Get ONMINE

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.

Shape
Shape
Stay Ahead

Explore More Insights

More Articles

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.

Shape

AI agent traffic drives first profitable year for Fastly

Fetcher bots, which retrieve content in real time when users make queries to AI assistants, show different concentration patterns. OpenAI’s ChatGPT and related bots generated 68% of fetcher bot requests. In some cases, fetcher bot request volumes exceeded 39,000 requests per minute to individual sites. AI agents check multiple websites

Read More »

EBW Warned of Faltering Gas Demand Heading into Holiday Weekend

In a U.S. natural gas focused EBW Analytics Group report sent to Rigzone by the EBW team on Friday, Eli Rubin, an energy analyst at the company, warned of “faltering demand” heading into the President’s Day holiday weekend. “The March contract tested as high as $3.316 yesterday before selling off after a bearish EIA [U.S. Energy Information Administration] storage surprise, and ahead of deteriorating heating demand into President’s Day holiday weekend and an 11 billion cubic foot per day drop into next Wednesday,” Rubin said in Friday’s report. “The threat of cold air in Western Canada and the Pacific Northwest moving into the U.S. remains a primary source of support,” he added. “If the market returns from the holiday weekend without this threat materializing, however, sub-$3.00 per million British thermal units may be in play as the year over year storage deficit flips to a 170 billion cubic foot surplus by late February,” he continued. In the report, Rubin went on to state that “steep storage refill demand east of the Rockies and loose supply/demand fundamentals during recent Marches may offer some medium-term support”. He added, however, that “storage exiting March near 1,800 billion cubic feet, with gathering production tailwinds and decelerating year over year LNG growth into mid to late 2026, suggest a bearish outlook for NYMEX gas futures”. In its latest weekly natural gas storage report, which was released on February 12 and included data for the week ending February 6, the EIA revealed that, according to its estimates, working gas in storage was 2,214 billion cubic feet as of February 6. “This represents a net decrease of 249 billion cubic feet from the previous week,” the EIA highlighted in the report. “Stocks were 97 billion cubic feet less than last year at this time and 130 billion

Read More »

North America Drops 6 Rigs Week on Week

North America dropped six rigs week on week, according to Baker Hughes’ latest North America rotary rig count, which was published on February 13. The total U.S. rig count remained unchanged week on week and the total Canada rig count dropped by six during the same period, pushing the total North America rig count down to 773, comprising 551 rigs from the U.S. and 222 rigs from Canada, the count outlined. Of the total U.S. rig count of 551, 531 rigs are categorized as land rigs, 17 are categorized as offshore rigs, and three are categorized as inland water rigs. The total U.S. rig count is made up of 409 oil rigs, 133 gas rigs, and nine miscellaneous rigs, according to Baker Hughes’ count, which revealed that the U.S. total comprises 481 horizontal rigs, 57 directional rigs, and 13 vertical rigs. Week on week, the U.S. land rig count dropped by one, its offshore rig count rose by one, and its inland water rig count remained unchanged, Baker Hughes highlighted. The U.S. oil rig count decreased by three week on week, while its gas rig count increased by three and its miscellaneous rig count remained unchanged, the count showed. The U.S. horizontal rig count dropped by two week on week, its directional rig count rose by two week on week, and its vertical rig count remained flat during the same period, the count revealed. A major state variances subcategory included in the rig count showed that, week on week, Texas dropped three rigs, Oklahoma and North Dakota each dropped one rig, Louisiana added two rigs, and New Mexico, Pennsylvania, and Wyoming each added one rig. A major basin variances subcategory included in the rig count showed that, week on week, the Permian basin dropped three rigs, the Williston basin dropped

Read More »

