Thanks to the legal discovery process, Google’s antitrust trial with the Department of Justice has provided a fascinating glimpse into the future of ChatGPT.An internal OpenAI strategy document titled “ChatGPT: H1 2025 Strategy” describes the company’s aspiration to build an “AI super assistant that deeply understands you and is your interface to the internet.” Although the document is heavily redacted in parts, it reveals that OpenAI aims for ChatGPT to soon develop into much more than a chatbot. “In the first half of next year, we’ll start evolving ChatGPT into a super-assistant: one that knows you, understands what you care about, and helps with any task that a smart, trustworthy, emotionally intelligent person with a computer could do,” reads the document from late 2024. “The timing is right. Models like 02 and 03 are finally smart enough to reliably perform agentic tasks, tools like computer use can boost ChatGPT’s ability to take action, and interaction paradigms like multimodality and generative UI allow both ChatGPT and users to express themselves in the best way for the task.”The document goes on to describe a “super assistant” as “an intelligent entity with T-shaped skills” for both widely applicable and niche tasks. “The broad part is all about making life easier: answering a question, finding a home, contacting a lawyer, joining a gym, planning vacations, buying gifts, managing calendars, keeping track of todos, sending emails.” It mentions coding as an early example of a more niche task.Even when reading around the redactions, it’s clear that OpenAI sees hardware as essential to its future, and that it wants people to think of ChatGPT as not just a tool, but a companion. This tracks with Sam Altman recently saying that young people are using ChatGPT like a “ life advisor.”“Today, ChatGPT is in our lives through existing form factors — our website, phone, and desktop apps,” another part of the strategy document reads. “But our vision for ChatGPT is to help you with all of your life, no matter where you are. At home, it should help answer questions, play music, and suggest recipes. On the go, it should help you get to places, find the best restaurants, or catch up with friends. At work, it should help you take meeting notes, or prepare for the big presentation. And on solo walks, it should help you reflect and wind down.” At the same time, OpenAI finds itself in a wobbly position. Its infrastructure isn’t able to handle ChatGPT’s rising usage, which explains Altman’s focus on building data centers. In a section of the document describing AI chatbot competition, the company writes that “we are leading here, but we can’t rest,” and that “growth and revenue won’t line up forever.” It acknowledges that there are “powerful incumbents who will leverage their distribution to advantage their own products,” and states that OpenAI will advocate for regulation that requires other platforms to allow people to set ChatGPT as the default assistant. (Coincidentally, Apple is rumored to soon let iOS users also select Google’s Gemini for Siri queries. Meta AI just hit one billion users as well, thanks mostly to its many hooks in Instagram, WhatsApp, and Facebook.) “We have what we need to win: one of the fastest-growing products of all time, a category-defining brand, a research lead (reasoning, multimodal), a compute lead, a world-class research team, and an increasing number of effective people with agency who are motivated to ship,” the OpenAI document states. “We don’t rely on ads, giving us flexibility on what to build. Our culture values speed, bold moves, and self-disruption. Maintaining these advantages is hard work but, if we do, they will last for a while.”ElsewhereApple chickens out: For the first time in a decade, Apple won’t have its execs participate in John Gruber’s annual post-WWDC live podcast. Gruber recently wrote the viral “something is rotten in the state of Cupertino” essay, which was widely discussed in Apple circles. Although he hasn’t publicly connected that critical piece to the company backing out of his podcast, it’s easy to see the throughline. It says a lot about the state of Apple when its leaders don’t even want to participate in what has historically been a friendly forum.Elon was high: As Elon Musk attempts to reframe the public’s view of him by doing interviews about SpaceX, The New York Times reports that last year, he was taking so much ketamine that it “was affecting his bladder.” He also reportedly “traveled with a daily medication box that held about 20 pills, including ones with the markings of the stimulant Adderall.” Both Musk and the White House have had multiple opportunities to directly refute this report, and they have not. Now, Musk is at least partially stepping away from DOGE along with key lieutenants like Steve Davis. DOGE may be a failure based on Musk’s own stated hopes for spending cuts, but his closeness to Trump has certainly helped rescue X from financial ruin and grown SpaceX’s business. Now, the more difficult work begins: saving Tesla. Overheard“The way we do ranking is sacrosanct to us.” - Google CEO Sundar Pichai on Decoder, explaining why the company’s search results won’t be changed for President Trump or anyone else. “Compared to previous technology changes, I’m a little bit more worried about the labor impact… Yes, people will adapt, but they may not adapt fast enough.” - Anthropic CEO Dario Amodei on CNN raising the alarm about the technology he is developing. “Meta is a very different company than it was nine years ago when they fired me.” - Anduril founder Palmer Luckey telling Ashlee Vance why he is linking up with Mark Zuckerberg to make headsets for the military. Personnel logThe flattening of Meta’s AI organization has taken effect, with VP Ahmad Al-Dahle no longer overseeing the entire group. Now, he co-leads “AGI Foundations” with Amir Frenkel, VP of engineering, while Connor Hayes runs all AI products. All three men now report to Meta CPO Chris Cox, who has diplomatically framed the changes as a way to “give each org more ownership.”Xbox co-founder J Allard is leading a new ‘breakthrough’ devices group at Amazon called ZeroOne. One of the devices will be smart home-related, according to job listings.C.J. Mahoney, a former Trump administration official, is being promoted to general counsel at Microsoft, which has also hired Lisa Monaco from the last Biden administration to lead global policy. Reed Hastings is joining the board of Anthropic “because I believe in their approach to AI development, and to help humanity progress.” (He’s joining Anthropic’s corporate board, not the supervising board of its public benefit trust that can hire and fire corporate directors.)Sebastian Barrios, previously SVP at Mercado Libre, is joining Roblox as SVP of engineering for several areas, including ads, game discovery, and the company’s virtual currency work.Fidji Simo’s replacement at Instacart will be chief business officer Chris Rogers, who will become the company’s next CEO on August 15th after she officially joins OpenAI.Link listMore to click on:If you haven’t already, don’t forget to subscribe to The Verge, which includes unlimited access to Command Line and all of our reporting.As always, I welcome your feedback, especially if you have thoughts on this issue or a story idea to share. You can respond here or ping me securely on Signal.Thanks for subscribing.See More:

