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.

ExxonMobil begins drilling wells in Guyana’s EEZ

bp completes sale of German refinery, associated assets

European independent refiner Klesch Group has completed its previously announced deal to acquire bp plc’s 265,000-b/d refinery and related assets in Gelsenkirchen and Horst and Scholven, Germany, which is operated as an integrated refining and petrochemical site. With the transaction finalized as of Aug. 3, Klesch has taken full ownership

Samsung offers future AI memory roadmap

It is already being used now in NAND flash memory for 3D stacking. Rather than spread the memory circuits out, they are stacked on top of each other like stories on a high-rise building. The technique was first introduced in 2014, with 24-layer NAND flash period last year it broke

AMD to buy Taalas, maker of model-specific AI chips for enterprise inference

Where model-specific silicon fits Those tradeoffs significantly narrow the range of enterprise workloads where model-specific silicon is likely to make economic sense. Dai sees the technology as best suited for mature, predictable inference workloads that run at massive scale and rely on relatively stable AI models, such as customer service

TotalEnergies to acquire Shell’s European onshore renewables portfolio

TotalEnergies SE has agreed to acquire Shell’s 4-Gw onshore renewables portfolio in Europe. The portfolio includes 500 Mw of solar and wind assets in operation or under construction, primarily in Italy and the Netherlands, as well as a 3.5-Gw pipeline of solar, wind, and battery storage projects in Italy, the UK, and Spain, the company said Aug. 3. In the Netherlands, the assets include 254.2 Mw of installed peak capacity across the Moerdijk, Heerenveen-Zuid, and Emmen (GZI Next) solar parks; the Sas van Gent-Zuid and Koegorspolder solar parks in Terneuzen; and the Pottendijk combined solar and wind park in Emmen. TotalEnergies will assume full ownership of the portfolio upon closing. The transaction is subject to regulatory approvals and is expected to be completed by yearend 2026. “This agreement reflects Shell’s continued focus on actively managing and further strengthening its electricity portfolio, in line with the strategy outlined during Capital Markets Day 2025,” said Machteld de Haan, president, downstream, renewables and energy solutions, Shell. De Haan said Shell is prioritizing investment in areas where it has competitive advantages, including asset-backed power trading and customer-focused energy solutions. Shell said it will continue to buy and sell onshore solar and wind power in Europe and will retain interests in projects including Holland Hydrogen 1, Northern Lights CCS in Norway, LNG, and carbon capture and storage activities. KKR acquires 50% interest in European renewables portfolio In another deal, TotalEnergies agreed to farm out a 50% interest in a largely developed 1.2-Gw onshore solar and wind portfolio in Europe to KKR. The company said the transaction is consistent with its strategy of selling 50% interests in renewable assets once they have been developed. The portfolio includes assets in Germany, Spain, France, and Poland. Electricity generated by the assets has already been sold to third parties or

Petrobras makes another gas discovery offshore Colombia

@import url(‘https://fonts.googleapis.com/css2?family=Inter:wght@100..900&display=swap’); .ebm-page__main h1, .ebm-page__main h2, .ebm-page__main h3, .ebm-page__main h4, .ebm-page__main h5, .ebm-page__main h6 { font-family: Inter; } body { line-height: 150%; letter-spacing: 0.025em; } button, .ebm-button-wrapper { font-family: Inter; } .label-style { text-transform: uppercase; color: var(–color-grey); font-weight: 600; font-size: 0.75rem; } .caption-style { font-size: 0.75rem; opacity: .6; } #onetrust-pc-sdk [id*=btn-handler], #onetrust-pc-sdk [class*=btn-handler] { background-color: #c19a06 !important; border-color: #c19a06 !important; } #onetrust-policy a, #onetrust-pc-sdk a, #ot-pc-content a { color: #c19a06 !important; } #onetrust-consent-sdk #onetrust-pc-sdk .ot-active-menu { border-color: #c19a06 !important; } #onetrust-consent-sdk #onetrust-accept-btn-handler, #onetrust-banner-sdk #onetrust-reject-all-handler, #onetrust-consent-sdk #onetrust-pc-btn-handler.cookie-setting-link { background-color: #c19a06 !important; border-color: #c19a06 !important; } #onetrust-consent-sdk .onetrust-pc-btn-handler { color: #c19a06 !important; border-color: #c19a06 !important; } Petroleo Brasileiro SA (Petrobras) discovered another new gas accumulation with the deepwater Sandia-1 exploratory well in GUA-OFF-0 block about 42 km offshore Colombia in 1,251 m of water. Sandia-1 spudded on June 12, 2026, reaching final depth on July 29, 2026. Proven gas intervals are being evaluated by means of well profiles and will later be characterized by laboratory analyses. The well is 18 km from the Sirius-1 (discoverer) and Sirius-2 (appraisal) wells and 9 km from the Copoazu-1 (discoverer) well, indicating that there is strong gas potential in this area of offshore Colombia, Petrobras said. Petrobras subsidiary Petrobras International Braspetro BV–Sucursal Colombia (PIB-COL; 44.44%) operates the GUA-OFF-0 block on behalf of partner Ecopetrol SA (55.56%).

