, so only four components. You can get them all for $300. Load them with open-source software that was developed by AFIT, the Air Force Institute of Technology. So Burt Peterson and Greg Captain [inaudible]. We worked together to build this fully autonomous 3D printing robot that taught itself how to print to better than manufacturer’s specifications. So that was a really fun advance for us, and now we’re trying to take that same idea and broaden it. So I’ll turn it back over to you.

Genkina: Yeah, okay. So maybe let’s talk a little bit about this automated research robot that you’ve made. So right now, it works with a 3D printer, but is the big picture that one day it’s going to give people access to that million dollar lab? How would that look like?

Maruyama: Right, so there are different models out there. One, we just did a workshop at the University of— sorry, North Carolina State University about that very problem, right? So there’s two models. One is to get low-cost scientific tools like the 3D printer. There’s a couple of different chemistry robots, one out of University of Maryland and NIST, one out of University of Washington that are in the sort of 300 to 1,000 dollars range that makes it accessible. The other part is kind of the user facility model. So in the US, the Department of Energy National Labs have many user facilities where you can apply to get time on very expensive instruments. Now we’re talking tens of millions. For example, Brookhaven has a synchrotron light source where you can sign up and it doesn’t cost you any money to use the facility. And you can get days on that facility. And so that’s already there, but now the advances are that by using this, autonomy, autonomous closed loop experimentation, that the work that you do will be much faster and much more productive. So, for example, on ARES, our Autonomous Research System at AFRL, we actually were able to do experiments so fast that a professor who came into my lab said, it just took me aside and said, “Hey, Benji, in a week’s worth of time, I did a dissertation’s worth of research.” So maybe five years worth of research in a week. So imagine if you keep doing that week after week after week, how fast research goes. So it’s very exciting.

Genkina: Yeah, so tell us a little bit about how that works. So what’s this system that has sped up five years of research into a week and made graduate students obsolete? Not yet, not yet. How does that work? Is that the 3D printer system or is that a—

Maruyama: So we started with our system to grow carbon nanotubes. And I’ll say, actually, when we first thought about it, your comment about graduate students being absolute— obsolete, sorry, is interesting and important because, when we first built our system that worked it 100 times faster than normal, I thought that might be the case. We called it sort of graduate student out of the loop. But when I started talking with people who specialize in autonomy, it’s actually the opposite, right? It’s actually empowering graduate students to go faster and also to do the work that they want to do, right? And so just to digress a little bit, if you think about farmers before the Industrial Revolution, what were they doing? They were plowing fields with oxen and beasts of burden and hand plows. And it was hard work. And now, of course, you wouldn’t ask a farmer today to give up their tractor or their combine harvester, right? They would say, of course not. So very soon, we expect it to be the same for researchers, that if you asked a graduate student to give up their autonomous research robot five years from now, they’ll say, “Are you crazy? This is how I get my work done.”

But for our original ARES system, it worked on the synthesis of carbon nanotubes. So that meant that what we’re doing is trying to take this system that’s been pretty well studied, but we haven’t figured out how to make it at scale. So at hundreds of millions of tons per year, sort of like polyethylene production. And part of that is because it’s slow, right? One experiment takes a day, but also because there are just so many different ways to do a reaction, so many different combinations of temperature and pressure and a dozen different gases and half the periodic table as far as the catalyst. It’s just too much to just brute force your way through. So even though we went from experiments where we could do 100 experiments a day instead of one experiment a day, just that combinatorial space was vastly overwhelmed our ability to do it, even with many research robots or many graduate students. So the idea of having artificial intelligence algorithms that drive the research is key. And so that ability to do an experiment, see what happened, and then analyze it, iterate, and constantly be able to choose the optimal next best experiment to do is where ARES really shines. And so that’s what we did. ARES taught itself how to grow carbon nanotubes at controlled rates. And we were the first ones to do that for material science in our 2016 publication.

Genkina: That’s very exciting. So maybe we can peer under the hood a little bit of this AI model. How does the magic work? How does it pick the next best point to take and why it’s better than you could do as a graduate student or researcher?

Maruyama: Yeah, and so I think it’s interesting, right? In science, a lot of times we’re taught to hold everything constant, change one variable at a time, search over that entire space, see what happened, and then go back and try something else, right? So we reduce it to one variable at a time. It’s a reductionist approach. And that’s worked really well, but a lot of the problems that we want to go after are simply too complex for that reductionist approach. And so the benefit of being able to use artificial intelligence is that high dimensionality is no problem, right? Tens of dimensions search over very complex high-dimensional parameter space, which is overwhelming to humans, right? Is just basically bread and butter for AI. The other part to it is the iterative part. The beauty of doing autonomous experimentation is that you’re constantly iterating. You’re constantly learning over what just happened. You might also say, well, not only do I know what happened experimentally, but I have other sources of prior knowledge, right? So for example, ideal gas law says that this should happen, right? Or Gibbs phase rule might say, this can happen or this can’t happen. So you can use that prior knowledge to say, “Okay, I’m not going to do those experiments because that’s not going to work. I’m going to try here because this has the best chance of working.”