From large language models (LLMs) to reasoning agents, today’s AI tools bring unprecedented computational demands. Trillion-parameter models, workloads running on-device, and swarms of agents collaborating to complete tasks all require a new paradigm of computing to become truly seamless and ubiquitous. Silicon’s mid-life crisis AI has evolved from classical ML to deep learning to generative AI. The most recent chapter, which took AI mainstream, hinges on two phases—training and inference—that are data and energy-intensive in terms of computation, data movement, and cooling. At the same time, Moore’s Law, which determines that the number of transistors on a chip doubles every two years, is reaching a physical and economic plateau. For the last 40 years, silicon chips and digital technology have nudged each other forward—every step ahead in processing capability frees the imagination of innovators to envision new products, which require yet more power to run. That is happening at light speed in the AI age. As models become more readily available, deployment at scale puts the spotlight on inference and the application of trained models for everyday use cases. This transition requires the appropriate hardware to handle inference tasks efficiently. Central processing units (CPUs) have managed general computing tasks for decades, but the broad adoption of ML introduced computational demands that stretched the capabilities of traditional CPUs. This has led to the adoption of graphics processing units (GPUs) and other accelerator chips for training complex neural networks, due to their parallel execution capabilities and high memory bandwidth that allow large-scale mathematical operations to be processed efficiently. But CPUs are already the most widely deployed and can be companions to processors like GPUs and tensor processing units (TPUs). AI developers are also hesitant to adapt software to fit specialized or bespoke hardware, and they favor the consistency and ubiquity of CPUs. Chip designers are unlocking performance gains through optimized software tooling, adding novel processing features and data types specifically to serve ML workloads, integrating specialized units and accelerators, and advancing silicon chip innovations, including custom silicon. AI itself is a helpful aid for chip design, creating a positive feedback loop in which AI helps optimize the chips that it needs to run. These enhancements and strong software support mean modern CPUs are a good choice to handle a range of inference tasks. Beyond silicon-based processors, disruptive technologies are emerging to address growing AI compute and data demands. The unicorn start-up Lightmatter, for instance, introduced photonic computing solutions that use light for data transmission to generate significant improvements in speed and energy efficiency. Quantum computing represents another promising area in AI hardware. While still years or even decades away, the integration of quantum computing with AI could further transform fields like drug discovery and genomics. The developments in ML theories and network architectures have significantly enhanced the efficiency and capabilities of AI models. Today, the industry is moving from monolithic models to agent-based systems characterized by smaller, specialized models that work together to complete tasks more efficiently at the edge—on devices like smartphones or modern vehicles. This allows them to extract increased performance gains, like faster model response times, from the same or even less compute. Researchers have developed techniques, including few-shot learning, to train AI models using smaller datasets and fewer training iterations. AI systems can learn new tasks from a limited number of examples to reduce dependency on large datasets and lower energy demands. Optimization techniques like quantization, which lower the memory requirements by selectively reducing precision, are helping reduce model sizes without sacrificing performance.  New system architectures, like retrieval-augmented generation (RAG), have streamlined data access during both training and inference to reduce computational costs and overhead. The DeepSeek R1, an open source LLM, is a compelling example of how more output can be extracted using the same hardware. By applying reinforcement learning techniques in novel ways, R1 has achieved advanced reasoning capabilities while using far fewer computational resources in some contexts. The integration of heterogeneous computing architectures, which combine various processing units like CPUs, GPUs, and specialized accelerators, has further optimized AI model performance. This approach allows for the efficient distribution of workloads across different hardware components to optimize computational throughput and energy efficiency based on the use case. As AI becomes an ambient capability humming in the background of many tasks and workflows, agents are taking charge and making decisions in real-world scenarios. These range from customer support to edge use cases, where multiple agents coordinate and handle localized tasks across devices. With AI increasingly used in daily life, the role of user experiences becomes critical for mass adoption. Features like predictive text in touch keyboards, and adaptive gearboxes in vehicles, offer glimpses of AI as a vital enabler to improve technology interactions for users. Edge processing is also accelerating the diffusion of AI into everyday applications, bringing computational capabilities closer to the source of data generation. Smart cameras, autonomous vehicles, and wearable technology now process information locally to reduce latency and improve efficiency. Advances in CPU design and energy-efficient chips have made it feasible to perform complex AI tasks on devices with limited power resources. This shift toward heterogeneous compute enhances the development of ambient intelligence, where interconnected devices create responsive environments that adapt to user needs. Seamless AI naturally requires common standards, frameworks, and platforms to bring the industry together. Contemporary AI brings new risks. For instance, by adding more complex software and personalized experiences to consumer devices, it expands the attack surface for hackers, requiring stronger security at both the software and silicon levels, including cryptographic safeguards and transforming the trust model of compute environments. More than 70% of respondents to a 2024 DarkTrace survey reported that AI-powered cyber threats significantly impact their organizations, while 60% say their organizations are not adequately prepared to defend against AI-powered attacks. Collaboration is essential to forging common frameworks. Universities contribute foundational research, companies apply findings to develop practical solutions, and governments establish policies for ethical and responsible deployment. Organizations like Anthropic are setting industry standards by introducing frameworks, such as the Model Context Protocol, to unify the way developers connect AI systems with data. Arm is another leader in driving standards-based and open source initiatives, including ecosystem development to accelerate and harmonize the chiplet market, where chips are stacked together through common frameworks and standards. Arm also helps optimize open source AI frameworks and models for inference on the Arm compute platform, without needing customized tuning.  How far AI goes to becoming a general-purpose technology, like electricity or semiconductors, is being shaped by technical decisions taken today. Hardware-agnostic platforms, standards-based approaches, and continued incremental improvements to critical workhorses like CPUs, all help deliver the promise of AI as a seamless and silent capability for individuals and businesses alike. Open source contributions are also helpful in allowing a broader range of stakeholders to participate in AI advances. By sharing tools and knowledge, the community can cultivate innovation and help ensure that the benefits of AI are accessible to everyone, everywhere. Learn more about Arm’s approach to enabling AI everywhere. This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. This content was researched, designed, and written entirely by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Jesse Armstrong loves to pull fictional stories out of reality. His universally acclaimed TV show Succession, for instance, was inspired by real-life media dynasties like the Murdochs and the Hearsts. Similarly, his newest film Mountainhead centers upon characters that share key traits with the tech world’s most powerful leaders: Elon Musk, Mark