How to build a better AI benchmark

Stay Ahead, Stay ONMINE

How to build a better AI benchmark

It’s not easy being one of Silicon Valley’s favorite benchmarks. SWE-Bench (pronounced “swee bench”) launched in November 2024 to evaluate an AI model’s coding skill, using more than 2,000 real-world programming problems pulled from the public GitHub repositories of 12 different Python-based projects. In the months since then, it’s quickly become one of the most popular tests in AI. A SWE-Bench score has become a mainstay of major model releases from OpenAI, Anthropic, and Google—and outside of foundation models, the fine-tuners at AI firms are in constant competition to see who can rise above the pack. The top of the leaderboard is a pileup between three different fine tunings of Anthropic’s Claude Sonnet model and Amazon’s Q developer agent. Auto Code Rover—one of the Claude modifications—nabbed the number two spot in November, and was acquired just three months later. Despite all the fervor, this isn’t exactly a truthful assessment of which model is “better.” As the benchmark has gained prominence, “you start to see that people really want that top spot,” says John Yang, a researcher on the team that developed SWE-Bench at Princeton University. As a result, entrants have begun to game the system—which is pushing many others to wonder whether there’s a better way to actually measure AI achievement. Developers of these coding agents aren’t necessarily doing anything as straightforward cheating, but they’re crafting approaches that are too neatly tailored to the specifics of the benchmark. The initial SWE-Bench test set was limited to programs written in Python, which meant developers could gain an advantage by training their models exclusively on Python code. Soon, Yang noticed that high-scoring models would fail completely when tested on different programming languages—revealing an approach to the test that he describes as “gilded.” “It looks nice and shiny at first glance, but then you try to run it on a different language and the whole thing just kind of falls apart,” Yang says. “At that point, you’re not designing a software engineering agent. You’re designing to make a SWE-Bench agent, which is much less interesting.” The SWE-Bench issue is a symptom of a more sweeping—and complicated—problem in AI evaluation, and one that’s increasingly sparking heated debate: The benchmarks the industry uses to guide development are drifting further and further away from evaluating actual capabilities, calling their basic value into question. Making the situation worse, several benchmarks, most notably FrontierMath and Chatbot Arena, have recently come under heat for an alleged lack of transparency. Nevertheless, benchmarks still play a central role in model development, even if few experts are willing to take their results at face value. OpenAI cofounder Andrej Karpathy recently described the situation as “an evaluation crisis”: the industry has fewer trusted methods for measuring capabilities and no clear path to better ones. “Historically, benchmarks were the way we evaluated AI systems,” says Vanessa Parli, director of research at Stanford University’s Institute for Human-Centered AI. “Is that the way we want to evaluate systems going forward? And if it’s not, what is the way?” A growing group of academics and AI researchers are making the case that the answer is to go smaller, trading sweeping ambition for an approach inspired by the social sciences. Specifically, they want to focus more on testing validity, which for quantitative social scientists refers to how well a given questionnaire measures what it’s claiming to measure—and, more fundamentally, whether what it is measuring has a coherent definition. That could cause trouble for benchmarks assessing hazily defined concepts like “reasoning” or “scientific knowledge”—and for developers aiming to reach the much-hyped goal of artificial general intelligence—but it would put the industry on firmer ground as it looks to prove the worth of individual models. “Taking validity seriously means asking folks in academia, industry, or wherever to show that their system does what they say it does,” says Abigail Jacobs, a University of Michigan professor who is a central figure in the new push for validity. “I think it points to a weakness in the AI world if they want to back off from showing that they can support their claim.” The limits of traditional testing If AI companies have been slow to respond to the growing failure of benchmarks, it’s partially because the test-scoring approach has been so effective for so long. One of the biggest early successes of contemporary AI was the ImageNet challenge, a kind of antecedent to contemporary benchmarks. Released in 2010 as an open challenge to researchers, the database held more than 3 million images for AI systems to categorize into 1,000 different classes. Crucially, the test was completely agnostic to methods, and any successful algorithm quickly gained credibility regardless of how it worked. When an algorithm called AlexNet broke through in 2012, with a then unconventional form of GPU training, it became one of the foundational results of modern AI. Few would have guessed in advance that AlexNet’s convolutional neural nets would be the secret to unlocking image recognition—but after it scored well, no one dared dispute it. (One of AlexNet’s developers, Ilya Sutskever, would go on to cofound OpenAI.) A large part of what made this challenge so effective was that there was little practical difference between ImageNet’s object classification challenge and the actual process of asking a computer to recognize an image. Even if there were disputes about methods, no one doubted that the highest-scoring model would have an advantage when deployed in an actual image recognition system. But in the 12 years since, AI researchers have applied that same method-agnostic approach to increasingly general tasks. SWE-Bench is commonly used as a proxy for broader coding ability, while other exam-style benchmarks often stand in for reasoning ability. That broad scope makes it difficult to be rigorous about what a specific benchmark measures—which, in turn, makes it hard to use the findings responsibly. Where things break down Anka Reuel, a PhD student who has been focusing on the benchmark problem as part of her research at Stanford, has become convinced the evaluation problem is the result of this push toward generality. “We’ve moved from task-specific models to general-purpose models,” Reuel says. “It’s not about a single task anymore but a whole bunch of tasks, so evaluation becomes harder.” Like the University of Michigan’s Jacobs, Reuel thinks “the main issue with benchmarks is validity, even more than the practical implementation,” noting: “That’s where a lot of things break down.” For a task as complicated as coding, for instance, it’s nearly impossible to incorporate every possible scenario into your problem set. As a result, it’s hard to gauge whether a model is scoring better because it’s more skilled at coding or because it has more effectively manipulated the problem set. And with so much pressure on developers to achieve record scores, shortcuts are hard to resist. For developers, the hope is that success on lots of specific benchmarks will add up to a generally capable model. But the techniques of agentic AI mean a single AI system can encompass a complex array of different models, making it hard to evaluate whether improvement on a specific task will lead to generalization. “There’s just many more knobs you can turn,” says Sayash Kapoor, a computer scientist at Princeton and a prominent critic of sloppy practices in the AI industry. “When it comes to agents, they have sort of given up on the best practices for evaluation.” In a paper from last July, Kapoor called out specific issues in how AI models were approaching the WebArena benchmark, designed by Carnegie Mellon University researchers in 2024 as a test of an AI agent’s ability to traverse the web. The benchmark consists of more than 800 tasks to be