OPEC+ approves September output hike, completes 2023 cuts rollback

OPEC+ has approved a fresh increase in oil production quotas for September of roughly 188,000 b/d, completing the phased reversal of voluntary supply cuts first introduced in 2023. The decision, confirmed in an official OPEC statement following a virtual meeting on Aug. 2, 2026, marks the sixth consecutive monthly increase by the group this year. Seven core members of the alliance—Saudi Arabia, Russia, Iraq, Kuwait, Kazakhstan, Algeria, and Oman—agreed to raise output targets. The move completes the unwinding of the 1.65-million b/d voluntary supply cut originally agreed in 2023, back when the group still included the United Arab Emirates (UAE), which exited OPEC in May. The group said the adjustment would also give participating countries an opportunity to accelerate compensation for previous overproduction, and it reiterated commitment to the OPEC+ Declaration of Cooperation, with compliance to be monitored by the Joint Ministerial Monitoring Committee (JMMC). While the September hike is now finalized, OPEC+ is widely expected to pause further increases starting in the fourth quarter. Though the group’s official statement gave no explicit guidance on fourth-quarter policy, OPEC+ sources cited by Reuters and analysts—including Rystad Energy’s Jorge Leon—say a pause is likely as the alliance assesses market conditions after finishing the restoration of the 2023 cuts. A separate layer of roughly 2 million b/d in cuts, dating to 2022, remains in place and is expected to continue through the end of 2026. The steady stream of monthly increases comes against a backdrop of major market disruption. Ongoing Middle East tensions—including disruptions tied to the Iran conflict and the Strait of Hormuz—have complicated the group’s ability to translate higher quotas into actual barrels reaching the market. Russia, in particular, continues to produce below its OPEC+ target of about 9.8 million b/d, with output near 9 million b/d amid repeated Ukrainian drone

Market Focus: Reading the oil market after the US-Iran MOU collapse

Drawing on nearly three decades of experience in energy trading and risk management, Kessler offers insight into the fallout from escalating Middle East tensions, the breakdown of US-Iran diplomatic efforts, and the critical role of the Strait of Hormuz, through which a significant share of global oil supplies traditionally flows. The discussion explores what it would take to achieve a meaningful de-escalation in the region and how market participants are assessing the risks. Kessler argues that restoring safe passage through the Strait of Hormuz will be central to any lasting stability, while Iran’s oil exports and broader economic pressures could influence future negotiations. He also shares his perspective on how OPEC+ is responding to disruptions, the alliance’s efforts to restore production, and the growing competitive pressure it faces from producers outside the Gulf region. Turning to North America, Kessler examines the outlook for US shale producers in a higher-price environment. With crude prices holding above $80/bbl, he discusses signs of increased drilling activity, stronger production growth potential, and the continued emphasis on hedging and capital discipline among operators. The conversation also highlights advances in drilling technology and efficiency that could enable US producers to respond more quickly to market opportunities while managing downside risk. Looking further ahead, the episode considers whether recent disruptions will accelerate a long-term shift away from traditional Middle East oil chokepoints. Kessler discusses the growing role of US, Canadian, African, and Latin American supplies, expanding export infrastructure, and the possibility that today’s high prices could ultimately lead to demand destruction, increased competition, and renewed market oversupply. For anyone following global crude markets, OPEC+ strategy, US shale growth, energy security, and future oil price trends, this conversation provides a timely and thought-provoking outlook on the evolving global energy landscape. About our guest Dennis Kissler, senior vice-president of