Zuckerberg, Sam Altman, and others.Mountainhead, which releases on HBO on May 31 at 8 p.m. ET, portrays four top tech executives who retreat to a Utah hideaway as the AI deepfake tools newly released by one of their companies wreak havoc across the world. As the believable deepfakes inflame hatred on social media and real-world violence, the comfortably-appointed quartet mulls a global governmental takeover, intergalactic conquest and immortality, before interpersonal conflict derails their plans.Armstrong tells TIME in a Zoom interview that he first became interested in writing a story about tech titans after reading books like Michael Lewis’ Going Infinite (about Sam Bankman-Fried) and Ashlee Vance’s Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future, as well as journalistic profiles of Peter Thiel, Marc Andreessen, and others. He then built the story around the interplay between four character archetypes—the father, the dynamo, the usurper, and the hanger-on—and conducted extensive research so that his fictional executives reflected real ones. His characters, he says, aren’t one-to-one matches, but “Frankenstein monsters with limbs sewn together.” These characters are deeply flawed and destructive, to say the least. Armstrong says he did not intend for the film to be a wholly negative depiction of tech leaders and AI development. “I do try to take myself out of it, but obviously my sense of what this tech does and could do infuses the piece. Maybe I do have some anxieties,” he says. Armstrong contends that the film is more so channeling fears that AI leaders themselves have warned about. “If somebody who knows the technology better than anyone in the world thinks there's a 1/5th chance that it's going to wipe out humanity—and they're some of the optimists—I think that's legitimately quite unnerving,” he says. Here’s how each of the characters in Mountainhead resembles real-world tech leaders. This article contains spoilers. Venis (Cory Michael Smith) is the dynamo.Cory Michael Smith in Mountainhead Macall Polay—HBOVenis is Armstrong’s “dynamo”: the richest man in the world, who has gained his wealth from his social media platform Traam and its 4 billion users. Venis is ambitious, juvenile, and self-centered, even questioning whether other people are as real as him and his friends. Venis’ first obvious comp is Elon Musk, the richest man in the real world. Like Musk, Venis is obsessed with going to outer space and with using his enormous war chest to build hyperscale data centers to create powerful anti-woke AI systems. Venis also has a strange relationship with his child, essentially using it as a prop to help him through his own emotional turmoil. Throughout the movie, others caution Venis to shut down his deepfake AI tools which have led to military conflict and the desecration of holy sites across the world. Venis rebuffs them and says that people just need to adapt to technological changes and focus on the cool art being made. This argument is similar to those made by Sam Altman, who has argued that OpenAI needs to unveil ChatGPT and other cutting-edge tools as fast as possible in order to show the public the power of the technology. Like Mark Zuckerberg, Venis presides over a massively popular social media platform that some have accused of ignoring harms in favor of growth. Just as Amnesty International accused Meta of having “substantially contributed” to human rights violations perpetrated against Myanmar’s Rohingya ethnic group, Venis complains of the UN being “up his ass for starting a race war.”Randall (Steve Carell) is the father.Steve Carell in Mountainhead Macall Polay—HBOThe group’s eldest member is Randall, an investor and technologist who resembles Marc Andreessen and Peter Thiel in his lofty philosophizing and quest for immortality. Like Andreessen, Randall is a staunch accelerationist who believes that U.S. companies need to develop AI as fast as possible in order to both prevent the Chinese from controlling the technology, and to ostensibly ignite a new American utopia in which productivity, happiness, and health flourish. Randall’s power comes from the fact that he was Venis’ first investor, just as Thiel was an early investor in Facebook. While Andreessen pens manifestos about technological advancement, Randall paints his mission in grandiose, historical terms, using anti-democratic, sci-fi-inflected language that resembles that of the philosopher Curtis Yarvin, who has been funded and promoted by Thiel over his career. Randall’s justification of murder through utilitarian and Kantian lenses calls to mind Sam Bankman-Fried’s extensive philosophizing, which included a declaration that he would roll the dice on killing everyone on earth if there was a 51% chance he would create a second earth. Bankman-Fried’s approach—in embracing risk and harm in order to reap massive rewards—led him to be convicted of massive financial fraud. Randall is also obsessed with longevity just like Thiel, who has railed for years against the “inevitability of death” and yearns for “super-duper medical treatments” that would render him immortal. Jeff (Ramy Youssef) is the usurper.Ramy Youssef in Mountainhead Macall Polay—HBOJeff is a technologist who often serves as the movie’s conscience, slinging criticisms about the other characters. But he’s also deeply embedded within their world, and he needs their resources, particularly Venis’ access to computing power, to thrive. In the end, Jeff sells out his values for his own survival and well-being. AI skeptics have lobbed similar criticisms at the leaders of the main AI labs, including Altman—who started OpenAI as a nonprofit before attempting to restructure the company—as well as Demis Hassabis and Dario Amodei. Hassabis is the CEO of Google Deepmind and a winner of the 2024 Nobel Prize in Chemistry; a rare scientist surrounded by businessmen and technologists. In order to try to achieve his AI dreams of curing disease and halting global warning, Hassabis enlisted with Google, inking a contract in 2014 in which he prohibited Google from using his technology for military applications. But that clause has since disappeared, and the AI systems developed under Hassabis are being sold, via Google, to militaries like Israel’s. Another parallel can be drawn between Jeff and Amodei, an AI researcher who defected from OpenAI after becoming worried that the company was cutting back its safety measures, and then formed his own company, Anthropic. Amodei has urged governments to create AI guardrails and has warned about the potentially catastrophic effects of the AI industry’s race dynamics. But some have criticized Anthropic for operating similarly to OpenAI, prioritizing scale in a way that exacerbates competitive pressures. Souper (Jason Schwartzman) is the hanger-on. Jason Schwartzman in Mountainhead Macall Polay—HBOEvery quartet needs its Turtle or its Ringo; a clear fourth wheel to serve as a punching bag for the rest of the group’s alpha males. Mountainhead’s hanger-on is Souper, thus named because he has soup kitchen money compared to the rest (hundreds of millions as opposed to billions of dollars). In order to prove his worth, he’s fixated on getting funding for a meditation startup that he hopes will eventually become an “everything app.” No tech exec would want to be compared to Souper, who has a clear inferiority complex. But plenty of tech leaders have emphasized the importance of meditation and mindfulness—including Twitter co-founder and Square CEO Jack Dorsey, who often goes on meditation retreats. Armstrong, in his interview, declined to answer specific questions about his characters’ inspirations, but conceded that some of the speculations were in the right ballpark. “For people who know the area well, it's a little bit of a fun house mirror in that you see something and are convinced that it's them,” he says. “I think all of those people featured in my research. There's bits of Andreessen and David Sacks and some of those philosopher types. It’s a good parlor game to choose your Frankenstein limbs.”