performed on a set of cloned websites mimicking Reddit, Wikipedia, and others. Kapoor and his team identified an apparent hack in the winning model, called STeP. STeP included specific instructions about how Reddit structures URLs, allowing STeP models to jump directly to a given user’s profile page (a frequent element of WebArena tasks). This shortcut wasn’t exactly cheating, but Kapoor sees it as “a serious misrepresentation of how well the agent would work had it seen the tasks in WebArena for the first time.” Because the technique was successful, though, a similar policy has since been adopted by OpenAI’s web agent Operator. (“Our evaluation setting is designed to assess how well an agent can solve tasks given some instruction about website structures and task execution,” an OpenAI representative said when reached for comment. “This approach is consistent with how others have used and reported results with WebArena.” STeP did not respond to a request for comment.) Further highlighting the problem with AI benchmarks, late last month Kapoor and a team of researchers wrote a paper that revealed significant problems in Chatbot Arena, the popular crowdsourced evaluation system. According to the paper, the leaderboard was being manipulated; many top foundation models were conducting undisclosed private testing and releasing their scores selectively. Today, even ImageNet itself, the mother of all benchmarks, has started to fall victim to validity problems. A 2023 study from researchers at the University of Washington and Google Research found that when ImageNet-winning algorithms were pitted against six real-world data sets, the architecture improvement “resulted in little to no progress,” suggesting that the external validity of the test had reached its limit. Going smaller For those who believe the main problem is validity, the best fix is reconnecting benchmarks to specific tasks. As Reuel puts it, AI developers “have to resort to these high-level benchmarks that are almost meaningless for downstream consumers, because the benchmark developers can’t anticipate the downstream task anymore.” So what if there was a way to help the downstream consumers identify this gap? In November 2024, Reuel launched a public ranking project called BetterBench, which rates benchmarks on dozens of different criteria, such as whether the code has been publicly documented. But validity is a central theme, with particular criteria challenging designers to spell out what capability their benchmark is testing and how it relates to the tasks that make up the benchmark. “You need to have a structural breakdown of the capabilities,” Reuel says. “What are the actual skills you care about, and how do you operationalize them into something we can measure?” The results are surprising. One of the highest-scoring benchmarks is also the oldest: the Arcade Learning Environment (ALE), established in 2013 as a way to test models’ ability to learn how to play a library of Atari 2600 games. One of the lowest-scoring is the Massive Multitask Language Understanding (MMLU) benchmark, a widely used test for general language skills; by the standards of BetterBench, the connection between the questions and the underlying skill was too poorly defined. BetterBench hasn’t meant much for the reputations of specific benchmarks, at least not yet; MMLU is still widely used, and ALE is still marginal. But the project has succeeded in pushing validity into the broader conversation about how to fix benchmarks. In April, Reuel quietly joined a new research group hosted by Hugging Face, the University of Edinburgh, and EleutherAI, where she’ll develop her ideas on validity and AI model evaluation with other figures in the field. (An official announcement is expected later this month.) Irene Solaiman, Hugging Face’s head of global policy, says the group will focus on building valid benchmarks that go beyond measuring straightforward capabilities. “There’s just so much hunger for a good benchmark off the shelf that already works,” Solaiman says. “A lot of evaluations are trying to do too much.” Increasingly, the rest of the industry seems to agree. In a paper in March, researchers from Google, Microsoft, Anthropic, and others laid out a new framework for improving evaluations—with validity as the first step. “AI evaluation science must,” the researchers argue, “move beyond coarse grained claims of ‘general intelligence’ towards more task-specific and real-world relevant measures of progress.” Measuring the “squishy” things To help make this shift, some researchers are looking to the tools of social science. A February position paper argued that “evaluating GenAI systems is a social science measurement challenge,” specifically unpacking how the validity systems used in social measurements can be applied to AI benchmarking. The authors, largely employed by Microsoft’s research branch but joined by academics from Stanford and the University of Michigan, point to the standards that social scientists use to measure contested concepts like ideology, democracy, and media bias. Applied to AI benchmarks, those same procedures could offer a way to measure concepts like “reasoning” and “math proficiency” without slipping into hazy generalizations. In the social science literature, it’s particularly important that metrics begin with a rigorous definition of the concept measured by the test. For instance, if the test is to measure how democratic a society is, it first needs to establish a definition for a “democratic society” and then establish questions that are relevant to that definition. To apply this to a benchmark like SWE-Bench, designers would need to set aside the classic machine learning approach, which is to collect programming problems from GitHub and create a scheme to validate answers as true or false. Instead, they’d first need to define what the benchmark aims to measure (“ability to resolve flagged issues in software,” for instance), break that into subskills (different types of problems or types of program that the AI model can successfully process), and then finally assemble questions that accurately cover the different subskills. It’s a profound change from how AI researchers typically approach benchmarking—but for researchers like Jacobs, a coauthor on the February paper, that’s the whole point. “There’s a mismatch between what’s happening in the tech industry and these tools from social science,” she says. “We have decades and decades of thinking about how we want to measure these squishy things about humans.” Even though the idea has made a real impact in the research world, it’s been slow to influence the way AI companies are actually using benchmarks. The last two months have seen new model releases from OpenAI, Anthropic, Google, and Meta, and all of them lean heavily on multiple-choice knowledge benchmarks like MMLU—the exact approach that validity researchers are trying to move past. After all, model releases are, for the most part, still about showing increases in general intelligence, and broad benchmarks continue to be used to back up those claims. For some observers, that’s good enough. Benchmarks, Wharton professor Ethan Mollick says, are “bad measures of things, but also they’re what we’ve got.” He adds: “At the same time, the models are getting better. A lot of sins are forgiven by fast progress.” For now, the industry’s long-standing focus on artificial general intelligence seems to be crowding out a more focused validity-based approach. As long as AI models can keep growing in general intelligence, then specific applications don’t seem as compelling—even if that leaves practitioners relying on tools they no longer fully trust. “This is the tightrope we’re walking,” says Hugging Face’s Solaiman. “It’s too easy to throw the system out, but evaluations are really helpful in understanding our models, even with these limitations.” Russell Brandom is a freelance writer covering artificial intelligence. He lives in Brooklyn with his wife and two cats. This story was supported by a grant from the Tarbell Center for AI Journalism.