Maurel & Prom to acquire Gran Tierra Energy’s assets in Colombia, Ecuador for $1.33 billion

The company said the predominantly operated portfolio comprises producing assets, development projects, and exploration acreage across Colombia’s Middle Magdalena Valley, Putumayo, and Llanos basins and Ecuador’s Oriente basin. Production is entirely oil-weighted and benefits from established processing, storage, and transportation infrastructure as well as access to multiple export routes. The principal Colombian assets include Acordionero, Costayaco, and Moqueta on the Chaza block, the Suroriente block centered on Cohembi, and the recently acquired interests in Tisquirama and San Roque. Growth opportunities in Colombia include continued development of Tisquirama, expansion of the Cohembi-Raju area, the Pegasus prospect, and longer-term potential associated with the La Luna formation. In Ecuador, the Chanangue, Charapa, Conejo, Iguana, Perico, and Espejo assets provide a combination of producing fields, discovered resources, and appraisal and exploration opportunities. Maurel & Prom said the assets represent a growth platform supported by existing discoveries and additional potential through waterflood application across the portfolio. For Gran Tierra Energy, the transaction serves as an exit from South America as part of the company’s plan to reduce debt and focus on growth opportunities in Canada and Azerbaijan. Maurel & Prom is a Paris-listed international oil and natural gas exploration and production company majority owned by PT Pertamina Internasional Eksplorasi dan Produksi (PIEP), a subsidiary of Indonesia’s national energy company, PT Pertamina (Persero). Closing, expected by yearend, is subject to shareholder approval, creditor consents, regulatory approvals, and other customary closing conditions.

HF Sinclair inks supply deals amid pending segment spinoff, refinery closure

HF Sinclair Corp. has lined up long-term supply arrangements to support the transition of the company’s lubricants and specialties products business in parallel with its recently announced plan to retire its 15,600-b/d base oil refining plant in Mississauga, Ont. After revealing a downstream integration strategy on July 28 involving the proposed closure of its Canadian refining business and transformation of its lubricants and specialties products segment, HF Sinclair confirmed on Aug. 3 that it entered into strategic long-term commercial agreements with suppliers SK On Co. Ltd.’s SK Enmove and Chevron USA Inc.’s Chevron Products Co. to establish a diversified North American base oil supply network. As part of the August agreement that aims to support maintaining base oil coverage after the Canadian refining assets are retired, SK Enmove and Chevron Products will supply HF Sinclair with Group III and Group II base oils, respectively, according to the companies. In return, HF Sinclair said its lubricants and specialties business will serve as a distributor for SK Enmove’s YUBASE Group III base oils in key regional markets in North America, as well as distribute Chevron-branded Group II base oils in Canada and select US regions. The supply arrangements come as part of HF Sinclair’s transformation and separation of its lubricants and specialties segments via the capital markets to create “two independent public companies in a manner that is tax-efficient for HF Sinclair and its shareholders,” according to the operator’s July 28 presentation to investors. HF Sinclair said it expects the new lubricants-specialties company will to drive organic growth and consolidate a highly fragmented global lubricants and specialties market, leading to reduced earnings volatility supported by diversified end markets and a differentiated finished and specialty mix of products. Subject to customary conditions and final approvals, HF Sinclair said the separation of the lubricants-specialty

Polish data center plans to send its waste heat to the neighbors

As Europe swelters in a heatwave, residents probably don’t want to hear about ways to make their homes even hotter, but that’s what Polish property developer Citylink is talking about, with plans to dump waste heat from a new data center in Wrocław into the municipal district heating network. Citylink is designing the data center so that heat from servers can be recovered instead of being dissipated via cooling systems — and as the data center grows, any increase in computing power will mean more energy available for recovery. The collaboration with local power company Kogeneracja will provide “valuable experience in designing and operating modern data centers, with a particular focus on infrastructure dedicated to AI nodes,” said Michał Starybrat, development director at Citylink.

