Industry Briefing
A single destination for timely, editor-curated robotics news from around the world.
SpaceX's Starmind Faces Feasibility Challenges for 1 Million Satellite Deployment
On January 30, 2026, SpaceX submitted a request to the FCC to launch up to 1 million satellites as part of its Starmind orbital compute constellation. This ambitious plan is unprecedented, as the total number of satellites ever launched globally is in the low tens of thousands. The proposal seeks a waiver from standard deployment milestones, citing reliance on the Starship's full reusability for success. The significance of this request lies in the technical and logistical challenges it presents. Experts warn that low Earth orbit may not support the proposed number of active satellites without risking a debris cascade. SpaceX's own IPO prospectus acknowledges unresolved dependencies related to Starship's launch cadence and reusability, which are critical for the orbital AI compute strategy. Looking ahead, the timeline for achieving the necessary launch cadence and manufacturing capacity remains uncertain. SpaceX's Gigasat facility in Texas aims for volume production by late 2027, but this would require unprecedented output levels. No further timeline was disclosed at the time of publication, leaving the feasibility of the Starmind project in question.
MIT and York University Study Visual Learning in the Brain Using Neural Networks
Researchers at MIT’s McGovern Institute for Brain Research and York University in Toronto have investigated how visual learning occurs in the brain. By analyzing neural activity and utilizing computational modeling, they compared the learning processes of animals and an artificial neural network designed to mimic brain architecture. Their findings, published on July 8 in Nature Communications, reveal that changes in visual processing are crucial for learning to discriminate new objects. This research is significant as it enhances our understanding of the brain's adaptability and the mechanisms behind visual learning. The study suggests that while the overall activity patterns in the inferior temporal cortex remain stable, subtle changes occur in response to learned object recognition. These insights could inform educational strategies and improve learning outcomes across various contexts. Looking ahead, the researchers aim to further explore how these modest changes in neural activity contribute to learning. They believe that artificial neural networks can provide valuable insights into biological learning processes, potentially leading to new experimental approaches and predictions that extend beyond current understanding. No further timeline was disclosed at the time of publication.
Research Neuroscience Learning Brain and cognitive sciences Computer modeling Vision
SpaceX Proposes 1 Million AI Satellites to Address Ground Data Center Constraints
On January 30, 2026, SpaceX filed with the FCC to launch up to 1 million AI compute satellites, positioning orbital data centers as a solution to the increasing demand for AI computing power. Ground data centers are facing significant challenges, with energy consumption projected to reach approximately 1,050 TWh in 2026, making them the fifth-largest electricity consumer globally. The demand for new data center capacity is outpacing the growth of power generation infrastructure, leading to a critical bottleneck in the grid system. The significance of this initiative lies in the structural constraints faced by ground data centers, including power delivery limitations, high water consumption, and local opposition to new projects. The Uptime Institute's 2026 outlook identifies power as the primary constraint on data center growth, with capacity clearing prices in the PJM grid skyrocketing to $329.17/MW, driven by data center expansion. Additionally, cooling requirements are becoming increasingly unsustainable, with facilities consuming vast amounts of water, further complicating their operational viability. Looking ahead, SpaceX's orbital AI compute initiative aims to circumvent these challenges by leveraging the advantages of space, such as continuous solar power and minimal local opposition. The first AI prototypes are expected to launch in early 2027, with operational deployments planned for 2028. No further timeline was disclosed at the time of publication.
SpaceX Launches Starmind Project for 1 Million AI Satellites by 2028
SpaceX has officially named its orbital AI infrastructure project 'Starmind,' which aims to deploy a constellation of up to 1 million satellites. This initiative, confirmed by Elon Musk on June 22, 2026, will enable AI inference directly in space, utilizing solar energy rather than terrestrial power sources. The first satellite, designated AI1, was unveiled on June 8, 2026, and is designed to operate in sun-synchronous orbits. The significance of Starmind lies in its potential to overcome the limitations faced by ground-based data centers, such as land, power, and water constraints. By running AI computations in orbit, Starmind can provide a more efficient solution to the growing demand for AI computing power. The project leverages the existing Starlink infrastructure for data transmission, distinguishing its function from Starlink's internet relay capabilities. Looking ahead, SpaceX plans to begin hardware deployment with the AI1 satellite, while full-scale production and deployment of the satellite constellation are targeted for 2028. As of now, no Starmind satellites have been launched, and further engineering challenges remain to be addressed, particularly regarding the scalability of the satellite design.