It’s not easy being one of Silicon Valley’s favorite benchmarks.

SWE-Bench (pronounced “swee bench”) launched in November 2024 to evaluate an AI model’s coding skill, using more than 2,000 real-world programming problems pulled from the public GitHub repositories of 12 different Python-based projects.

In the months since then, it’s quickly become one of the most popular tests in AI. A SWE-Bench score has become a mainstay of major model releases from OpenAI, Anthropic, and Google—and outside of foundation models, the fine-tuners at AI firms are in constant competition to see who can rise above the pack. The top of the leaderboard is a pileup between three different fine tunings of Anthropic’s Claude Sonnet model and Amazon’s Q developer agent. Auto Code Rover—one of the Claude modifications—nabbed the number two spot in November, and was acquired just three months later.

Despite all the fervor, this isn’t exactly a truthful assessment of which model is “better.” As the benchmark has gained prominence, “you start to see that people really want that top spot,” says John Yang, a researcher on the team that developed SWE-Bench at Princeton University. As a result, entrants have begun to game the system—which is pushing many others to wonder whether there’s a better way to actually measure AI achievement.

Developers of these coding agents aren’t necessarily doing anything as straightforward cheating, but they’re crafting approaches that are too neatly tailored to the specifics of the benchmark. The initial SWE-Bench test set was limited to programs written in Python, which meant developers could gain an advantage by training their models exclusively on Python code. Soon, Yang noticed that high-scoring models would fail completely when tested on different programming languages—revealing an approach to the test that he describes as “gilded.”

“It looks nice and shiny at first glance, but then you try to run it on a different language and the whole thing just kind of falls apart,” Yang says. “At that point, you’re not designing a software engineering agent. You’re designing to make a SWE-Bench agent, which is much less interesting.”

The SWE-Bench issue is a symptom of a more sweeping—and complicated—problem in AI evaluation, and one that’s increasingly sparking heated debate: The benchmarks the industry uses to guide development are drifting further and further away from evaluating actual capabilities, calling their basic value into question. Making the situation worse, several benchmarks, most notably FrontierMath and Chatbot Arena, have recently come under heat for an alleged lack of transparency. Nevertheless, benchmarks still play a central role in model development, even if few experts are willing to take their results at face value. OpenAI cofounder Andrej Karpathy recently described the situation as “an evaluation crisis”: the industry has fewer trusted methods for measuring capabilities and no clear path to better ones.

“Historically, benchmarks were the way we evaluated AI systems,” says Vanessa Parli, director of research at Stanford University’s Institute for Human-Centered AI. “Is that the way we want to evaluate systems going forward? And if it’s not, what is the way?”

A growing group of academics and AI researchers are making the case that the answer is to go smaller, trading sweeping ambition for an approach inspired by the social sciences. Specifically, they want to focus more on testing validity, which for quantitative social scientists refers to how well a given questionnaire measures what it’s claiming to measure—and, more fundamentally, whether what it is measuring has a coherent definition. That could cause trouble for benchmarks assessing hazily defined concepts like “reasoning” or “scientific knowledge”—and for developers aiming to reach the much–hyped goal of artificial general intelligence—but it would put the industry on firmer ground as it looks to prove the worth of individual models.

“Taking validity seriously means asking folks in academia, industry, or wherever to show that their system does what they say it does,” says Abigail Jacobs, a University of Michigan professor who is a central figure in the new push for validity. “I think it points to a weakness in the AI world if they want to back off from showing that they can support their claim.”

The limits of traditional testing

If AI companies have been slow to respond to the growing failure of benchmarks, it’s partially because the test-scoring approach has been so effective for so long.

One of the biggest early successes of contemporary AI was the ImageNet challenge, a kind of antecedent to contemporary benchmarks. Released in 2010 as an open challenge to researchers, the database held more than 3 million images for AI systems to categorize into 1,000 different classes.

Crucially, the test was completely agnostic to methods, and any successful algorithm quickly gained credibility regardless of how it worked. When an algorithm called AlexNet broke through in 2012, with a then unconventional form of GPU training, it became one of the foundational results of modern AI. Few would have guessed in advance that AlexNet’s convolutional neural nets would be the secret to unlocking image recognition—but after it scored well, no one dared dispute it. (One of AlexNet’s developers, Ilya Sutskever, would go on to cofound OpenAI.)

A large part of what made this challenge so effective was that there was little practical difference between ImageNet’s object classification challenge and the actual process of asking a computer to recognize an image. Even if there were disputes about methods, no one doubted that the highest-scoring model would have an advantage when deployed in an actual image recognition system.

But in the 12 years since, AI researchers have applied that same method-agnostic approach to increasingly general tasks. SWE-Bench is commonly used as a proxy for broader coding ability, while other exam-style benchmarks often stand in for reasoning ability. That broad scope makes it difficult to be rigorous about what a specific benchmark measures—which, in turn, makes it hard to use the findings responsibly.