From large language models (LLMs) to reasoning agents, today’s AI tools bring unprecedented computational demands. Trillion-parameter models, workloads running on-device, and swarms of agents collaborating to complete tasks all require a new paradigm of computing to become truly seamless and ubiquitous. Silicon’s mid-life crisis AI has evolved from classical ML to deep learning to generative AI. The most recent chapter, which took AI mainstream, hinges on two phases—training and inference—that are data and energy-intensive in terms of computation, data movement, and cooling. At the same time, Moore’s Law, which determines that the number of transistors on a chip doubles every two years, is reaching a physical and economic plateau. For the last 40 years, silicon chips and digital technology have nudged each other forward—every step ahead in processing capability frees the imagination of innovators to envision new products, which require yet more power to run. That is happening at light speed in the AI age. As models become more readily available, deployment at scale puts the spotlight on inference and the application of trained models for everyday use cases. This transition requires the appropriate hardware to handle inference tasks efficiently. Central processing units (CPUs) have managed general computing tasks for decades, but the broad adoption of ML introduced computational demands that stretched the capabilities of traditional CPUs. This has led to the adoption of graphics processing units (GPUs) and other accelerator chips for training complex neural networks, due to their parallel execution capabilities and high memory bandwidth that allow large-scale mathematical operations to be processed efficiently. But CPUs are already the most widely deployed and can be companions to processors like GPUs and tensor processing units (TPUs). AI developers are also hesitant to adapt software to fit specialized or bespoke hardware, and they favor the consistency and ubiquity of CPUs. Chip designers are unlocking performance gains through optimized software tooling, adding novel processing features and data types specifically to serve ML workloads, integrating specialized units and accelerators, and advancing silicon chip innovations, including custom silicon. AI itself is a helpful aid for chip design, creating a positive feedback loop in which AI helps optimize the chips that it needs to run. These enhancements and strong software support mean modern CPUs are a good choice to handle a range of inference tasks. Beyond silicon-based processors, disruptive technologies are emerging to address growing AI compute and data demands. The unicorn start-up Lightmatter, for instance, introduced photonic computing solutions that use light for data transmission to generate significant improvements in speed and energy efficiency. Quantum computing represents another promising area in AI hardware. While still years or even decades away, the integration of quantum computing with AI could further transform fields like drug discovery and genomics. The developments in ML theories and network architectures have significantly enhanced the efficiency and capabilities of AI models. Today, the industry is moving from monolithic models to agent-based systems characterized by smaller, specialized models that work together to complete tasks more efficiently at the edge—on devices like smartphones or modern vehicles. This allows them to extract increased performance gains, like faster model response times, from the same or even less compute. Researchers have developed techniques, including few-shot learning, to train AI models using smaller datasets and fewer training iterations. AI systems can learn new tasks from a limited number of examples to reduce dependency on large datasets and lower energy demands. Optimization techniques like quantization, which lower the memory requirements by selectively reducing precision, are helping reduce model sizes without sacrificing performance.  New system architectures, like retrieval-augmented generation (RAG), have streamlined data access during both training and inference to reduce computational costs and overhead. The DeepSeek R1, an open source LLM, is a compelling example of how more output can be extracted using the same hardware. By applying reinforcement learning techniques in novel ways, R1 has achieved advanced reasoning capabilities while using far fewer computational resources in some contexts. The integration of heterogeneous computing architectures, which combine various processing units like CPUs, GPUs, and specialized accelerators, has further optimized AI model performance. This approach allows for the efficient distribution of workloads across different hardware components to optimize computational throughput and energy efficiency based on the use case. As AI becomes an ambient capability humming in the background of many tasks and workflows, agents are taking charge and making decisions in real-world scenarios. These range from customer support to edge use cases, where multiple agents coordinate and handle localized tasks across devices. With AI increasingly used in daily life, the role of user experiences becomes critical for mass adoption. Features like predictive text in touch keyboards, and adaptive gearboxes in vehicles, offer glimpses of AI as a vital enabler to improve technology interactions for users. Edge processing is also accelerating the diffusion of AI into everyday applications, bringing computational capabilities closer to the source of data generation. Smart cameras, autonomous vehicles, and wearable technology now process information locally to reduce latency and improve efficiency. Advances in CPU design and energy-efficient chips have made it feasible to perform complex AI tasks on devices with limited power resources. This shift toward heterogeneous compute enhances the development of ambient intelligence, where interconnected devices create responsive environments that adapt to user needs. Seamless AI naturally requires common standards, frameworks, and platforms to bring the industry together. Contemporary AI brings new risks. For instance, by adding more complex software and personalized experiences to consumer devices, it expands the attack surface for hackers, requiring stronger security at both the software and silicon levels, including cryptographic safeguards and transforming the trust model of compute environments. More than 70% of respondents to a 2024 DarkTrace survey reported that AI-powered cyber threats significantly impact their organizations, while 60% say their organizations are not adequately prepared to defend against AI-powered attacks. Collaboration is essential to forging common frameworks. Universities contribute foundational research, companies apply findings to develop practical solutions, and governments establish policies for ethical and responsible deployment. Organizations like Anthropic are setting industry standards by introducing frameworks, such as the Model Context Protocol, to unify the way developers connect AI systems with data. Arm is another leader in driving standards-based and open source initiatives, including ecosystem development to accelerate and harmonize the chiplet market, where chips are stacked together through common frameworks and standards. Arm also helps optimize open source AI frameworks and models for inference on the Arm compute platform, without needing customized tuning.  