And within that, there are many different machine learning or artificial intelligence algorithms. Bayesian optimization is a popular one to help you choose what experiment is best. There’s also new AI that people are trying to develop to get better search.

Genkina: Cool. And so the software part of this autonomous robot is available for anyone to download, which is also really exciting. So what would someone need to do to be able to use that? Do they need to get a 3D printer and a Raspberry Pi and set it up? And what would they be able to do with it? Can they just build carbon nanotubes or can they do more stuff?

Maruyama: Right. So what we did, we built ARES OS, which is our open source software, and we’ll make sure to get you the GitHub link so that anyone can download it. And the idea behind ARES OS is that it provides a software framework for anyone to build their own autonomous research robot. And so the 3D printing example will be out there soon. But it’s the starting point. Of course, if you want to build your own new kind of robot, you still have to do the software development, for example, to link the ARES framework, the core, if you will, to your particular hardware, maybe your particular camera or 3D printer, or pipetting robot, or spectrometer, whatever that is. We have examples out there and we’re hoping to get to a point where it becomes much more user-friendly. So having direct Python connects so that you don’t— currently it’s programmed in C#. But to make it more accessible, we’d like it to be set up so that if you can do Python, you can probably have good success in building your own research robot.

Genkina: Cool. And you’re also working on a educational version of this, I understand. So what’s the status of that and what’s different about that version?

Maruyama: Yeah, right. So the educational version is going to be-- its sort of composition of a combination of hardware and software. So what we’re starting with is a low-cost 3D printer. And we’re collaborating now with the University at Buffalo, Materials Design Innovation Department. And we’re hoping to build up a robot based on a 3D printer. And we’ll see how it goes. It’s still evolving. But for example, it could be based on this very inexpensive $200 3D printer. It’s an Ender 3D printer. There’s another printer out there that’s based on University of Washington’s Jubilee printer. And that’s a very exciting development as well. So professors Lilo Pozzo and Nadya Peek at the University of Washington built this Jubilee robot with that idea of accessibility in mind. And so combining our ARES OS software with their Jubilee robot hardware is something that I’m very excited about and hope to be able to move forward on.

Genkina: What’s this Jubilee 3D printer? How is it different from a regular 3D printer?

Maruyama: It’s very open source. Not all 3D printers are open source and it’s based on a gantry system with interchangeable heads. So for example, you can get not just a 3D printing head, but other heads that might do things like do indentation, see how stiff something is, or maybe put a camera on there that can move around. And so it’s the flexibility of being able to pick different heads dynamically that I think makes it super useful. For the software, right, we have to have a good, accessible, user-friendly graphical user interface, a GUI. That takes time and effort, so we want to work on that. But again, that’s just the hardware software. Really to make ARES a good educational platform, we need to make it so that a teacher who’s interested can have the lowest activation barrier possible, right? We want she or he to be able to pull a lesson plan off of the internet, have supporting YouTube videos, and actually have the material that is a fully developed curriculum that’s mapped against state standards.

So that, right now, if you’re a teacher who— let’s face it, teachers are already overwhelmed with all that they have to do, putting something like this into their curriculum can be a lot of work, especially if you have to think about, well, I’m going to take all this time, but I also have to meet all of my teaching standards, all the state curriculum standards. And so if we build that out so that it’s a matter of just looking at the curriculum and just checking off the boxes of what state standards it maps to, then that makes it that much easier for the teacher to teach.

Genkina: Great. And what do you think is the timeline? Do you expect to be able to do this sometime in the coming year?

Maruyama: That’s right. These things always take longer than hoped for than expected, but we’re hoping to do it within this calendar year and very excited to get it going. And I would say for your listeners, if you’re interested in working together, please let me know. We’re very excited about trying to involve as many people as we can.

Genkina: Great. Okay, so you have the educational version, and you have the more research geared version, and you’re working on making this educational version more accessible. Is there something with the research version that you’re working on next, how you’re hoping to upgrade it, or is there something you’re using it for right now that you’re excited about?