Where things break down

Anka Reuel, a PhD student who has been focusing on the benchmark problem as part of her research at Stanford, has become convinced the evaluation problem is the result of this push toward generality. “We’ve moved from task-specific models to general-purpose models,” Reuel says. “It’s not about a single task anymore but a whole bunch of tasks, so evaluation becomes harder.”

Like the University of Michigan’s Jacobs, Reuel thinks “the main issue with benchmarks is validity, even more than the practical implementation,” noting: “That’s where a lot of things break down.” For a task as complicated as coding, for instance, it’s nearly impossible to incorporate every possible scenario into your problem set. As a result, it’s hard to gauge whether a model is scoring better because it’s more skilled at coding or because it has more effectively manipulated the problem set. And with so much pressure on developers to achieve record scores, shortcuts are hard to resist.

For developers, the hope is that success on lots of specific benchmarks will add up to a generally capable model. But the techniques of agentic AI mean a single AI system can encompass a complex array of different models, making it hard to evaluate whether improvement on a specific task will lead to generalization. “There’s just many more knobs you can turn,” says Sayash Kapoor, a computer scientist at Princeton and a prominent critic of sloppy practices in the AI industry. “When it comes to agents, they have sort of given up on the best practices for evaluation.”

In a paper from last July, Kapoor called out specific issues in how AI models were approaching the WebArena benchmark, designed by Carnegie Mellon University researchers in 2024 as a test of an AI agent’s ability to traverse the web. The benchmark consists of more than 800 tasks to be performed on a set of cloned websites mimicking Reddit, Wikipedia, and others. Kapoor and his team identified an apparent hack in the winning model, called STeP. STeP included specific instructions about how Reddit structures URLs, allowing STeP models to jump directly to a given user’s profile page (a frequent element of WebArena tasks).

This shortcut wasn’t exactly cheating, but Kapoor sees it as “a serious misrepresentation of how well the agent would work had it seen the tasks in WebArena for the first time.” Because the technique was successful, though, a similar policy has since been adopted by OpenAI’s web agent Operator. (“Our evaluation setting is designed to assess how well an agent can solve tasks given some instruction about website structures and task execution,” an OpenAI representative said when reached for comment. “This approach is consistent with how others have used and reported results with WebArena.” STeP did not respond to a request for comment.)

Further highlighting the problem with AI benchmarks, late last month Kapoor and a team of researchers wrote a paper that revealed significant problems in Chatbot Arena, the popular crowdsourced evaluation system. According to the paper, the leaderboard was being manipulated; many top foundation models were conducting undisclosed private testing and releasing their scores selectively.

Today, even ImageNet itself, the mother of all benchmarks, has started to fall victim to validity problems. A 2023 study from researchers at the University of Washington and Google Research found that when ImageNet-winning algorithms were pitted against six real-world data sets, the architecture improvement “resulted in little to no progress,” suggesting that the external validity of the test had reached its limit.

Going smaller

For those who believe the main problem is validity, the best fix is reconnecting benchmarks to specific tasks. As Reuel puts it, AI developers “have to resort to these high-level benchmarks that are almost meaningless for downstream consumers, because the benchmark developers can’t anticipate the downstream task anymore.” So what if there was a way to help the downstream consumers identify this gap?

In November 2024, Reuel launched a public ranking project called BetterBench, which rates benchmarks on dozens of different criteria, such as whether the code has been publicly documented. But validity is a central theme, with particular criteria challenging designers to spell out what capability their benchmark is testing and how it relates to the tasks that make up the benchmark.

“You need to have a structural breakdown of the capabilities,” Reuel says. “What are the actual skills you care about, and how do you operationalize them into something we can measure?”

The results are surprising. One of the highest-scoring benchmarks is also the oldest: the Arcade Learning Environment (ALE), established in 2013 as a way to test models’ ability to learn how to play a library of Atari 2600 games. One of the lowest-scoring is the Massive Multitask Language Understanding (MMLU) benchmark, a widely used test for general language skills; by the standards of BetterBench, the connection between the questions and the underlying skill was too poorly defined.

BetterBench hasn’t meant much for the reputations of specific benchmarks, at least not yet; MMLU is still widely used, and ALE is still marginal. But the project has succeeded in pushing validity into the broader conversation about how to fix benchmarks. In April, Reuel quietly joined a new research group hosted by Hugging Face, the University of Edinburgh, and EleutherAI, where she’ll develop her ideas on validity and AI model evaluation with other figures in the field. (An official announcement is expected later this month.)

Irene Solaiman, Hugging Face’s head of global policy, says the group will focus on building valid benchmarks that go beyond measuring straightforward capabilities. “There’s just so much hunger for a good benchmark off the shelf that already works,” Solaiman says. “A lot of evaluations are trying to do too much.”

Increasingly, the rest of the industry seems to agree. In a paper in March, researchers from Google, Microsoft, Anthropic, and others laid out a new framework for improving evaluations—with validity as the first step.

“AI evaluation science must,” the researchers argue, “move beyond coarse grained claims of ‘general intelligence’ towards more task-specific and real-world relevant measures of progress.”

Measuring the “squishy” things

To help make this shift, some researchers are looking to the tools of social science. A February position paper argued that “evaluating GenAI systems is a social science measurement challenge,” specifically unpacking how the validity systems used in social measurements can be applied to AI benchmarking.

The authors, largely employed by Microsoft’s research branch but joined by academics from Stanford and the University of Michigan, point to the standards that social scientists use to measure contested concepts like ideology, democracy, and media bias. Applied to AI benchmarks, those same procedures could offer a way to measure concepts like “reasoning” and “math proficiency” without slipping into hazy generalizations.