How far AI goes to becoming a general-purpose technology, like electricity or semiconductors, is being shaped by technical decisions taken today. Hardware-agnostic platforms, standards-based approaches, and continued incremental improvements to critical workhorses like CPUs, all help deliver the promise of AI as a seamless and silent capability for individuals and businesses alike. Open source contributions are also helpful in allowing a broader range of stakeholders to participate in AI advances. By sharing tools and knowledge, the community can cultivate innovation and help ensure that the benefits of AI are accessible to everyone, everywhere. Learn more about Arm’s approach to enabling AI everywhere. This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. This content was researched, designed, and written entirely by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Alibaba Group has introduced QwenLong-L1, a new framework that enables large language models (LLMs) to reason over extremely long inputs. This development could unlock a new wave of enterprise applications that require models to understand and draw insights from extensive documents such as detailed corporate filings, lengthy financial statements, or complex legal contracts. The challenge of long-form reasoning for AI Recent advances in large reasoning models (LRMs), particularly through reinforcement learning (RL), have significantly improved their problem-solving capabilities. Research shows that when trained with RL fine-tuning, LRMs acquire skills similar to human “slow thinking,” where they develop sophisticated strategies to tackle complex tasks. However, these improvements are primarily seen when models work with relatively short pieces of text, typically around 4,000 tokens. The ability of these models to scale their reasoning to much longer contexts (e.g., 120,000 tokens) remains a major challenge. Such long-form reasoning requires a robust understanding of the entire context and the ability to perform multi-step analysis. “This limitation poses a significant barrier to practical applications requiring interaction with external knowledge, such as deep research, where LRMs must collect and process information from knowledge-intensive environments,” the developers of QwenLong-L1 write in their paper. The researchers formalize these challenges into the concept of “long-context reasoning RL.” Unlike short-context reasoning, which often relies on knowledge already stored within the model, long-context reasoning RL requires models to retrieve and ground relevant information from lengthy inputs accurately. Only then can they generate chains of reasoning based on this incorporated information.  Training models for this through RL is tricky and often results in inefficient learning and unstable optimization processes. Models struggle to converge on good solutions or lose their ability to explore diverse reasoning paths. QwenLong-L1: A multi-stage approach QwenLong-L1 is a reinforcement learning framework designed to help LRMs transition from proficiency with short texts to robust generalization across long contexts. The framework enhances existing short-context LRMs through a carefully structured, multi-stage process: Warm-up Supervised Fine-Tuning (SFT): The model first undergoes an SFT phase, where it is trained on examples of long-context reasoning. This stage establishes a solid foundation, enabling the model to ground information accurately from long inputs. It helps develop fundamental capabilities in understanding context, generating logical reasoning chains, and extracting answers. Curriculum-Guided Phased RL: At this stage, the model is trained through multiple phases, with the target length of the input documents gradually increasing. This systematic, step-by-step approach helps the model stably adapt its reasoning strategies from shorter to progressively longer contexts. It avoids the instability often seen when models are abruptly trained on very long texts. Difficulty-Aware Retrospective Sampling: The final training stage incorporates challenging examples from the preceding training phases, ensuring the model continues to learn from the hardest problems. This prioritizes difficult instances and encourages the model to explore more diverse and complex reasoning paths. QwenLong-L1 process Source: arXiv Beyond this structured training, QwenLong-L1 also uses a distinct reward system. While training for short-context reasoning tasks often relies on strict rule-based rewards (e.g., a correct answer in a math problem), QwenLong-L1 employs a hybrid reward mechanism. This combines rule-based verification, which ensures precision by checking for strict adherence to correctness criteria, with an “LLM-as-a-judge.” This judge model compares the semanticity of the generated answer with the ground truth, allowing for more flexibility and better handling of the diverse ways correct answers can be expressed when dealing with long, nuanced documents. Putting QwenLong-L1 to the test The Alibaba team evaluated QwenLong-L1 using document question-answering (DocQA) as the primary task. This scenario is highly relevant to enterprise needs, where AI must understand dense documents to answer complex questions.  Experimental results across seven long-context DocQA benchmarks showed QwenLong-L1’s capabilities. Notably, the QWENLONG-L1-32B model (based on DeepSeek-R1-Distill-Qwen-32B) achieved performance comparable to Anthropic’s Claude-3.7 Sonnet Thinking, and outperformed models like OpenAI’s o3-mini and Qwen3-235B-A22B. The smaller QWENLONG-L1-14B model also outperformed Google’s Gemini 2.0 Flash Thinking and Qwen3-32B.  Source: arXiv An important finding relevant to real-world applications is how RL training results in the model developing specialized long-context reasoning behaviors. The paper notes that models trained with QwenLong-L1 become better at “grounding” (linking answers to specific parts of a document), “subgoal setting” (breaking down complex questions), “backtracking” (recognizing and correcting their own mistakes mid-reasoning), and “verification” (double-checking their answers). For instance, while a base model might get sidetracked by irrelevant details in a financial document or get stuck in a loop of over-analyzing unrelated information, the QwenLong-L1 trained model demonstrated an ability to engage in effective self-reflection. It could successfully filter out these distractor details, backtrack from incorrect paths, and arrive at the correct answer. Techniques like QwenLong-L1 could significantly expand the utility of AI in the enterprise. Potential applications include legal tech (analyzing thousands of pages of legal documents), finance (deep research on annual reports and financial filings for risk assessment or investment opportunities) and customer service (analyzing long customer interaction histories to provide more informed support). The researchers have released the code for the QwenLong-L1 recipe and the weights for the trained models. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

From large language models (LLMs) to reasoning agents, today’s AI tools bring unprecedented computational demands. Trillion-parameter models, workloads running on-device, and swarms of agents collaborating to complete tasks all require a new paradigm of computing to become truly seamless and ubiquitous. First, technical progress in hardware and silicon design is critical to pushing the boundaries of compute. Second, advances in machine learning (ML) allow AI systems to achieve increased efficiency with smaller computational demands. Finally, the integration, orchestration, and adoption of AI into applications, devices, and systems is crucial to delivering tangible impact and value. Silicon’s mid-life crisis AI has evolved from classical ML to deep learning to generative AI. The most recent chapter, which took AI mainstream, hinges on two phases—training and inference—that are data and energy-intensive in terms of computation, data movement, and cooling. At the same time, Moore’s Law, which determines that the number of transistors on a chip doubles every two years, is reaching a physical and economic plateau. For the last 40 years, silicon chips and digital technology have nudged each other forward—every step ahead in processing capability frees the imagination of innovators to envision new products, which require yet more power to run. That is happening at light speed in the AI age. As models become more readily available, deployment at scale puts the spotlight on inference and the