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 shifts IT roles from operator to orchestrator

The report indicates that IT roles are becoming more strategic and automation-driven, with 52% of respondents citing increases in both areas. Roles are also becoming more cross-functional (47%) and complex (41%), reflecting the integration of AI into broader business processes. AI is also affecting how IT teams allocate time. Respondents

Read More »

Apply Now: 2026 Waste to Energy and Materials Technical Assistance for State, Local, and Tribal Governments

The U.S. Department of Energy’s Alternative Fuels and Feedstocks Office (AFFO), formerly known as the Bioenergy Technologies Office, and the National Laboratory of the Rockies (NLR) are launching the 2026 Waste to Energy and Materials Technical Assistance Program for state, local, and Tribal governments. The scope of this year’s program has been expanded to include additional municipal solid waste materials such as electronics, industrial wastewater, and other byproducts.  U.S. waste streams present significant logistical and economic challenges for states, counties, municipalities, and Tribal governments. However, waste is also a resource that can be used as an unconventional additional source of energy, advanced materials, and critical minerals. This program provides no-cost technical assistance to states, counties, municipalities, and Tribal governments with the most relevant data to guide decision-making—providing local solutions to the various aspects of waste management, taking into consideration current handling practices, costs, and infrastructure. It is designed to help officials evaluate the most sensible end uses for their waste, whether repurposing it for on-site heat and power, upgrading it into transportation fuels, or using it for material and mineral recovery. Program technical assistance includes: Waste resource information Infrastructure considerations Techno-economic comparison of energy, material, and mineral recovery options Evaluation and sharing of case studies (to the extent possible) from similar communities/projects The 2026 Waste to Energy and Materials Technical Assistance application portal is now open and applications will be accepted through May 30, 2026. For information on applicant eligibility and how to apply, please visit NLR’s technical assistance webpage. Timeline for Technical Assistance Opportunity Date Action April 15, 2026 Application Portal Opens May 30, 2026 Application Portal Closes  July – August 2026 Selections Made and Recipients Informed  Learn more about AFFO-supported waste to energy and materials technical assistance. If you have further questions, please see frequently asked questions or contact the Waste to

Read More »

Energy Deputy Secretary Danly Commends FERC Action on Large Load Interconnection Reform

WASHINGTON—U.S. Deputy Secretary of Energy James P. Danly issued the following statement after the Federal Energy Regulatory Commission (FERC or Commission) announced it will take action by June 2026 on the large load interconnection proceeding initiated at the direction of U.S. Secretary of Energy Chris Wright: “FERC’s announcement today demonstrates Chairman Swett’s commitment to implement Secretary Wright’s directive that the Commission ensure the timely and orderly integration of large electric loads that deliver on President Trump’s goal of American energy dominance. “I expect that the Commission will act quickly and decisively to improve interconnection processes, support the co-location of load and generation, and accelerate the addition of new generation to ensure that supply is built alongside demand—delivering affordable, reliable, and secure energy for all Americans. “Having served at FERC as commissioner and chairman, I understand FERC’s role in ensuring the reliability of the nation’s bulk power system, and I commend Chairman Swett for focusing on affordability and reliability.”                                                                                               ###  

Read More »

Petrobras discovers hydrocarbons in Campos basin presalt offshore Brazil

@import url(‘https://fonts.googleapis.com/css2?family=Inter:[email protected]&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; } Petrobras has discovered presence in the Campos basin presalt offshore Brazil during exploration in sector SC-AP4, block CM-477. Samples taken from the well, 1-BRSA-1404DC-RJS, will be sent for laboratory analysis with the aim of characterizing the conditions of the reservoirs and fluids found to enable continued evaluation of the area’s potential, the company said in a release Apr. 13. The discovery well was drilled 201 km off the coast of the state of Rio de Janeiro in water depth of 2,984 m. The hydrocarbon-bearing interval was confirmed through electrical profiles, gas evidence, and fluid sampling. Petrobras is the operator of block CM-477 with 70% interest. bp plc holds the remaining 30%.