There’s a number of things that we are very excited about the possibility of carbon nanotubes being produced at very large scale. So right now, people may remember carbon nanotubes as that great material that sort of never made it and was very overhyped. But there’s a core group of us who are still working on it because of the important promise of that material. So it’s material that is super strong, stiff, lightweight, electrically conductive. Much better than silicon as a digital electronics compute material. All of those great things, except we’re not making it at large enough scale. It’s actually used pretty significantly in lithium-ion batteries. It’s an important application. But other than that, it’s sort of like where’s my flying car? It’s never panned out. But there’s, as I said, a group of us who are working to really produce carbon nanotubes at much larger scale. So large scale for nanotubes now is sort of in the kilogram or ton scale. But what we need to get to is hundreds of millions of tons per year production rates. And why is that? Well, there’s a great effort that came out of ARPA-E. So the Department of Energy Advanced Research Projects Agency and the E is for Energy in that case.

So they funded a collaboration between Shell Oil and Rice University to pyrolyze methane, so natural gas into hydrogen for the hydrogen economy. So now that’s a clean burning fuel plus carbon. And instead of burning the carbon to CO2, which is what we now do, right? We just take natural gas and feed it through a turbine and generate electric power instead of— and that, by the way, generates so much CO2 that it’s causing global climate change. So if we can do that pyrolysis at scale, at hundreds of millions of tons per year, it’s literally a save the world proposition, meaning that we can avoid so much CO2 emissions that we can reduce global CO2 emissions by 20 to 40 percent. And that is the save the world proposition. It’s a huge undertaking, right? That’s a big problem to tackle, starting with the science. We still don’t have the science to efficiently and effectively make carbon nanotubes at that scale. And then, of course, we have to take the material and turn it into useful products. So the batteries is the first example, but thinking about replacing copper for electrical wire, replacing steel for structural materials, aluminum, all those kinds of applications. But we can’t do it. We can’t even get to that kind of development because we haven’t been able to make the carbon nanotubes at sufficient scale.

So I would say that’s something that I’m working on now that I’m very excited about and trying to get there, but it’s going to take some good developments in our research robots and some very smart people to get us there.

Genkina: Yeah, it seems so counterintuitive that making everything out of carbon is good for lowering carbon emissions, but I guess that’s the break.

Maruyama: Yeah, it is interesting, right? So people talk about carbon emissions, but really, the molecule that’s causing global warming is carbon dioxide, CO2, which you get from burning carbon. And so if you take that methane and parallelize it to carbon nanotubes, that carbon is now sequestered, right? It’s not going off as CO2. It’s staying in solid state. And not only is it just not going up into the atmosphere, but now we’re using it to replace steel, for example, which, by the way, steel, aluminum, copper production, all of those things emit lots of CO2 in their production, right? They’re energy intensive as a material production. So it’s kind of ironic.

Genkina: Okay, and are there any other research robots that you’re excited about that you think are also contributing to this democratization of science process?

Maruyama: Yeah, so we talked about Jubilee, the NIST robot, which is from Professor Ichiro Takeuchi at Maryland and Gilad Kusne at NIST, National Institute of Standards and Technology. Theirs is fun too. It’s LEGO as. So it’s actually based on a LEGO robotics platform. So it’s an actual chemistry robot built out of Legos. So I think that’s fun as well. And you can imagine, just like we have LEGO robot competitions, we can have autonomous research robot competitions where we try and do research through these robots or competitions where everybody sort of starts with the same robot, just like with LEGO robotics. So that’s fun as well. But I would say there’s a growing number of people doing these kinds of, first of all, low-cost science, accessible science, but in particular low-cost autonomous experimentation.

Genkina: So how far are we from a world where a high school student has an idea and they can just go and carry it out on some autonomous research system at some high-end lab?

Maruyama: That’s a really good question. I hope that it’s going to be in 5 to 10 years, that it becomes reasonably commonplace. But it’s going to take still some significant investment to get this going. And so we’ll see how that goes. But I don’t think there are any scientific impediments to getting this done. There is a significant amount of engineering to be done. And sometimes we hear, oh, it’s just engineering. The engineering is a significant problem. And it’s work to get some of these things accessible, low cost. But there are lots of great efforts. There are people who have used CDs, compact discs to make spectrometers out of. There are lots of good examples of citizen science out there. But it’s, I think, at this point, going to take investment in software, in hardware to make it accessible, and then importantly, getting students really up to speed on what AI is and how it works and how it can help them. And so I think it’s actually really important. So again, that’s the democratization of science is if we can make it available to everyone and accessible, then that helps people, everyone contribute to science. And I do believe that there are important contributions to be made by ordinary citizens, by people who aren’t you know PhDs working in a lab.

And I think there’s a lot of science out there to be done. If you ask working scientists, almost no one has run out of ideas or things they want to work on. There’s many more scientific problems to work on than we have the time where people are funding to work on. And so if we make science cheaper to do, then all of a sudden, more people can do science. And so those questions start to be resolved. And so I think that’s super important. And now we have, instead of, just those of us who work in big labs, you have millions, tens of millions, up to a billion people, that’s the billion scientist idea, who are contributing to the scientific community. And that, to me, is so powerful that many more of us can contribute than just the few of us who do it right now.