application of trained models for everyday use cases. This transition requires the appropriate hardware to handle inference tasks efficiently. Central processing units (CPUs) have managed general computing tasks for decades, but the broad adoption of ML introduced computational demands that stretched the capabilities of traditional CPUs. This has led to the adoption of graphics processing units (GPUs) and other accelerator chips for training complex neural networks, due to their parallel execution capabilities and high memory bandwidth that allow large-scale mathematical operations to be processed efficiently. But CPUs are already the most widely deployed and can be companions to processors like GPUs and tensor processing units (TPUs). AI developers are also hesitant to adapt software to fit specialized or bespoke hardware, and they favor the consistency and ubiquity of CPUs. Chip designers are unlocking performance gains through optimized software tooling, adding novel processing features and data types specifically to serve ML workloads, integrating specialized units and accelerators, and advancing silicon chip innovations, including custom silicon. AI itself is a helpful aid for chip design, creating a positive feedback loop in which AI helps optimize the chips that it needs to run. These enhancements and strong software support mean modern CPUs are a good choice to handle a range of inference tasks. Beyond silicon-based processors, disruptive technologies are emerging to address growing AI compute and data demands. The unicorn start-up Lightmatter, for instance, introduced photonic computing solutions that use light for data transmission to generate significant improvements in speed and energy efficiency. Quantum computing represents another promising area in AI hardware. While still years or even decades away, the integration of quantum computing with AI could further transform fields like drug discovery and genomics. Understanding models and paradigms The developments in ML theories and network architectures have significantly enhanced the efficiency and capabilities of AI models. Today, the industry is moving from monolithic models to agent-based systems characterized by smaller, specialized models that work together to complete tasks more efficiently at the edge—on devices like smartphones or modern vehicles. This allows them to extract increased performance gains, like faster model response times, from the same or even less compute. Researchers have developed techniques, including few-shot learning, to train AI models using smaller datasets and fewer training iterations. AI systems can learn new tasks from a limited number of examples to reduce dependency on large datasets and lower energy demands. Optimization techniques like quantization, which lower the memory requirements by selectively reducing precision, are helping reduce model sizes without sacrificing performance.  New system architectures, like retrieval-augmented generation (RAG), have streamlined data access during both training and inference to reduce computational costs and overhead. The DeepSeek R1, an open source LLM, is a compelling example of how more output can be extracted using the same hardware. By applying reinforcement learning techniques in novel ways, R1 has achieved advanced reasoning capabilities while using far fewer computational resources in some contexts. The integration of heterogeneous computing architectures, which combine various processing units like CPUs, GPUs, and specialized accelerators, has further optimized AI model performance. This approach allows for the efficient distribution of workloads across different hardware components to optimize computational throughput and energy efficiency based on the use case. Orchestrating AI As AI becomes an ambient capability humming in the background of many tasks and workflows, agents are taking charge and making decisions in real-world scenarios. These range from customer support to edge use cases, where multiple agents coordinate and handle localized tasks across devices. With AI increasingly used in daily life, the role of user experiences becomes critical for mass adoption. Features like predictive text in touch keyboards, and adaptive gearboxes in vehicles, offer glimpses of AI as a vital enabler to improve technology interactions for users. Edge processing is also accelerating the diffusion of AI into everyday applications, bringing computational capabilities closer to the source of data generation. Smart cameras, autonomous vehicles, and wearable technology now process information locally to reduce latency and improve efficiency. Advances in CPU design and energy-efficient chips have made it feasible to perform complex AI tasks on devices with limited power resources. This shift toward heterogeneous compute enhances the development of ambient intelligence, where interconnected devices create responsive environments that adapt to user needs. Seamless AI naturally requires common standards, frameworks, and platforms to bring the industry together. Contemporary AI brings new risks. For instance, by adding more complex software and personalized experiences to consumer devices, it expands the attack surface for hackers, requiring stronger security at both the software and silicon levels, including cryptographic safeguards and transforming the trust model of compute environments. More than 70% of respondents to a 2024 DarkTrace survey reported that AI-powered cyber threats significantly impact their organizations, while 60% say their organizations are not adequately prepared to defend against AI-powered attacks. Collaboration is essential to forging common frameworks. Universities contribute foundational research, companies apply findings to develop practical solutions, and governments establish policies for ethical and responsible deployment. Organizations like Anthropic are setting industry standards by introducing frameworks, such as the Model Context Protocol, to unify the way developers connect AI systems with data. Arm is another leader in driving standards-based and open source initiatives, including ecosystem development to accelerate and harmonize the chiplet market, where chips are stacked together through common frameworks and standards. Arm also helps optimize open source AI frameworks and models for inference on the Arm compute platform, without needing customized tuning.  How far AI goes to becoming a general-purpose technology, like electricity or semiconductors, is being shaped by technical decisions taken today. Hardware-agnostic platforms, standards-based approaches, and continued incremental improvements to critical workhorses like CPUs, all help deliver the promise of AI as a seamless and silent capability for individuals and businesses alike. Open source contributions are also helpful in allowing a broader range of stakeholders to participate in AI advances. By sharing tools and knowledge, the community can cultivate innovation and help ensure that the benefits of AI are accessible to everyone, everywhere. Learn more about Arm’s approach to enabling AI everywhere. This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. This content was researched, designed, and written entirely by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Perplexity this week released Perplexity Labs, a new tool for Pro users that can craft reports, spreadsheets, dashboards, and visual representations, to meet users’ increased demand for AI productivity tools with greater autonomy and ever more sophisticated capabilities. The platform, a rival to Anthropic Claude, OpenAI’s ChatGPT, and Google Gemini, can even work on its own for 10 minutes (or more) as it reasons through complicated assignments. “Labs underscores a broader shift toward multi-agent AI systems that plan, execute, and refine full workflows,” said Thomas Randall, research lead for AI at Info-Tech Research Group. Designed to handle more complex assignments Perplexity launched Perplexity Search, its proprietary search