Read More »

bp to operate blocks offshore Namibia through acquisition

@import url(‘https://fonts.googleapis.com/css2?family=Inter:[email protected]&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; } Map from bp plc <!–> –> bp plc aims to become operator of three exploration blocks offshore Namibia through acquisition of a 60% interest from Eco Atlantic Oil & Gas. Subject to Namibian government and joint venture partner approvals, bp will operate blocks PEL97, PEL99, and PEL100 in Walvis basin.   In a release Apr. 13, bp said entering the blocks builds on its recent exploration successes in Namibia through Azule Energy, a 50-50 joint venture between bp and Eni. Eco Atlantic will remain a partner, along with Namibia’s national oil company NAMCOR, following the deal’s closing, which is subject to closing conditions.

Read More »

ConocoPhillips sends team to Venezuela to evaluate oil, gas opportunities

ConocoPhillips sent a team to Venezuela to evaluate oil and gas opportunities, the company confirmed to Oil & Gas Journal Apr. 13. In an email to OGJ, a company spokesperson said “ConocoPhillips can confirm that we sent a small evaluation team to Venezuela during the week of Apr. 6 to better understand the potential for in-country oil and gas opportunities.” Asked what clarity the company seeks, the spokesperson said the team “will evaluate Venezuela against other international opportunities as part of our disciplined investment framework.” The operator left Venezuela in 2007 after then-President Hugo Chavez’s government reverted privately run oil fields to state control. ConocoPhillips, along with ExxonMobil, refused the government’s terms and took claims to the World Bank’s International Centre for the Settlement of Investment Disputes (ICSID). ConocoPhillips is owed about $12 billion following two judgements, an amount still sought by the company, which, prior to the expropriation of its interests, held a 50.1% interest in Petrozuata, a 40% interest in Hamaca, and a 32.5% interest in Corocoro heavy oil projects in Venezuela. In January, following the removal of Venezuela’s leader Nicolas Maduro, US President Donald Trump urged oil and gas companies to spend billions to rebuild Venezuela’s energy sector. ExxonMobil, which also exited the country in 2007, ​sent a technical team to Venezuela in March to ⁠evaluate the infrastructure and investment opportunities. In a discussion at CERAWeek by S&P Global in Houston in March, ConocoPhillips’ chief executive officer, Ryan Lance, said Venezuela needs to “completely rewire” ​its fiscal system to attract new ‌investment. The South American country holds a large cache of proven oil reserves, but has faced decades of production challenges due to mismanagement, underinvestment, and sanctions.

Read More »

TotalEnergies, TPAO sign MoU to assess exploration opportunities

@import url(‘https://fonts.googleapis.com/css2?family=Inter:[email protected]&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; } TotalEnergies EP New Ventures SA has signed a memorandum of understanding (MoU) with Türkiye Petrolleri Anonim Ortaklığı (TPAO) for potential collaboration. The MoU provides a framework for technical collaboration, including a joint assessment of hydrocarbon exploration opportunities in the Black Sea region of Türkiye as well as internationally. In February of this year, TPAO signed an MoU with Chevron Business Development EMEA Ltd., a subsidiary of Chevron, providing an opportunity to “identify and evaluate cooperation opportunities that may arise in international projects and in oil exploration and production license areas in onshore and offshore fields in Türkiye.”

Read More »

Equinix offering targets automated AI-centric network operations

Another component, Fabric Application Connect, functions as a private, dedicated connectivity marketplace for AI services. It lets enterprises access inference, training, storage, and security providers over private connections, bypassing the public Internet and limiting data exposure during AI development and deployment. Operational visibility is provided through Fabric Insights, an AI-powered monitoring layer that analyzes real-time network telemetry to detect anomalies and predict potential issues before they impact workloads. Fabric Insights integrates with security information and event management (SIEM) platforms such as Splunk and Datadog and feeds data directly into Fabric Super-Agent to support automated remediation. Fabric Intelligence operates on top of Equinix’s global infrastructure footprint, which includes hundreds of data centers across dozens of metropolitan markets. The platform is positioned as part of Equinix Fabric, a connectivity portfolio used by thousands of customers worldwide to link cloud providers, enterprises, and network services. Fabric Intelligence is available now to preview.

