People reported that photos weren’t loading, its banking service XMoney wasn’t working, and search wasn’t returning any results, among other things.
Read more Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation.
Read more The chief executive of the United States Institute of Peace (USIP) says Elon Musk’s DOGE left the nonprofit’s Washington, D.C., headquarters in disarray.
Read more Venture capitalist Mary Meeker just dropped a 340-page report to describe the speed at which AI is being developed, adopted, spent on, and used.
Read 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.
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.
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
This benchmark used Reddit’s AITA to test how much AI models suck up to us
Back in April, OpenAI announced it was rolling back an update to its GPT-4o model that made ChatGPT’s responses to user queries too sycophantic.
An AI model that acts in an overly agreeable and flattering way is more than just annoying. It could reinforce users’ incorrect beliefs, mislead people, and spread misinformation that can be dangerous—a particular risk when increasing numbers of young people are using ChatGPT as a life advisor. And because sycophancy is difficult to detect, it can go unnoticed until a model or update has already been deployed.
A new benchmark called Elephant that measures the sycophantic tendencies of major AI models could help companies avoid these issues in the future. But just knowing when models are sycophantic isn’t enough; you need to be able to do something about it. And that’s trickier. Read the full story.
—Rhiannon Williams
The AI Hype Index
Separating AI reality from hyped-up fiction isn’t always easy. That’s why we’ve created the AI Hype Index—a simple, at-a-glance summary of everything you need to know about the state of the industry. Take a look at this month’s edition of the index here.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 Anduril is partnering with Meta to build an advanced weapons system
EagleEye’s VR headsets will enhance soldiers’ hearing and vision. (WSJ $)
+ Palmer Luckey wants to turn “warfighters into technomancers.” (TechCrunch)
+ Luckey and Mark Zuckerberg have buried the hatchet, then. (Insider $)
+ Palmer Luckey on the Pentagon’s future of mixed reality. (MIT Technology Review)
2 A new Texas law requires app stores to verify users’ ages
It’s following in Utah’s footsteps, which passed a similar bill in March. (NYT $)
+ Apple has pushed back on the law. (CNN)
3 What happens to DOGE now?
It has lost its leader and a top lieutenant within the space of a week. (WSJ $)
+ Musk’s departure raises questions over how much power it will wield without him. (The Guardian)
+ DOGE’s tech takeover threatens the safety and stability of our critical data. (MIT Technology Review)
4 NASA’s ambitions of a 2027 moon landing are looking less likely
It needs SpaceX’s Starship, which keeps blowing up. (WP $)
+ Is there a viable alternative? (New Scientist $)
5 Students are using AI to generate nude images of each other
It’s a grave and growing problem that no one has a solution for. (404 Media)
6 Google AI Overviews doesn’t know what year it is
A year after its introduction, the feature is still making obvious mistakes. (Wired $)
+ Google’s new AI-powered search isn’t fit to handle even basic queries. (NYT $)
+ The company is pushing AI into everything. Will it pay off? (Vox)
+ Why Google’s AI Overviews gets things wrong. (MIT Technology Review)
7 Hugging Face has created two humanoid robots 
The machines are open source, meaning anyone can build software for them. (TechCrunch)
8 A popular vibe coding app has a major security flaw
Despite being notified about it months ago. (Semafor)
+ Any AI coding program catering to amateurs faces the same issue. (The Information $)
+ What is vibe coding, exactly? (MIT Technology Review)
9 AI-generated videos are becoming way more realistic
But not when it comes to depicting gymnastics. (Ars Technica)
10 This electronic tattoo measures your stress levels
Consider it a mood ring for your face. (IEEE Spectrum)
Quote of the day
“I think finally we are seeing Apple being dragged into the child safety arena kicking and screaming.”
—Sarah Gardner, CEO of child safety collective Heat Initiative, tells the Washington Post why Texas’ new app store law could signal a turning point for Apple.
One more thing
House-flipping algorithms are coming to your neighborhood
When Michael Maxson found his dream home in Nevada, it was not owned by a person but by a tech company, Zillow. When he went to take a look at the property, however, he discovered it damaged by a huge water leak. Despite offering to handle the costly repairs himself, Maxson discovered that the house had already been sold to another family, at the same price he had offered.
During this time, Zillow lost more than $420 million in three months of erratic house buying and unprofitable sales, leading analysts to question whether the entire tech-driven model is really viable. For the rest of us, a bigger question remains: Does the arrival of Silicon Valley tech point to a better future for housing or an industry disruption to fear? Read the full story.
—Matthew Ponsford
We can still have nice things
A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line or skeet ’em at me.)
+ A 100-mile real-time ultramarathon video game that lasts anywhere up to 27 hours is about as fun as it sounds.
+ Here’s how edible glitter could help save the humble water vole from extinction.
+ Cleaning massive statues is not for the faint-hearted ($)
+ When is a flute teacher not a flautist? When he’s a whistleblower.
Back in April, OpenAIannounced it was rolling back an update to its GPT-4o model that made ChatGPT’s responses to user queries too sycophantic.
An AI model that acts in an overly agreeable and flattering way is more than just annoying. It could reinforce users’ incorrect beliefs, mislead people, and spread misinformation that can be dangerous—a particular risk when increasing numbers of young people are using ChatGPT as a life advisor. And because sycophancy is difficult to detect, it can go unnoticed until a model or update has already been deployed, as OpenAI found out.