In the social science literature, it’s particularly important that metrics begin with a rigorous definition of the concept measured by the test. For instance, if the test is to measure how democratic a society is, it first needs to establish a definition for a “democratic society” and then establish questions that are relevant to that definition.

To apply this to a benchmark like SWE-Bench, designers would need to set aside the classic machine learning approach, which is to collect programming problems from GitHub and create a scheme to validate answers as true or false. Instead, they’d first need to define what the benchmark aims to measure (“ability to resolve flagged issues in software,” for instance), break that into subskills (different types of problems or types of program that the AI model can successfully process), and then finally assemble questions that accurately cover the different subskills.

It’s a profound change from how AI researchers typically approach benchmarking—but for researchers like Jacobs, a coauthor on the February paper, that’s the whole point. “There’s a mismatch between what’s happening in the tech industry and these tools from social science,” she says. “We have decades and decades of thinking about how we want to measure these squishy things about humans.”

Even though the idea has made a real impact in the research world, it’s been slow to influence the way AI companies are actually using benchmarks.

The last two months have seen new model releases from OpenAI, Anthropic, Google, and Meta, and all of them lean heavily on multiple-choice knowledge benchmarks like MMLU—the exact approach that validity researchers are trying to move past. After all, model releases are, for the most part, still about showing increases in general intelligence, and broad benchmarks continue to be used to back up those claims.

For some observers, that’s good enough. Benchmarks, Wharton professor Ethan Mollick says, are “bad measures of things, but also they’re what we’ve got.” He adds: “At the same time, the models are getting better. A lot of sins are forgiven by fast progress.”

For now, the industry’s long-standing focus on artificial general intelligence seems to be crowding out a more focused validity-based approach. As long as AI models can keep growing in general intelligence, then specific applications don’t seem as compelling—even if that leaves practitioners relying on tools they no longer fully trust.

“This is the tightrope we’re walking,” says Hugging Face’s Solaiman. “It’s too easy to throw the system out, but evaluations are really helpful in understanding our models, even with these limitations.”

Russell Brandom is a freelance writer covering artificial intelligence. He lives in Brooklyn with his wife and two cats.

This story was supported by a grant from the Tarbell Center for AI Journalism.

Stay Ahead

Explore More Insights

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy, bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

Cisco unveils integrated edge platform for AI

Announced at Cisco’s Partner Summit, Unified Edge will likely be part of many third-party packages that can be configured in a variety of ways, Cisco stated. “The platform is customer definable. For example, if a customer has a workload and they’ve decided they want to use Nutanix, they can go

Infoblox bolsters Universal DDI Platform with multi-cloud integrations

Universal DDI for Microsoft Management integration enables enterprises to gain control of their DNS and DHCP by centrally managing DNS and DHCP hosted on Microsoft server platforms. Integration with Google Cloud Internal Range applies consistent IPAM policies across Google Cloud, on-premises, and other cloud environments, which helps enterprise IT to

Agentic AI: What now, what next?

Agentic AI burst onto the scene with its promises of streamliningoperations and accelerating productivity. But what’s real and what’s hype when it comes to deploying agentic AI? This Special Report examines the state of agentic AI, the challenges organizations are facing in deploying it, and the lessons learned from success

Wolfgang Wendt: “I like to describe IBM as an ‘older start-up’”

Wolfgang Wendt, Managing Director IBM DACH, has worked for Big Blue for more than 30 years. Parts of the IBM System Z mainframe system are developed at IBM Germany in Böblingen.

Uniper Completes Sale of Ruhr District Heating Unit

Uniper SE said Monday it had completed the divestment of Uniper Waerme GmbH, a district heating network serving over 14,400 customers in Germany’s Ruhr area, to Steag Iqony Group’s Iqony Fernwaerme GmbH. Waerme is among the assets that the German power and gas utility agreed to sell to fulfill fair-competition guardrails imposed by the European Commission in approving Uniper’s bailout by the government in late 2022. Waerme has a network of over 750 kilometers (466.03 miles), according to Uniper. Waerme “is an expert in the efficient use of heat that is generated during electricity production in combined heat and power plants”, Uniper said in a press release. “In addition, they use a variety of other environmentally friendly alternatives for heat generation. This includes heat from mine gas, waste heat from industrial processes and heat generated in electric boilers and smaller decentralized CHP plants”. In the initial announcement of the transaction on August 4, Waerme managing director Nikola Feldmann said, “We are pleased to have found a buyer in the Steag Iqony Group that will continue on the path we have taken and be a reliable employer for our around 130 colleagues”. The companies did not disclose the transaction price. Earlier Uniper said it had signed an agreement to sell the Datteln IV coal-run power plant in North Rhine-Westphalia to ResInvest Group AS, toward the satisfaction of the bailout conditions. Commissioned 2020, the Datteln plant has a net output of 1,052 megawatts (MW). It supplies electricity and district heating to households, as well as traction power to rail operator Deutsche Bahn, according to Uniper. The over 100 employees at the site will transfer to Czechia’s ResInvest, Uniper said in an online statement September 19. The transaction is subject to regulatory approvals, it said. In July Uniper said it had sold its 18.26