Aramco Commits to 1 MMtpa for 20 Years from Commonwealth LNG

Saudi Arabian Oil Co (Aramco) has signed a 20-year agreement to buy one million metric tons per annum (MMtpa) of liquefied natural gas from the under-development Commonwealth LNG in Cameron Parish, Louisiana. “Commonwealth is advancing toward a final investment decision with line of sight to secure its remaining capacity”, said a joint statement by the offtake parties. “Aramco Trading joins Glencore, JERA, PETRONAS, Mercuria and EQT among international energy companies entering into long-term offtake contracts with the platform”. Early this month Commonwealth announced a 20-year deal to supply one MMtpa to Geneva, Switzerland-based energy and commodities trader Mercuria. Commonwealth LNG is a project of Kimmeridge Energy Management Co LLC and Mubadala Investment Co through their joint venture Caturus HoldCo LLC. Expected to start operation 2030, Commonwealth LNG is designed to produce up to 9.5 million metric tons a year of LNG. “This agreement highlights the strong international demand for U.S. LNG and underscores how our longstanding relationships and capabilities position Caturus to serve global markets”, said Caturus chief executive David Lawler. “Our contract with Aramco Trading underscores the differentiated value Caturus can bring through our global reach in offering wellhead to water services”, Lawler added. Mohammed K. Al Mulhim, Aramco Trading president and CEO, said, “This agreement reflects Aramco Trading’s efforts to secure a reliable, long-term energy supply for global markets while strengthening our presence in the LNG sector”. The Gulf Coast project is permitted to ship up to 9.5 MMtpa of LNG, equivalent to around 1.21 billion cubic feet per day of gas according to Kimmeridge. The United States Energy Department granted the project authorization to export to countries without a free trade agreement (FTA) with the U.S. in August 2025 and FTA authorization in April 2020. The developers expect the first phase of the project to generate around

Read More »

Enbridge Q4 Profit Up YoY

Enbridge Inc has reported CAD 1.95 billion ($1.43 billion) in earnings and CAD 1.92 billion in adjusted earnings for the fourth quarter of 2025, up from CAD 493 million and CAD 1.64 billion for the same three-month period in 2024 respectively. Q4 2025 income per share of CAD 0.88 ($0.63), adjusted for extraordinary items, beat the Zacks Consensus Estimate of $0.6. Calgary-based Enbridge, which operates oil and gas pipelines in Canada and the United States, earlier bumped up its quarterly dividend by three percent against the prior rate to CAD 0.97. The annualized rate for 2026 is CAD 3.88 per share. Q4 2025 adjusted EBITDA rose 1.62 percent year-on-year to CAD 5.21 billion “due primarily to favorable gas transmission contracting and Venice Extension entering service, colder weather and higher rates and customer growth at Enbridge Gas Ontario, partially offset by the absence in 2025 of equity earnings related to investment tax credits from our investment in Fox Squirrel Solar”, Enbridge said in an online statement. United States gas transmission contributed CAD 997 million to segment adjusted EBITDA, down from CAD 1 billion for Q4 2024. The U.S. figure benefited from the startup of the Venice Extension Project, which expands the Texas Eastern system’s capacity to deliver gas to Gulf Coast markets, and Enbridge’s acquisition of a stake in the Matterhorn Express Pipeline. Enbridge also recognized “favorable contracting and successful rate case settlements on our U.S. Gas Transmission assets”, partially offset by the timing of operating costs. Adjusted EBITDA from Canadian gas transmission increased from CAD 157 million for Q4 2024 to CAD 190 million for Q4 2025, helped by “higher revenues at Aitken Creek due to favorable storage spreads”. Liquid pipelines logged CAD 2.45 billion in adjusted EBITDA, up from CAD 2.4 billion for Q4 2024. The Mainline System, which carries

Read More »

Analyst Highlights Focus of IEW Event

Focus at the London International Energy Week (IEW) last week was the balancing of geopolitics versus assessed surplus of oil globally in 2026. That’s what Skandinaviska Enskilda Banken AB (SEB) Chief Commodities Analyst Bjarne Schieldrop noted in a SEB report sent to Rigzone on Monday morning, adding that one delegate at the event stated that “if OPEC doesn’t cut, we’ll have $45 per barrel in June”. “That may be true,” Schieldrop said in the report. “But OPEC+ is meeting every month, taking a measure of the state of the global oil market and then decides what to do on the back of that. The group has been very explicit that they may cut, increase, or keep production steady depending on their findings,” he added. “We believe they will and thus we do not buy into $45 per barrel by June because, if need-be, they will trim production as they say they will,” he continued, pointing out that OPEC+ is next scheduled to meet on March 1 “to discuss production for April”. Schieldrop highlighted in the report that, in its February oil market report, the International Energy Agency (IEA) “restated its view that the world will only need 25.7 million barrels per day of crude from OPEC in 2026 versus a recent production by the group of 28.8 million barrels per day”. “I.e. that to keep the market balanced the group will need to cut production by some three million barrels per day,” he said. “Though strategic stock building around the world needs to be deducted from that. And the appetite for such stock building could be solid given elevated geopolitical risks. Thus what will flow to commercial stocks in the end remains to be seen,” he stated. Schieldrop went on to note in the report that increased Iranian tension could drive Brent