The Data Center Industry’s Permission to Build

The data center industry has spent the past several years announcing the future. Gigawatts. AI factories. New regions. New power architectures. Campuses at a scale that would have seemed extraordinary before generative AI reset the industry’s expectations. Now the public has entered the room. Communities are asking harder questions about who pays for electrical infrastructure, where the water comes from, how much noise reaches neighboring properties and what remains locally after construction crews leave. Utilities are being pressed to protect ratepayers from speculative load and costly system upgrades. Elected officials who once treated data centers primarily as economic-development wins are finding that the politics have changed. The defining question is no longer whether demand is real. It is whether the data center industry can keep earning the permission required to build at the scale it has promised. I mean permission in a broader sense than zoning approval, an environmental permit or a signed utility agreement. I mean the political and social room to develop infrastructure measured in hundreds of megawatts and billions of dollars—often in places whose residents have only recently begun to understand what is being proposed around them. That room is narrowing. A Different Kind of Constraint On July 18, opponents organized 142 demonstrations across 42 states in what Reuters described as the first coordinated national protest against the data center buildout. The movement crossed familiar political boundaries, bringing together environmental advocates, rural landowners and residents concerned about power prices, water, noise and the pace of development. A June Reuters/Ipsos poll found that 57% of respondents would oppose a data center in their community. Only 14% said they would be comfortable with one nearby. Those findings deserve the industry’s full attention. New York has imposed a one-year pause on certain environmental approvals for new hyperscale data centers while

NVIDIA’s Reported $50B Lease and the Nuclear-Powered AI Factory

Aalo and Crusoe Pursue the Nuclear-Powered AI Factory The Aalo-Crusoe partnership addresses the industry’s power problem by bringing power generation directly to the compute. In this case, skipping intermediary power stages such as minimal grid or custom BTM gas turbine solutions and going straight to nuclear. Aalo Atomics and Crusoe said they plan to deploy a Crusoe Spark modular data center running Crusoe Cloud at Idaho National Laboratory in 2027. The proof-of-concept project is intended to demonstrate an AI workload operating on power from an Aalo advanced reactor. Crusoe continues to expand their other data center campus projects. The companies then intend to deploy Aalo Pods, Aalo’s 50-megawatt-electric nuclear power plants, at Crusoe data centers by the end of 2029. Aalo has already begun work on a second reactor beside its initial test unit at the Idaho site. That reactor is expected to produce electricity for the Crusoe installation. On July 4, 2026, Aalo’s zero-power Critical Test Reactor reached criticality, sustaining a nuclear chain reaction without generating commercial electricity. The test reactor contains a full-scale core and components analogous to those planned for the 10-megawatt-electric Aalo-X power reactor being built next door, but it operates before sodium coolant and electricity-generating systems are added. Aalo plans to continue experiments with the Critical Test Reactor to refine its reactor-physics models, characterize control behavior and generate data supporting development and licensing of the full-power Aalo-X system. Advanced nuclear announcements sometimes blur the line between a successful test, an electricity-producing demonstration and a commercially licensed fleet. Aalo has achieved an important technical milestone, but substantial work remains before reactors can be manufactured, licensed, financed and operated at commercial data center sites. The pairing with Crusoe should be noted because it connects a reactor developer with a company that can provide the data center load,

Data Center Jobs: Engineering, Construction, Commissioning, Sales, Field Service and Facility Tech Jobs Available in Major Data Center Hotspots