Aramco Q3 Profit Tops Estimates

Saudi Aramco posted a surprise increase in third-quarter profit as a production boost helped mitigate the impact of lower crude prices and helped the oil giant break a yearslong streak of falling earnings. Adjusted net income for the period rose about 1 percent to 104.9 billion riyals ($28 billion) from a year earlier, surpassing analyst estimates compiled by Bloomberg. Free cash flow exceeded the dividend payout for the first time in about two years, while net debt eased compared with three months ago. The latest results follow a sequence of lower quarterly profit at Aramco over the past couple of years, and follow the firm’s move to raise output as part of an OPEC+ policy that’s helped it counter muted crude prices. The world’s biggest oil exporter is a lynchpin of the Saudi economy, with revenue from oil sales and hefty dividend payouts supporting the kingdom’s multitrilion-dollar economic rejig. Oil prices in London have declined 13 percent this year to about $65 a barrel, well below the more than $90 that the International Monetary Fund says Saudi Arabia needs to balance its budget. That’s translated into pullbacks in some major infrastructure and tourism projects in the kingdom, while Aramco has also slowed some domestic refining and chemical plans as it focuses on a mega natural gas development. Aramco sold its oil at about $70 a barrel in the third quarter, compared with nearly $79 a year earlier. But liquids production increased 3.8 percent to 10.8 million barrels a day, while natural gas output rose 5 percent. The company’s “ability to quickly ramp-up production and capture rising demand drove our strong third quarter performance,” Chief Financial Officer Ziad Al-Murshed said in the statement. Aramco’s free cash flow – funds left over from operations after accounting for investments and expenses – rose to $23.6 billion in the quarter. That

Eni, Petronas Seal Deal for Indonesia-Malaysia JV

Malaysia’s Petroliam Nasional Bhd (Petronas) and Italy’s Eni SpA on Monday announced a binding agreement to combine their assets in Indonesia and Malaysia into a company equally owned by the state-controlled oil and gas producers. The independent company, to be called NewCo, will focus on natural gas-producing and development assets including in Indonesia’s Kutei Basin, Eni said in an online statement Monday. Announcing the prior memorandum of understanding February 27, Eni and Petronas said the combination would create a “major” liquefied natural gas player in the Asian market. Last year Indonesian authorities approved Eni’s development plan for the offshore Gehem and Geng North fields, which includes the construction of a new production hub with an output capacity of about two billion cubic feet a day of gas and 80,000 barrels of oil per day of condensates. Geng North was discovered 2023 under the North Ganal production sharing contract (PSC). Gehem meanwhile came under Eni when it acquired Chevron Corp’s operating stake in the Rapak PSC, as well as the Ganal and Makassar Straits PSCs, in 2023. The government also approved at the time Eni’s development plan for the Gendalo and Gandang fields, which would be tied back to existing infrastructure. Gendalo and Gandang are under the Ganal block, which is separate from the North Ganal block. Eni at the time also secured a 20-year extension to the Indonesia Deepwater Development gas project, which consists of the Ganal and Rapak blocks, as announced by the company August 23, 2024. The Ganal, North Ganal, Makassar Straits and Rapak blocks sit in the Kutei Basin off the coast of East Kalimantan province. The province is on the Indonesian side of Borneo, an island shared with Brunei and Malaysia. NewCo will have an initial production capacity of over 300,000 barrels of oil equivalent a

Adnoc CEO Says AI Raises Energy Investment Needs to $4T

The global energy industry needs annual investment of $4 trillion as the boom in data centers and artificial intelligence increases demand, according to Sultan Al Jaber, head of the biggest crude producer in the United Arab Emirates. Long-term demand growth will outlast immediate concerns over an oil supply glut, Al Jaber, chief executive officer of Abu Dhabi National Oil Co., said at a conference in Abu Dhabi on Monday. Investors will need to develop the resources to drive the coming data boom, including revamping power grids, he said. Oil producers like the UAE are boosting output capacity, even as most analysts warn of a coming glut in crude supply that will weigh further on prices next year. Brent crude is down nearly 13% so far this year. Geopolitical concerns that have delayed shipments and threatened supply disruptions have prevented prices from dropping further. “Near-term uncertainty is real, while long-term demand remains strong,” Al Jaber said. “Our response to meet that demand should focus on data, not the drama.” In a nod to the short-term weakness in markets, the OPEC+ producers group said on Sunday that it will pause production increases during the first quarter of next year after a modest increase this month and in December. Oil demand is set to remain above 100 million barrels a day beyond 2040, requiring added investment, the CEO said. The UAE aims to play a role in this energy investment push, with Al Jaber declaring that the country is “open for business” and that Adnoc’s international arm XRG is seeking more deals.

Oil Holds Steady on OPEC+ Pause

Oil was little changed Monday as traders weighed the OPEC+ alliance’s plan to pause its output revival next quarter on anticipation demand will slow, while the market is seen headed for oversupply. West Texas Intermediate rose about 0.1% to settle above $61 a barrel after fluctuating between small gains and losses through the day, extending a string of marginal increases. The Organization of the Petroleum Exporting Countries and its partners said the decision on Sunday to halt production hikes from January reflects an expectation for a seasonal slowdown. The move comes against a backdrop of widespread forecasts for excess supplies next year that could weigh down prices. The US benchmark has slumped about 9% over the past three months as OPEC+ ramped up output in an apparent effort to regain market share, while producers outside the group also increased production. Prices recently bounced from a five-month low after tighter US sanctions on two major Russian oil producers over the war in Ukraine raised some questions about supply from Moscow. “The decision to halt quota hikes during 1Q does not materially change our production forecasts but still sends an important signal,” Morgan Stanley analysts including Martijn Rats and Charlotte Firkins wrote. “The group is still adjusting supply in response to market conditions.” The eight key members of OPEC+ are left with roughly 1.2 million barrels a day of their current supply tranche still to restore. Actual output increases have fallen short of advertised volumes, as some members offset earlier overproduction and others struggle to pump more. Following the OPEC+ move, Morgan Stanley raised its near-term price forecast for Brent while also maintaining a warning for a “substantial surplus.” The United Arab Emirates, meanwhile, on Monday added to the chorus of producers who have come out to downplay glut concerns. Traders will