Genkina: Okay, that’s a great place to end on, I think. So, today we spoke to Dr. Benji Maruyama, a material scientist at AFRL, about his efforts to democratize scientific discovery through automated research robots. For IEEE Spectrum, I’m Dina Genkina, and I hope you’ll join us next time on Fixing the Future.

The first commercial silicon chip that includes open-source, built-in hardware security was announced today by the OpenTitan coalition.

This milestone represents another step in the growth of the open hardware movement. Open hardware has been gaining steam since the development of the popular open-source processor architecture RISC-V.

RISC-V gives an openly available prescription for how a computer can operate efficiently at the most basic level. OpenTitan goes beyond RISC-V’s open-source instruction set by delivering an open-source design for the silicon itself. Although other open-source silicon has been developed, this is the first one to include the design-verification stage and to produce a fully functional commercial chip, the coalition claims.

Utilizing a RISC-V based processor core, the chip, called Earl Grey, includes a number of built-in hardware security and cryptography modules, all working together in a self-contained microprocessor. The project began back in 2019 by a coalition of companies, started by Google and shepherded by the nonprofit lowRISC in Cambridge, United Kingdom. Modeled after open-source software projects, it has been developed by contributors from around the world, both official affiliates with the project and independent coders. Today’s announcement is the culmination of five years of work.

Open source “just takes over because it has certain valuable properties... I think we’re seeing the beginning of this now with silicon.”—Dominic Rizzo, zeroRISC

“This chip is very, very exciting,” says OpenTitan cocreator and CEO of coalition partner zeroRISC Dominic Rizzo. “But there’s a much bigger thing here, which is the development of this whole new type of methodology. Instead of a traditional…command and control style structure, this is distributed.”

The methodology they have developed is called Silicon Commons. Open-source hardware design faces challenges that open-source software didn’t, such as greater costs, a smaller professional community, and inability to supply bug fixes in patches after the product is released, explains lowRISC CEO Gavin Ferris. The Silicon Commons framework provides rules for documentation, predefined interfaces, and quality standards, as well as the governance structure laying out how the different partners make decisions as a collective.

Another key to the success of the project, Ferris says, was picking a problem that all the partners would have an incentive to continue participating in over the course of the five years of development. Hardware security was the right fit for the job because of its commercial importance as well as its particular fit to the open-source model. There’s a notion in cryptography known as Kerckhoffs’s principle, which states that the only thing that should actually be secret in a cryptosystem is the secret key itself. Open-sourcing the entire protocol makes sure the cryptosystem conforms to this rule.

What Is a Hardware Root-of-Trust?

OpenTitan uses a hardware security protocol known as a root of trust (RoT). The idea is to provide an on-chip source of cryptographic keys that is inaccessible remotely. Because it’s otherwise inaccessible, the system can trust that it hasn’t been tampered with, providing a basis to build security on. “Root of Trust means that at the end of the day, there is something that we both believe in,” explains Ravi Subrahmanyan, senior director of integrated circuit design at Analog Devices, who was not involved in the effort. Once there is something both people agree on, a trusted secure connection can be established.

Conventional, proprietary chips can also leverage RoT technology. Open-sourcing it provides an extra layer of trust, proponents argue. Since anyone can inspect and probe the design, the theory is that bugs are more likely to get noticed and the bug fixes can be verified. “The openness is a good thing,” says Subrahmanyan. “Because for example, let’s say a proprietary implementation has some problem. I won’t necessarily know, right? I’m at their mercy as to whether they’re going to tell me or not.”

This kind of on-chip security is especially relevant in devices forming the Internet of Things (IoT), which suffer from unaddressed security challenges. ZeroRISC and its partners will open up sales to IoT markets via an early-access program, and they anticipate broad adoption in that sphere.

Rizzo and Ferris believe their chip has a template for open-source hardware development that other collaborations will replicate. On top of providing transparent security, open-sourcing saves companies money by allowing them to reuse hardware components rather than having to independently develop proprietary versions of the same thing. It also opens the door for many more partners to participate in the effort, including academic institutions such as OpenTitan coalition partner ETH Zurich. Thanks to academic involvement, OpenTitan was able to incorporate cryptography protocols that are safe against future quantum computers.

“Once the methodology has been proven, others will pick it up,” Rizzo says. “If you look at what’s happened with open-source software, first, people thought it was kind of an edge pursuit, and then it ended up running almost every mobile phone. It just takes over because it has certain valuable properties. And so I think we’re seeing the beginning of this now with silicon.”