Each month Data Center Frontier, in partnership with Pkaza, posts some of the hottest data center career opportunities in the market. Here’s a look at some of the latest data center jobs posted on the Data Center Frontier jobs board, powered by Pkaza Critical Facilities Recruiting. Looking for Data Center Candidates? Check out Pkaza’s Active Candidate / Featured Candidate Hotlist CFD Engineer – Data Center Mechanical DesignNew York, NY (remote)This position is also available as a remote role anywhere in the US in addition to key markets such as Cedar Rapids, IA; Kansas City, CA or White Plains, NY. Our client is an engineering design and commissioning company that has a national footprint and specializes in MEP critical facilities design. They provide design, commissioning, consulting and management expertise in the critical facilities space. They have a mindset to provide reliability, energy efficiency, and sustainable design expertise when providing these consulting services for enterprise, colocation and hyperscale companies. This career-growth minded opportunity offers exciting projects with leading-edge technology and innovation as well as competitive salaries and benefits. Electrical Commissioning Agent – Data CentersColumbus, OH (limited travel) Non-traveling CxA positions available in: Indianapolis, IN; Cedar Rapids, IA; Phoenix, AZ; Atlanta, GA and Austin, TX. Traveling CxA based really near any major airport, otherwise traveling to: New York, NY; White Plains, NY; Dallas, TX; Richmond, VA; Montvale, NJ; Charlotte, NC; Salt Lake City, UT; Kansas City, MO; Chesterton, IN or Chicago, IL. ***Also looking for a lead EE and ME CxA Agents and CxA PMs.*** This opportunity is with a leading EPC company of data center design / build / commissioning solutions. This company provides a complete life cycle of solutions that are custom-fit to the requirements of their client’s mission-critical facilities. This opportunity provides a career-growth minded role with exciting projects with leading-edge technology

Navigating Virginia’s Data Center Boom: Policy Shifts, Local Projects, and Future Challenges

Virginia’s newest high profile data center story is no longer the announcement of the next generation AI data center campus, it is now how the state is beginning to set the trend for legislative process to protect its communities while still encouraging the data center industry development. On August 3, state Senators Richard Stuart, a Republican, and Russet Perry, a Democrat, called on Gov. Abigail Spanberger to convene a special legislative session to address groundwater strain. Their request followed a state study warning that eastern Virginia’s groundwater supply is constrained and that large new industrial withdrawals may be difficult to sustain. The debate has expanded into calls for a broader pause: Senator Glen Sturtevant has asked for an immediate statewide moratorium on new data center development, while Senate President Pro Tempore Louise Lucas has said such a moratorium deserves serious consideration. Those proposals are not yet law, but they are the clearest indication that Virginia’s policy discussion has moved beyond incremental regulation. The Commonwealth spent years treating data centers primarily as an economic-development and tax-base success. It is now evaluating them simultaneously as power, water, land-use, air-quality and ratepayer issues. That shift is especially important for projects outside Northern Virginia, where developers are increasingly pursuing large sites in communities with less experience reviewing hyperscale infrastructure. The calls for a special session arrive only weeks after a significant package of data center laws and budget provisions took effect July 1. Virginia’s new budget established what the administration describes as a first-of-its-kind electricity consumption tax on data centers. The charge is 1.1 cents per kilowatt-hour, began July 1 and is capped at $600 million in annual collections, with excess revenue refunded to data center taxpayers. The compromise preserved Virginia’s sales-and-use-tax exemption for qualifying data center equipment, avoiding the abrupt repeal sought by

Land and Expand: The Gigawatt Credibility Test

The midsummer wave of U.S. data center development is not defined by a single market, developer or technology company. It stretches from the Georgia coast to West Texas, from the industrial Midwest to the Mississippi River. What links the projects announced since early June is not just their scale, it is the realization that scale alone is not enough. Developers are still announcing multibillion-dollar campuses and gigawatt power requirements, but the language surrounding those announcements has changed. Companies are emphasizing who will pay for new generation and transmission, how cooling systems will limit water consumption, what communities will receive beyond temporary construction employment, and when contracted customers will begin occupying capacity. In several cases, the announcement is less about acquiring land than proving that a project has become commercially and electrically credible. As we have seen progressing through the industry, the latest announcements point toward campuses that combine compute, power, financing and community agreements in one development package. OpenAI Goes Direct in Georgia OpenAI, on July 22 disclosed Project Camellia, a long-term data center development in Effingham County, Georgia. OpenAI said it is designing and developing the campus itself and has contracted with Georgia Power for 3.2 gigawatts of electricity, to be delivered in phases from 2028 through 2032. The project has been reported as a roughly $20 billion investment on approximately 1,400 acres, making it one of the largest individual data center proposals currently moving through the U.S. pipeline. Project Camellia is notable not only for its size but for OpenAI’s more direct role. The company has traditionally secured capacity through cloud providers and infrastructure partners. By taking responsibility for designing and developing the Georgia campus, OpenAI is signaling that control over power, schedule and facility design has become strategically important as AI companies compete for increasingly scarce large-scale

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