Gas continues to dominate Entergy plans as data center pipeline grows

7-12 GW Data center pipeline, up from 5-10 GW last quarter. 4.5 GW New power generating secured for “large growth opportunities.” $41 billion Estimated capital spending through 2029. Pending and signed hyperscaler deals Entergy is a vertically-integrated electric utility with five operating companies in four states: Arkansas, Louisiana, Mississippi and Texas. Its weather-adjusted retail sales rose 4.4% compared to last year. Profit increased from $645 million in the third quarter of 2024 to $694 million in the third quarter of 2025. In 2024, commercial and industrial customers made up 69% of Entergy’s sales, which were concentrated in Louisiana. The company has agreements for several notable large load projects with Google and Meta. In general, it expects substantial load growth from commercial and industrial customers through 2029, though not the breakneck pace utilities like Dominion Energy expect in data center hotspots such as northern Virginia. In its third-quarter investor update, Entergy said its data center customer pipeline increased 2 GW from the previous quarter to land in the 7 GW to 12 GW range. Also in the quarter, Energy secured 4.5 GW of power generation equipment to serve new load expected to come online this decade. In August, the Louisiana Public Service Commission approved the generation and transmission resources needed to support Meta’s planned 2-GW campus in northeastern Louisiana. Those include three combined-cycle gas plants with combined nameplate capacity of about 2.2 GW. As part of the same proceeding, the commission also authorized Entergy to procure up to 1.5 GW of solar resources for the project. Entergy executives said Wednesday that the Meta agreement and additional expected load growth are reflected in the utility’s five-year capital plan, which lays out $41 billion in estimated spending through 2029. Entergy has also submitted an application to the Arkansas Public Service Commission for the Cypress

OpenAI spends even more money it doesn’t have

The aim, said Gogia, “is continuity, not cost efficiency. These deals are forward leaning, relying on revenue forecasts that remain speculative. In that context, OpenAI must continue to draw heavily on outside capital, whether through venture rounds, debt, or a future public offering.” He pointed out, “the company’s recent legal and corporate restructuring was designed to open the doors to that capital. Removing Microsoft’s exclusivity makes room for more vendors but also signals that no one provider can meet OpenAI’s demands. In several cases, suppliers are stepping in with financing arrangements that link product sales to future performance. While these strategies help close funding gaps, they introduce fragility. What looks like revenue is often pre-paid consumption, not realized margin.” Execution risks, he said, add to the concern. “Building and energizing enough data centers to meet OpenAI’s projected needs is not a function of ambition alone. It requires grid access, cooling capacity, and regional stability. Microsoft has acknowledged that it lacks the power infrastructure to fully deploy the GPUs it owns. Without physical readiness, all of these agreements sit on shaky ground.” Lots of equity swapping going on Scott Bickley, advisory fellow at Info-Tech Research Group, said he has not only been astounded by the funding announcements over the last few months, but is also appalled, primarily, he said, “because of the disconnect to what this does to the underlying technology stocks and their market prices versus where the technology is at from a development and ROI perspective … and from a boots on the ground perspective.” He added that while the financial pledges involve “huge, staggering numbers, most of them are tied up in ways that are not necessarily going to require all the cash to come from OpenAI. In a lot of cases, there is equity swapping. You have

Verizon to build high-capacity fiber network to link AWS AI data centers

“AI will be essential to the future of business and society, driving innovation that demands a network to match,” Scott Lawrence, senior vice president and chief product officer at Verizon Business said in a statement. “This deal with Amazon demonstrates our continued commitment to meet the growing demands of AI workloads for the businesses and developers building our future.” This is not the first time that two companies have partnered. Verizon has previously adopted AWS as a preferred public cloud provider for its digital transformation efforts. The collaboration also extends to joint development of private mobile edge computing solutions, delivering secure, dedicated connectivity for enterprise customers. These efforts have been targeted at industries such as manufacturing, healthcare, retail, and entertainment.

Supermicro Unveils Data Center Building Blocks to Accelerate AI Factory Deployment

Supermicro has introduced a new business line, Data Center Building Block Solutions (DCBBS), expanding its modular approach to data center development. The offering packages servers, storage, liquid-cooling infrastructure, networking, power shelves and battery backup units (BBUs), DCIM and automation software, and on-site services into pre-validated, factory-tested bundles designed to accelerate time-to-online (TTO) and improve long-term serviceability. This move represents a significant step beyond traditional rack integration; a shift toward a one-stop, data-center-scale platform aimed squarely at the hyperscale and AI factory market. By providing a single point of accountability across IT, power, and thermal domains, Supermicro’s model enables faster deployments and reduces integration risk—the modern equivalent of a “single throat to choke” for data center operators racing to bring GB200/NVL72-class racks online. What’s New in DCBBS DCBBS extends Supermicro’s modular design philosophy to an integrated catalog of facility-adjacent building blocks, not just IT nodes. By including critical supporting infrastructure—cooling, power, networking, and lifecycle software—the platform helps operators bring new capacity online more quickly and predictably. According to Supermicro, DCBBS encompasses: Multi-vendor AI system support: Compatibility with NVIDIA, AMD, and Intel architectures, featuring Supermicro-designed cold plates that dissipate up to 98% of component-level heat. In-rack liquid-cooling designs: Coolant distribution manifolds (CDMs) and CDUs rated up to 250 kW, supporting 45 °C liquids, alongside rear-door heat exchangers, 800 GbE switches (51.2 Tb/s), 33 kW power shelves, and 48 V battery backup units. Liquid-to-Air (L2A) sidecars: Each row can reject up to 200 kW of heat without modifying existing building hydronics—an especially practical design for air-to-liquid retrofits. Automation and management software: SuperCloud Composer for rack-scale and liquid-cooling lifecycle management SuperCloud Automation Center for firmware, OS, Kubernetes, and AI pipeline enablement Developer Experience Console for self-service workflows and orchestration End-to-end services: Design, validation, and on-site deployment options—including four-hour response service levels—for both greenfield builds

Investments Anchor Vertiv’s Growth Strategy as AI-Driven Data Center Orders Surge 60% YoY