Read More »

Hungary Asks Croatia to Allow Russian Crude Shipments

Hungary requested that Croatia allow the shipment of Russian crude via the Adriatic pipeline while a key route through Ukraine remains blocked. Hungarian Foreign Minister Peter Szijjarto and Slovak Economy Minister Denisa Sakova jointly wrote to the Croatian government in Zagreb with the request, Szijjarto said in a statement Sunday. Oil transit along the Druzhba pipeline via Ukraine has been halted since late last month amid large-scale Russian attacks on Ukraine’s energy infrastructure, with the governments in Budapest and Kyiv in a standoff over the fallout. Budapest relies on the Druzhba pipeline connecting Hungary with Russia through war-torn Ukraine for most of its oil flows. Hungarian Prime Minister Viktor Orban, who has remained committed to buying Russian energy sources for his landlocked country, has also frequently engaged in debate with neighboring Croatia over the capacity of the Adriatic pipeline.  Energy policy is also likely to feature in Orban’s talks in Budapest with US Secretary of State Marco Rubio on Monday. Orban has found an ally in Slovak counterpart Robert Fico, who on Sunday echoed his views that Ukraine was using the Druzhba pipeline for political leverage, which officials in Kyiv have denied. What do you think? We’d love to hear from you, join the conversation on the Rigzone Energy Network. The Rigzone Energy Network is a new social experience created for you and all energy professionals to Speak Up about our industry, share knowledge, connect with peers and industry insiders and engage in a professional community that will empower your career in energy.

Read More »

Arista laments ‘horrendous’ memory situation

Digging in on campus Arista has been clear about its plans to grow its presence campus networking environments. Last Fall, Ullal said she expects Arista’s campus and WAN business would grow from the current $750 million-$800 million run rate to $1.25 billion, representing a 60% growth opportunity for the company. “We are committed to our aggressive goal of $1.25 billion for ’26 for the cognitive campus and branch. We have also successfully deployed in many routing edge, core spine and peering use cases,” Ullal said. “In Q4 2025, Arista launched our flagship 7800 R4 spine for many routing use cases, including DCI, AI spines with that massive 460 terabits of capacity to meet the demanding needs of multiservice routing, AI workloads and switching use cases. The combined campus and routing adjacencies together contribute approximately 18% of revenue.” Ethernet leads the way “In terms of annual 2025 product lines, our core cloud, AI and data center products built upon our highly differentiated Arista EOS stack is successfully deployed across 10 gig to 800 gigabit Ethernet speeds with 1.6 terabit migration imminent,” Ullal said. “This includes our portfolio of EtherLink AI and our 7000 series platforms for best-in-class performance, power efficiency, high availability, automation, agility for both the front and back-end compute, storage and all of the interconnect zones.” Ullal said she expects Ethernet will get even more of a boost later this year when the multivendor Ethernet for Scale-Up Networking (ESUN) specification is released.  “We have consistently described that today’s configurations are mostly a combination of scale out and scale up were largely based on 800G and smaller ratings. Now that the ESUN specification is well underway, we need a good solid spec. Otherwise, we’ll be shipping proprietary products like some people in the world do today. And so we will tie our

Read More »