Read More »

Blue Owl Builds a Capital Platform for the Hyperscale AI Era

Capital as a Service: The Hyperscaler Shift This is not just another project financing. It points to a model in which hyperscalers can externalize a significant portion of the capital required for AI campuses while retaining operational control. Under the Hyperion structure, Meta provides construction and property management, while Blue Owl supplies capital at scale alongside infrastructure expertise. Reuters described the transaction as Meta’s largest private capital deal to date, with the campus projected to exceed 2 gigawatts of capacity. For Blue Owl, it marks a shift in role: from backing developers serving hyperscalers to working directly with a hyperscaler to structure ownership more efficiently at scale. Hyperion also helps explain why this model is gaining traction. Hyperscalers are now deploying capital at a pace that makes flexibility a strategic priority. Structures like the Meta–Blue Owl JV allow them to continue expanding infrastructure without fully absorbing the balance-sheet impact of each new campus. Analyst commentary cited by Reuters suggested the arrangement could help Meta mitigate risk and avoid concentrating too much capital in land, buildings, and long-lived infrastructure, preserving capacity for additional facilities and ongoing AI investment. That is the service Blue Owl is effectively providing. Not just capital, but balance-sheet flexibility at a time when AI infrastructure demand is stretching even the largest technology companies. With major tech firms projected to spend hundreds of billions annually on AI infrastructure, that capability is becoming central to how the next generation of campuses gets built. The Capital Baseline Resets In early 2026, hyperscalers effectively reset the capital baseline for the sector. Alphabet projected $175 billion to $185 billion in annual capex, citing continued constraints across servers, data centers, and networking. Amazon pointed to roughly $200 billion, up from $131 billion the prior year, while noting persistent demand pressure in AWS. Meta

Read More »

OpenAI pulls out of a second Stargate data center deal

“OpenAI is embattled on several fronts. Anthropic has been doing very well in the enterprise, and OpenAI’s cash burn might be a problem if it wants to go public at an astronomical $800 billion+ valuation. This is especially true with higher energy prices due to geopolitics, and the public and regulators increasingly skeptical of AI companies, especially outside of the United States,” Roberts said. “I see these moves as OpenAI tightening its belt a bit and being more deliberate about spending as it moves past the interesting tech demo stage of its existence and is expected to provide a real return for investors.” He added, “I expect it’s a symptom of a broader problem, which is that OpenAI has thrown some good money after bad in bets that didn’t work out, like the Sora platform it just shut down, and it’s under increasing pressure to translate its first-mover advantage into real upside for its investors. Spending operational money instead of capital money might give it some flexibility in the short term, and perhaps that’s what this is about.” All in all, he noted, “on a scale of business-ending event to nothingburger, I would put it somewhere in the middle, maybe a little closer to nothingburger.” Acceligence CIO Yuri Goryunov agreed with Roberts, and said, “OpenAI has a problem with commercialization and runaway operating costs, for sure. They are trying to rightsize their commitments and make sure that they deliver on their core products before they run out of money.” Goryunov described OpenAI’s arrangement with Microsoft in Norway as “prudent financial engineering” that allows it to access the data center resources without having to tie up too much capital. “It’s financial discipline. OpenAI [executives] are starting to behave like grownups.” Forrester senior analyst Alvin Nguyen echoed those thoughts. 

Read More »

DCF Tours: SDC Manhattan, 375 Pearl St.