A new benchmark that measures the sycophantic tendencies of major AI models could help AI companies avoid these issues in the future. The team behind Elephant, from Stanford, Carnegie Mellon, and the University of Oxford, found that LLMs consistently exhibit higher rates of sycophancy than humans do.
“We found that language models don’t challenge users’ assumptions, even when they might be harmful or totally misleading,” says Myra Cheng, a PhD student at Stanford University who worked on the research, which has not been peer-reviewed. “So we wanted to give researchers and developers the tools to empirically evaluate their models on sycophancy, because it’s a problem that is so prevalent.”
It’s hard to assess how sycophantic AI models are because sycophancy comes in many forms. Previous research has tended to focus on how chatbots agree with users even when what the human has told the AI is demonstrably wrong—for example, they might state that Nice, not Paris, is the capital of France.
While this approach is still useful, it overlooks all the subtler, more insidious ways in which models behave sycophantically when there isn’t a clear ground truth to measure against. Users typically ask LLMs open-ended questions containing implicit assumptions, and those assumptions can trigger sycophantic responses, the researchers claim. For example, a model that’s asked “How do I approach my difficult coworker?” is more likely to accept the premise that a coworker is difficult than it is to question why the user thinks so.
To bridge this gap, Elephant is designed to measure social sycophancy—a model’s propensity to preserve the user’s “face,” or self-image, even when doing so is misguided or potentially harmful. It uses metrics drawn from social science to assess five nuanced kinds of behavior that fall under the umbrella of sycophancy: emotional validation, moral endorsement, indirect language, indirect action, and accepting framing.
To do this, the researchers tested it on two data sets made up of personal advice written by humans. This first consisted of 3,027 open-ended questions about diverse real-world situations taken from previous studies. The second data set was drawn from 4,000 posts on Reddit’s AITA (“Am I the Asshole?”) subreddit, a popular forum among users seeking advice. Those data sets were fed into eight LLMs from OpenAI (the version of GPT-4o they assessed was earlier than the version that the company later called too sycophantic), Google, Anthropic, Meta, and Mistral, and the responses were analyzed to see how the LLMs’ answers compared with humans’.
Overall, all eight models were found to be far more sycophantic than humans, offering emotional validation in 76% of cases (versus 22% for humans) and accepting the way a user had framed the query in 90% of responses (versus 60% among humans). The models also endorsed user behavior that humans said was inappropriate in an average of 42% of cases from the AITA data set.
But just knowing when models are sycophantic isn’t enough; you need to be able to do something about it. And that’s trickier. The authors had limited success when they tried to mitigate these sycophantic tendencies through two different approaches: prompting the models to provide honest and accurate responses, and training a fine-tuned model on labeled AITA examples to encourage outputs that are less sycophantic. For example, they found that adding “Please provide direct advice, even if critical, since it is more helpful to me” to the prompt was the most effective technique, but it only increased accuracy by 3%. And although prompting improved performance for most of the models, none of the fine-tuned models were consistently better than the original versions.
“It’s nice that it works, but I don’t think it’s going to be an end-all, be-all solution,” says Ryan Liu, a PhD student at Princeton University who studies LLMs but was not involved in the research. “There’s definitely more to do in this space in order to make it better.”
Gaining a better understanding of AI models’ tendency to flatter their users is extremely important because it gives their makers crucial insight into how to make them safer, says Henry Papadatos, managing director at the nonprofit SaferAI. The breakneck speed at which AI models are currently being deployed to millions of people across the world, their powers of persuasion, and their improved abilities to retain information about their users add up to “all the components of a disaster,” he says. “Good safety takes time, and I don’t think they’re spending enough time doing this.”
While we don’t know the inner workings of LLMs that aren’t open-source, sycophancy is likely to be baked into models because of the ways we currently train and develop them. Cheng believes that models are often trained to optimize for the kinds of responses users indicate that they prefer. ChatGPT, for example, gives users the chance to mark a response as good or bad via thumbs-up and thumbs-down icons. “Sycophancy is what gets people coming back to these models. It’s almost the core of what makes ChatGPT feel so good to talk to,” she says. “And so it’s really beneficial, for companies, for their models to be sycophantic.” But while some sycophantic behaviors align with user expectations, others have the potential to cause harm if they go too far—particularly when people do turn to LLMs for emotional support or validation.
“We want ChatGPT to be genuinely useful, not sycophantic,” an OpenAI spokesperson says. “When we saw sycophantic behavior emerge in a recent model update, we quickly rolled it back and shared an explanation of what happened. We’re now improving how we train and evaluate models to better reflect long-term usefulness and trust, especially in emotionally complex conversations.”
Cheng and her fellow authors suggest that developers should warn users about the risks of social sycophancy and consider restricting model usage in socially sensitive contexts. They hope their work can be used as a starting point to develop safer guardrails.
She is currently researching the potential harms associated with these kinds of LLM behaviors, the way they affect humans and their attitudes toward other people, and the importance of making models that strike the right balance between being too sycophantic and too critical. “This is a very big socio-technical challenge,” she says. “We don’t want LLMs to end up telling users, ‘You are the asshole.’”
Buyers of BlackRock and Fidelity’s spot Ether ETFs are at an average unrealized loss of around 21%, says Glassnode.
OFAC claims Funnull Technology purchased a code repository used by web developers and altered it to redirect legitimate websites to scam websites.
The party leader leaned into pro-crypto talking points at the Bitcoin 2025 conference, announcing plans to introduce a digital asset bill if Reform UK wins in the next election.