From NIMBY to YIMBY: A Playbook for Data Center Community Acceptance

Across many conversations at the start of this year, at PTC and other conferences alike, the word on everyone’s lips seems to be “community.” For the data center industry, that single word now captures a turning point from just a few short years ago: we are no longer a niche, back‑of‑house utility, but a front‑page presence in local politics, school board budgets, and town hall debates. That visibility is forcing a choice in how we tell our story—either accept a permanent NIMBY-reactive framework, or actively build a YIMBY narrative that portrays the real value digital infrastructure brings to the markets and surrounding communities that host it. Speaking regularly with Ilissa Miller, CEO of iMiller Public Relations about this topic, there is work to be done across the ecosystem to build communications. Miller recently reflected: “What we’re seeing in communities isn’t a rejection of digital infrastructure, it’s a rejection of uncertainty driven by anxiety and fear. Most local leaders have never been given a framework to evaluate digital infrastructure developments the way they evaluate roads, water systems, or industrial parks. When there’s no shared planning language, ‘no’ becomes the safest answer.” A Brief History of “No” Community pushback against data centers is no longer episodic; it has become organized, media‑savvy, and politically influential in key markets. In Northern Virginia, resident groups and environmental organizations have mobilized against large‑scale campuses, pressing counties like Loudoun and Prince William to tighten zoning, question incentives, and delay or reshape projects.1 Loudoun County’s move in 2025 to end by‑right approvals for new facilities, requiring public hearings and board votes, marked a watershed moment as the world’s densest data center market signaled that communities now expect more say over where and how these campuses are built. Prince William County’s decision to sharply increase its tax rate on

Read More »

Nomads at the Frontier: PTC 2026 Signals the Digital Infrastructure Industry’s Moment of Execution

Each January, the Pacific Telecommunications Council conference serves as a barometer for where digital infrastructure is headed next. And according to Nomad Futurist founders Nabeel Mahmood and Phillip Koblence, the message from PTC 2026 was unmistakable: The industry has moved beyond hype. The hard work has begun. In the latest episode of The DCF Show Podcast, part of our ongoing ‘Nomads at the Frontier’ series, Mahmood and Koblence joined Data Center Frontier to unpack the tone shift emerging across the AI and data center ecosystem. Attendance continues to grow year over year. Conversations remain energetic. But the character of those conversations has changed. As Mahmood put it: “The hype that the market started to see is actually resulting a bit more into actions now, and those conversations are resulting into some good progress.” The difference from prior years? Less speculation. More execution. From Data Center Cowboys to Real Deployments Koblence offered perhaps the sharpest contrast between PTC conversations in 2024 and those in 2026. Two years ago, many projects felt speculative. Today, developers are arriving with secured power, customers, and construction underway. “If 2024’s PTC was data center cowboys — sites that in someone’s mind could be a data center — this year was: show me the money, show me the power, give me accurate timelines.” In other words, the market is no longer rewarding hypothetical capacity. It is demanding delivered capacity. Operators now speak in terms of deployments already underway, not aspirational campuses still waiting on permits and power commitments. And behind nearly every conversation sits the same gating factor. Power. Power Has Become the Industry’s Defining Constraint Whether discussions centered on AI factories, investment capital, or campus expansion, Mahmood and Koblence noted that every conversation eventually returned to energy availability. “All of those questions are power,” Koblence said.

Read More »

Cooling Consolidation Hits AI Scale: LiquidStack, Submer, and the Future of Data Center Thermal Strategy

As AI infrastructure scales toward ever-higher rack densities and gigawatt-class campuses, cooling has moved from a technical subsystem to a defining strategic issue for the data center industry. A trio of announcements in early February highlights how rapidly the cooling and AI infrastructure stack is consolidating and evolving: Trane Technologies’ acquisition of LiquidStack; Submer’s acquisition of Radian Arc, extending its reach from core data centers into telco edge environments; and Submer’s partnership with Anant Raj to accelerate sovereign AI infrastructure deployment across India. Layered atop these developments is fresh guidance from Oracle Cloud Infrastructure explaining why closed-loop, direct-to-chip cooling is becoming central to next-generation facility design, particularly in regions where water use has become a flashpoint in community discussions around data center growth. Taken together, these developments show how the industry is moving beyond point solutions toward integrated, scalable AI infrastructure ecosystems, where cooling, compute, and deployment models must work together across hyperscale campuses and distributed edge environments alike. Trane Moves to Own the Cooling Stack The most consequential development comes from Trane Technologies, which on February 10 announced it has entered into a definitive agreement to acquire LiquidStack, one of the pioneers and leading innovators in data center liquid cooling. The acquisition significantly strengthens Trane’s ambition to become a full-service thermal partner for data center operators, extending its reach from plant-level systems all the way down to the chip itself. LiquidStack, headquartered in Carrollton, Texas, built its reputation on immersion cooling and advanced direct-to-chip liquid solutions supporting high-density deployments across hyperscale, enterprise, colocation, edge, and blockchain environments. Under Trane, those technologies will now be scaled globally and integrated into a broader thermal portfolio. In practical terms, Trane is positioning itself to deliver cooling across the full thermal chain, including: • Central plant equipment and chillers.• Heat rejection and controls