engine, in December 2022, just after ChatGPT dropped, and earlier this year released Deep Research (now to be rebranded as Research), which scours the web, reads papers, reasons through materials, and creates comprehensive reports for users. The company says that Perplexity Labs is like “having a team” that can bring projects from ideation to reality. The platform creates reports, spreadsheets, dashboards and simple web apps. It can perform at least 10 minutes of self-supervised work, uses web browsing, writes and executes code to handle tasks like organizing data or applying formulas, and can create charts and images. “In some respects, this is a continuation of Perplexity’s original capabilities as an AI-driven search engine that provides deeper answers,” said Hyoun Park, CEO and chief analyst at Amalgam Insights. Indeed, Perplexity explained that Labs was designed to handle more complex assignments than Deep Research. “While Deep Research remains the fastest way to obtain comprehensive answers to in-depth questions —  typically within 3 or 4 minutes — Labs is designed to invest more time (10  minutes or longer) and leverage additional tools, such as advanced file generation and mini-app creation,” Perplexity wrote in a blog post. “This expanded capability empowers you to develop a broader array of deliverables for your projects.” With its longer research workflow, Perplexity Labs can generate spreadsheets, visual representations, and high-quality reports, the company said. It iteratively searches through hundreds of sources, reasons about that data, and refines its approach as it gets deeper into a project, similar to the way in which a human researcher might approach a new area of study. To create interactive dashboards without the need for coding expertise or external development tools like Ploty and Dash, users just describe what they’re visualizing in natural language, and Labs will generate it in real-time. Dashboards could, for instance, visualize business finances or other complex datasets, incorporating clickable elements to allow non-technical users to quickly act on insights. In one example from the blog, Perplexity prompted Labs from the position of a leader at a tech consulting firm looking to create a potential customer list. It specified that it wanted to partner with US B2B companies in seed, series A, or series B stages, and asked Labs to list 20 relevant companies and include key details including contact information. Labs compiled a comprehensive dataset of potential customers, organizing them by stage (A, B, or seed) and identified their core focus, intended customers, and funding to date. The platform cited links from Forbes, YCombinator, and Exploding Topics that it had used as sources. When further prompted, it crafted introductory emails to the CEOs of the series A startups. To simplify workflows, Labs arranges generated files in a dedicated tab for easy access, supports integration with other tools such as Google Sheets, and allows users to pull out and format citations to bring credibility to its research. Finished materials can be exported as PDFs or documents, or converted into a shareable Perplexity Page. Pro subscribers ($20 a month) can now work with Labs on Web, iOS, and Android; Mac and Windows apps are coming soon. A good fit for enterprise users? This new capability joins an increasingly competitive space, as users look for AI productivity tools that are ever-more performant and can handle more and more tasks autonomously. Park pointed out that Perplexity Labs is a response to tools and models such as OpenAI o1-pro (launched in March), Claude Opus 4 (released in May), and Google’s recent Flow and Firebase announcements. “There is a massive Hunger Games in the AI world right now,” said Park. “Every major vendor is ferociously trying to one-up each other in providing more functionality, either in a native model or with an agency of AI agents designed to work together and create digital assets such as documents, apps, and videos.” However, Perplexity Labs does provide differentiation from other providers in the market, Info-Tech’s Randall noted. In particular, Perplexity is betting that users will prefer a “low-cost, open, tool-agnostic sandbox” for web crawling, code execution, and the creation of finished artifacts including mini web apps. “These capabilities cannot yet be found in other enterprise platforms, such as Microsoft or Google offerings,” said Randall. But enterprises should approach Perplexity Labs with a governance-first mindset, he emphasized. Assets live in Perplexity’s cloud and, for now, lack the private data grounding and compliance controls that CIOs expect, and that they find in tools such as Microsoft Copilot or Google Gemini. From an enterprise perspective, Park noted, the biggest challenge is that every asset-creating model and agent is “still opaque” when it comes to understanding the assumptions, training, and reliability of assumptions used to create a document or app. He compared it to the way the iPhone bypassed BlackBerry and Windows through “sheer consumer delight.” “At some point, AI vendors seeking serious business usage will need to provide more transparency and governance tools to the business world, just as mobile device management and mobile security solutions eventually came to the iPhone,” said Park. Otherwise, businesses may be compelled to build their own clunky but secure versions of Perplexity Labs, “which are guaranteed to be less accurate and useful just [based on] the history of business apps trying to imitate viral consumer apps,” he said.

Google’s AI assistant, Gemini, is gaining a more prominent place in your inbox with the launch of email summary cards, which will appear at the top of your emails. The company announced Thursday that users would no longer have to tap an option to summarize an email with AI. Instead, the AI will now automatically summarize the content when needed, without requiring user interaction. When Gemini launched in the side panel of Gmail last year, one of the features allowed users to summarize their long email threads, along with other tools like those to draft email messages or see suggested responses, among other things. Now, Google is putting the AI to work on your inbox, whether or not it’s something you want to use. The update is another example of how AI is quickly infiltrating the software and services people use the most, even though AI summaries aren’t always reliable. When Apple rolled out AI summaries for app push notifications, for example, the BBC found the feature made repeated mistakes when summarizing news headlines. Apple ended up pausing the AI summaries for news apps. Google’s own AI Overviews feature for Search has also repeatedly made mistakes, offering poor quality and inaccurate information at times. With the new email summary cards, Gemini will list a longer email’s key points and will then continue to update that synopsis as replies arrive. The feature won’t replace the option to manually click a button to summarize an email, Google notes. That will still appear as a chip at the top of the email and in Gmail’s Gemini side panel. Techcrunch event Save now through June 4 for TechCrunch Sessions: AI Save $300 on your ticket to TC Sessions: AI—and get 50% off a second. Hear from leaders at OpenAI, Anthropic, Khosla Ventures, and more during a full day of expert insights, hands-on workshops, and high-impact networking. These low-rate deals disappear when the doors open on June 5. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW The feature is initially available only for emails in English. Depending on your region, the summary cards may be turned on or off by default. (For instance, smart features are turned off in the EU, the UK, Switzerland, and Japan, Google’s help documentation notes.) Others can choose to enable or disable the feature from Gmail’s Settings under “Smart features.” Workplace admins can also opt to disable the personalization settings for users from the Admin console.