Power: Redundant utility design in a power-constrained market The tour made equally clear that in Manhattan, power is still the central gating factor. The brochure describes SDC Manhattan as offering 18MW of aggregate power delivered to the building, backed by redundant electrical and mechanical systems, backup generators, and Tier III-type concurrent maintainability. The December 2025 press release updated that picture in a more market-facing way, noting that Sabey is one of the only colocation providers in Manhattan with available power, including nearly a megawatt of turnkey power and 7MW of utility power across two powered shell spaces. Bajrushi’s explanation of the electrical topology helped show how Sabey has made that possible. Standing on the third floor, he described a ring bus tying together four Con Edison feeds. Bajrushi said the feeds all originate from the same substation but take different paths into the building, creating redundancy outside the building as well as within it. He added that if one feed fails, the ring bus remains unaffected, and that only one feed is needed to power everything currently in operation. He also noted that Sabey has the ability to add two more feeds in the future if expansion calls for it. That matters in a city where available utility capacity is hard to come by and where many data center conversations end not with square footage but with a megawatt number. Bajrushi also noted that physical space is not the core constraint at 375 Pearl. He said the building still has plenty of room for future buildouts, including open areas that could become additional white space, chiller capacity, or other infrastructure. The bigger question, he suggested, is how and when power and supporting systems get installed. That observation aligns neatly with Sabey’s press release. The company is effectively arguing that SDC

Read More »

Maine to put brakes on big data centers as AI expansion collides with power limits

Mills has pushed for an exemption protecting a proposed $550 million project at the former Androscoggin paper mill in Jay, arguing it would reuse existing infrastructure without straining the grid. Lawmakers rejected that exemption. Mills’ office did not immediately respond to a request for comment. A national wave, an unanswered federal question Maine is one of at least 12 states now weighing moratorium or restraint legislation, alongside more than 300 data center bills filed across 30-plus states in the current session, according to legislative tracking firm MultiState. The shared concern is energy cost. Data centers could consume up to 12% of total US electricity by 2028, according to the US Department of Energy. On March 25, Senator Bernie Sanders and Alexandria Ocasio-Cortez introduced the AI Data Center Moratorium Act in Congress, which would impose a nationwide freeze on all new data center construction until Congress passes AI safety legislation. The Trump administration has pursued a different path from the legislative approach being taken in states. On March 4, Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed the White House’s Ratepayer Protection Pledge, a voluntary commitment by hyperscalers to fund their own power generation rather than pass grid costs to ratepayers. The pledge, published in the Federal Register on March 9, carries no penalties for noncompliance or auditing requirements.

Read More »

Cisco just made two moves to own the AI infrastructure stack

In a world of autonomous agents, identity and access become the de facto safety rails. Astrix is designed to inventory these non-human identities, map their permissions, detect toxic combinations, and remediate overprivileged access before it becomes an exploit or a data leak. That capability integrates directly with Cisco’s broader zero-trust and identity-centric security strategy, in which the network enforces policy based on who or what the entity is, not on which subnet it resides in. How this strengthens Cisco’s secure networking story Cisco has positioned itself as the vendor that can deliver “AI-ready, secure networks” spanning campus, data center, cloud, and edge. Galileo and Astrix extend that narrative from infrastructure into AI behavior and identity governance: The network becomes the high‑performance, policy‑enforcing substrate for AI traffic and data. Splunk plus Galileo becomes the observability plane for AI agents, linking AI incidents to network and application signals. Security plus Astrix becomes the identity and permission-control layer that constrains what AI agents can actually do within the environment. This is the core of Cisco’s emerging “Secure AI” posture: not just using AI to improve security but securing AI itself as it is embedded across every workflow, API, and device. For customers, that means AI initiatives can be brought under the same operational and compliance disciplines already used for networks and apps, rather than existing as unmanaged risk islands. Why this matters to Cisco customers Most large Cisco accounts are exactly the enterprises now experimenting with AI agents in contact centers, IT operations, and business workflows. They face three practical problems: They cannot see what agents are doing end‑to‑end, or measure quality beyond offline benchmarks. They lack a coherent model for managing the identities, secrets, and permissions those agents depend on. Their security and networking teams are often disconnected from AI projects happening in lines of business.

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