Read More »

Infrastructure Maturity Defines the Next Phase of AI Deployment

The State of Data Infrastructure Global Report 2025 from Hitachi Vantara arrives at a moment when the data center industry is undergoing one of the most profound structural shifts in its history. The transition from enterprise IT to AI-first infrastructure has moved from aspiration to inevitability, forcing operators, developers, and investors to confront uncomfortable truths about readiness, resilience, and risk. Although framed around “AI readiness,” the report ultimately tells an infrastructure story: one that maps directly onto how data centers are designed, operated, secured, and justified economically. Drawing on a global survey of more than 1,200 IT leaders, the report introduces a proprietary maturity model that evaluates organizations across six dimensions: scalability, reliability, security, governance, sovereignty, and sustainability. Respondents are then grouped into three categories—Emerging, Defined, and Optimized—revealing a stark conclusion: most organizations are not constrained by access to AI models or capital, but by the fragility of the infrastructure supporting their data pipelines. For the data center industry, the implications are immediate, shaping everything from availability design and automation strategies to sustainability planning and evolving customer expectations. In short, extracting value from AI now depends less on experimentation and more on the strength and resilience of the underlying infrastructure. The Focus of the Survey: Infrastructure, Not Algorithms Although the report is positioned as a study of AI readiness, its primary focus is not models, training approaches, or application development, but rather the infrastructure foundations required to operate AI reliably at scale. Drawing on responses from more than 1,200 organizations, Hitachi Vantara evaluates how enterprises are positioned to support production AI workloads across six dimensions as stated above: scalability, reliability, security, governance, sovereignty, and sustainability. These factors closely reflect the operational realities shaping modern data center design and management. The survey’s central argument is that AI success is no longer

Read More »

AI’s New Land Grab: Meta’s Indiana Megaproject and the Rise of Europe’s Neocloud Challengers

While Meta’s Indiana campus anchors hyperscale expansion in the United States, Europe recorded its own major infrastructure milestone this week as Amsterdam-based AI infrastructure provider Nebius unveiled plans for a 240-megawatt data center campus in Béthune, France, near Lille in the country’s northern industrial corridor. When completed, the campus will rank among Europe’s largest AI-focused data center facilities and positions northern France as a growing node in the continent’s expanding AI infrastructure map. The development repurposes a former Bridgestone tire manufacturing site, reflecting a broader trend across Europe in which legacy industrial properties, already equipped with heavy power access, transport links, and industrial zoning, are being converted into large-scale digital infrastructure hubs. Located within reach of connectivity and enterprise corridors linking Paris, Brussels, London, and Amsterdam, the site allows Nebius to serve major European markets while avoiding the congestion and power constraints increasingly shaping Tier 1 data center hubs. Industrial Infrastructure Becomes Digital Infrastructure Developers increasingly view former industrial sites as ideal for AI campuses because they often provide: • Existing grid interconnection capacity built for heavy industry• Transport and logistics infrastructure already in place• Industrial zoning that reduces permitting friction• Large contiguous parcels suited to phased campus expansion For regions like Hauts-de-France, redevelopment projects also offer economic transition opportunities, replacing legacy manufacturing capacity with next-generation digital infrastructure investment. Local officials have positioned the project as part of broader efforts to reposition northern France as a logistics and technology hub within Europe. The Neocloud Model Gains Ground Beyond the site itself, Nebius’ expansion illustrates the rapid emergence of neocloud infrastructure providers, companies building GPU-intensive AI capacity without operating full hyperscale cloud ecosystems. These firms increasingly occupy a strategic middle ground: supplying AI compute capacity to enterprises, startups, and even hyperscalers facing short-term infrastructure constraints. Nebius’ rise over the past year

Read More »

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.

Read More »

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

Read More »

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

Read More »

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

Read More »
More Articles