Reading Time: 5 minutes Last week, Google unveiled many interesting updates around Gemini at Google I/O 2025. The key highlight was the launch of Gemini 2.5 Pro, their most intelligent model yet. Although there were many updates, we’ve focused on those that will be beneficial for marketers like you. So, relax and read on to discover all the insights. Future of Google Search Shopping Google is reimagining how shoppers shop online—from product discovery to trial to checkout. If you sell online and list your products on Google, this is a huge update. The traditional method of shopping involved scanning hundreds of websites, reading reviews, and then buying a product. But with Google AI Mode or the Gemini app, shoppers can ask complicated questions, and Google will do the heavy lifting. Source – Google  For example, rather than searching for “best rugs for sofa”, shoppers can use Google AI mode and ask something like this, “I have a light gray sofa, I need a rug to brighten my room”. They can also add insights like “I have four kids at my house and they like to bring friends over”. But that’s not it, there’s more. Try it on Now, shoppers can try on clothes virtually to get a feel for how a style might look on them. Source – Google  They just need to upload your full-size image. While you might have seen something similar on other platforms, Google takes it to the next level with – Advanced 3D Shape Understanding Segmentation Diffusion There’s also a completely new way for checkout. Agentic Checkout Shoppers are always on the hunt for the best deals. Google’s new Agentic Checkout feature simplifies online shopping. It allows shoppers to specify their desired price, color, and size for a product, and Google will alert them when the item meets or drops below that price. When the price is dropped, shoppers can instantly check out with a click of a button, and all the details of your product (including size and colour) will already be added. Source – Google  Google examines numerous websites, sending alerts when its findings align with your specified preferences. If a price reduction occurs, a single click enables immediate checkout, as all your product specifics, such as size and color, are already saved. What does it mean for marketers? If you sell products online and list them on Google, you have a completely new channel through which your customers can discover, try, and purchase your products. However, optimize your products according to Google’s guidelines to ensure that they appear for relevant queries. Apparel companies must take advantage of the Try it on feature to provide their customers with a rich, personalized experience. Overall, E-commerce marketers can drive a lot of awareness, engagement, and conversions using the above updates. Gemini 2.5 Pro, AI Overviews and the New “AI Mode” Gemini 2.5 Pro is Google’s most intelligent, secure, and efficient LLM yet. Marketers can continue to use it for complicated tasks and deep research. It powers Google AI Overviews and the latest AI mode inside the Gemini App. Google AI Overviews is used by over 1.5 billion users every month across 200 countries. Rather than just providing links, AI Overviews provide direct, conversational answers and summaries. This change means that users get more information without clicking through, changing how visibility and traffic work. Seeing the rapid usage of AI Overviews, Google has launched something called “AI Mode.” They’ve reimagined Search from the ground up. Source – Google   Unlike a normal Google search, users can ask longer and more complex questions. The queries can be 2-3 times longer, and they can even ask follow-up questions. And the best part: It’ll take into account their personal information for even more personalized answers. AI Mode has already been launched in the US. What does it mean for marketers? To summarise, marketers now need to shift gears from just getting people to their website/app to making sure their content shows up in these AI-generated snippet/answers. Google has updated its generative AI capabilities, including image, audio, and video generation. Veo 3 now generates audio alongside video. Below is an example. What does it mean for marketers? Google is really pushing boundaries when it comes to what you can generate using AI. Marketers should definitely explore the various generative AI capabilities for improved campaign creation. However, if you’re looking to craft customer campaigns that resonate with your target customers and get higher engagement, you need to check out our own Generative AI platform, Merlin AI. There’s nothing better than an AI that constantly analyzes your previous campaigns to give you personalized output. Glance leveraged it for 50% faster campaign go-live times. Read the complete story here. Gemini Now Supports MCP The Gemini SDK now supports the Model Context Protocol (MCP), which was introduced by Anthropic. In addition, Google is bringing agentic capabilities to Search, Chrome, and the Gemini App. They call this “Agent Mode.” MCP is an emerging technical standard that allows AI models (like those used by Google) to connect more easily and securely with external, real-world data sources and tools. What does it mean for marketers? If you’re serious about building AI agents and you’ve been using Gemini, this is big news. The support for MCP opens up a new world of possibilities. But if you’re just getting started, you can read our latest on AI Agents for Customer Engagement. AI Agents can assist marketers with everything from campaign creation to analytics to decision-making. To help drive awareness, we’re also hosting a live masterclass titled “Agentic AI and Customer Engagement in 2025 and Beyond”. In this live panel discussion, we’ll cut through the noise and explore the exciting, practical ways to leverage AI in customer engagement. You can save your seat here. Android XR The most interesting launch was Google’s push into the XR space, the Gemini-powered Android XR platform. Designed for smart glasses and VR headsets, Android XR leverages Gemini for real-time translation, live messaging, contextual navigation, and more. Source – Google  What does it mean for marketers? Android XR is an emerging channel for marketers to consider. Its potential for rapid adoption suggests it could become a significant additional marketing avenue in the near future. Keep an eye out for future updates. Wrapping Up For marketers, the above updates mean that embracing AI-powered tools and strategies is no longer optional but essential for success. By leveraging these innovations, marketers can drive better results, improve customer engagement, and stay ahead of the curve in the ever-evolving world. The post Google I/O 2025 – What Marketers Need to Know appeared first on MoEngage.