New Acquisitions and Partner Awards Vertiv’s third-quarter financial performance was underscored by a series of strategic acquisitions and ecosystem recognitions that expand the company’s technological capabilities and market reach amid AI-driven demand. Acquisition of Waylay NV: AI and Hyperautomation for Infrastructure Intelligence On August 26, Vertiv announced its acquisition of Waylay NV, a Belgium-based developer of generative AI and hyperautomation software. The move bolsters Vertiv’s portfolio with AI-driven monitoring, predictive services, and performance optimization for digital infrastructure. Waylay’s automation platform integrates real-time analytics, orchestration, and workflow automation across diverse connected assets and cloud services—enabling predictive maintenance, uptime optimization, and energy management across power and cooling systems. “With the addition of Waylay’s technology and software-focused team, Vertiv will accelerate its vision of intelligent infrastructure—data-driven, proactive, and optimized for the world’s most demanding environments,” said CEO Giordano Albertazzi. Completion of Great Lakes Acquisition: Expanding White Space Integration Just days earlier, as alluded to above, Vertiv finalized its $200 million acquisition of Great Lakes Data Racks & Cabinets, a U.S.-based manufacturer of enclosures and integrated rack systems. The addition expands Vertiv’s capabilities in high-density, factory-integrated white space solutions; bridging power, cooling, and IT enclosures for hyperscale and edge data centers alike. Great Lakes’ U.S. and European manufacturing footprint complements Vertiv’s global reach, supporting faster deployment cycles and expanded configuration flexibility. Albertazzi noted that the acquisition “enhances our ability to deliver comprehensive infrastructure solutions, furthering Vertiv’s capabilities to customize at scale and configure at speed for AI and high-density computing environments.” 2024 Partner Awards: Recognizing the Ecosystem Behind Growth Vertiv also spotlighted its partner ecosystem in August with its 2024 North America Partner Awards. The company recognized 11 partners for 2024 performance, growth, and AI execution across segments: Partner of the Year – SHI for launching a customer-facing high-density AI & Cyber Labs featuring

QuEra’s Quantum Leap: From Neutral-Atom Breakthroughs to Hybrid HPC Integration

The race to make quantum computing practical – and commercially consequential – took a major step forward this fall, as Boston-based QuEra Computing announced new research milestones, expanded strategic funding, and an accelerating roadmap for hybrid quantum-classical supercomputing. QuEra’s Chief Commercial Officer Yuval Boger joined the Data Center Frontier Show to discuss how neutral-atom quantum systems are moving from research labs into high-performance computing centers and cloud environments worldwide. NVIDIA Joins Google in Backing QuEra’s $230 Million Round In early September, QuEra disclosed that NVentures, NVIDIA’s venture arm, has joined Google and others in expanding its $230 million Series B round. The investment deepens what has already been one of the most active collaborations between quantum and accelerated-computing companies. “We already work with NVIDIA, pairing our scalable neutral-atom architecture with its accelerated-computing stack to speed the arrival of useful, fault-tolerant quantum machines,” said QuEra CEO Andy Ory. “The decision to invest in us underscores our shared belief that hybrid quantum-classical systems will unlock meaningful value for customers sooner than many expect.” The partnership spans hardware, software, and go-to-market initiatives. QuEra’s neutral-atom machines are being integrated into NVIDIA’s CUDA-Q software platform for hybrid workloads, while the two companies collaborate at the NVIDIA Accelerated Quantum Center (NVAQC) in Boston, linking QuEra hardware with NVIDIA’s GB200 NVL72 GPU clusters for simulation and quantum-error-decoder research. Meanwhile, at Japan’s AIST ABCI-Q supercomputing center, QuEra’s Gemini-class quantum computer now operates beside more than 2,000 H100 GPUs, serving as a national testbed for hybrid workflows. A jointly developed transformer-based decoder running on NVIDIA’s GPUs has already outperformed classical maximum-likelihood error-correction models, marking a concrete step toward practical fault-tolerant quantum computing. For NVIDIA, the move signals conviction that quantum processing units (QPUs) will one day complement GPUs inside large-scale data centers. For QuEra, it widens access to the

How CoreWeave and Poolside Are Teaming Up in West Texas to Build the Next Generation of AI Data Centers

In the evolving landscape of artificial-intelligence infrastructure, a singular truth is emerging: access to cutting-edge silicon and massive GPU clusters is no longer enough by itself. For companies chasing the frontier of multi-trillion-parameter model training and agentic AI deployment, the bottleneck increasingly lies not just in compute, but in the seamless integration of compute + power + data center scale. The latest chapter in this story is the collaboration between CoreWeave and Poolside, culminating in the launch of Project Horizon, a 2-gigawatt AI-campus build in West Texas. Setting the Stage: Who’s Involved, and Why It Matters CoreWeave (NASDAQ: CRWV) has positioned itself as “The Essential Cloud for AI™” — a company founded in 2017, publicly listed in March 2025, and aggressively building out its footprint of ultra-high-performance infrastructure. One of its strategic moves: in July 2025 CoreWeave struck a definitive agreement to acquire Core Scientific (NASDAQ: CORZ) in an all-stock transaction. Through that deal, CoreWeave gains grip over approximately 1.3 GW of gross power across Core Scientific’s nationwide data center footprint, plus more than 1 GW of expansion potential. That acquisition underlines a broader trend: AI-specialist clouds are no longer renting space and power; they’re working to own or tightly control it. Poolside, founded in 2023, is a foundation-model company with an ambitious mission: building artificial general intelligence (AGI) and deploying enterprise-scale agents. According to Poolside’s blog: “When people ask what it takes to build frontier AI … the focus is usually on the model … but that’s only half the story. The other half is infrastructure. If you don’t control your infrastructure, you don’t control your destiny—and you don’t have a shot at the frontier.” Simply put: if you’re chasing multi-trillion-parameter models, you need both the compute horsepower and the power infrastructure; and ideally, tight vertical integration. Together, the

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs). In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle

Stay Ahead, Stay ONMINE