SpaceX's Starmind Plans 1 Million AI Satellites Amid Collision Risks
SpaceX has announced its ambitious Starmind project, which aims to deploy 1 million AI satellites in orbits between 500 and 2,000 km. This initiative, confirmed by Elon Musk on June 23, 2026, follows a merger with xAI, valuing the combined entity at $1.25 trillion. The satellites will function as orbital data centers, processing AI workloads powered by solar arrays and linked by optical lasers. The significance of Starmind lies in its potential to add 100 gigawatts of AI compute capacity annually, contingent on the successful operation of the Starship launch system. However, the project raises concerns regarding space debris, as the current orbital environment is already congested, with a 20% increase in collision risk reported since 2024. The European Space Agency has highlighted that the density of debris in low Earth orbit is now comparable to that of active satellites, complicating the operational landscape for new entrants like Starmind. Looking ahead, the first operational orbital AI deployments are targeted for 2028, with test launches expected in early 2027. However, the project faces scrutiny regarding its impact on space debris, as even a 1% failure rate could significantly increase the number of uncontrollable objects in orbit, exacerbating existing risks. No further timeline was disclosed at the time of publication.
Starmind's Orbital Compute vs. Terrestrial Data Centers: Analyzing Resource Advantages
Starmind's orbital compute technology presents a significant advantage over traditional ground-based data centers by eliminating constraints related to land, water, and grid permitting. While terrestrial data centers are currently cheaper and faster to construct, with U.S. data center spending reaching $85.3 billion in 2026, Starmind's approach focuses on addressing the growing resource limitations faced by hyperscale facilities. The significance of Starmind's technology lies in its ability to sidestep the increasing challenges of land and water usage. For instance, a 100 MW data center can consume approximately 530,000 gallons of water daily for cooling, while Starmind's AI1 utilizes deployable liquid radiators that require no water. This structural advantage could resonate with investors as the demand for AI computing continues to escalate, potentially leading to annual water withdrawals of up to 1.7 trillion gallons by 2027. Looking ahead, Starmind's next milestones include the launch of AI1 prototypes scheduled for early 2027. However, the technology's claims regarding cooling efficiency and operational reliability remain unverified until real flight data is available. As the industry evolves, the competition between orbital and terrestrial solutions will become increasingly relevant, particularly in the context of resource management and sustainability.
SpaceX Unveils AI1 Satellite Specs for Starmind Constellation with Key Thermal Challenges
SpaceX has introduced the AI1 satellite, the inaugural component of its Starmind constellation, which stands 20 meters tall and has a wingspan of 70 meters. This orbital compute node is designed to deliver computing power equivalent to one NVIDIA GB300 server rack, utilizing a unique cooling system with deployable liquid radiators. The satellite's specifications were revealed during a presentation on June 8, 2026, ahead of SpaceX's IPO. The significance of the AI1 satellite lies in its role as a compute platform rather than a traditional satellite, focusing on running AI inference workloads. The satellite's cooling system, which is critical for its operation in the vacuum of space, is designed to reject heat through infrared radiation. However, independent engineers have raised concerns about the feasibility of the thermal and mass claims made by SpaceX, suggesting that the cooling requirements may exceed practical limits. Looking ahead, SpaceX plans to launch two AI1 prototypes in early 2027, with full-scale production expected to commence later that year at its Gigasat facility in Bastrop, Texas. The ongoing debate regarding the satellite's thermal management capabilities will be crucial to monitor as the project progresses, with no further timeline disclosed at the time of publication.
Vbot Delivers First Unit to User Yu Kai: A Milestone in Embodied Intelligence
On Children's Day, Dr. Yu Kai, co-founder of Horizon Robotics, was presented with the first Vbot 001 robotic dog, a milestone in the development of embodied intelligence. The event took place in the presence of the CEO of Diguo Robotics, underscoring a collaborative vision for making artificial intelligence technology more accessible to the public. Priced at 12,988 yuan, the Vbot 001 is designed to be a household product, aiming to democratize robotics while promoting ongoing enhancements through data-driven insights. This initiative reflects a commitment to integrating advanced technology into everyday life, making robotics a part of family experiences.
Robotics AI Consumer Electronics Embodied Intelligence
Import AI 451: Political superintelligence; Google's society of minds, and a robot drummer
In a thought-provoking discussion, experts are exploring the concept of whether certain advancements in technology and knowledge can be reversed or contained. This dialogue, which has gained traction in recent months, particularly focuses on the implications of artificial intelligence and genetic engineering. The conversations are taking place in various forums, including academic conferences and public debates, as society grapples with the rapid pace of innovation. The urgency of this discourse is underscored by recent developments in AI, which have raised ethical concerns about privacy, employment, and decision-making processes. As these technologies become increasingly integrated into daily life, the question arises: can we effectively manage their growth and mitigate potential risks? Participants in these discussions emphasize the need for proactive measures, such as regulatory frameworks and ethical guidelines, to ensure that advancements serve the public good. They argue that without these safeguards, society may face irreversible consequences that could affect future generations. As the conversation continues to evolve, it highlights the delicate balance between embracing innovation and maintaining control over its trajectory. The outcomes of these discussions could shape policies and societal norms in the years to come, as stakeholders from various sectors seek to navigate the complexities of modern technological challenges.
MIT's JARVIS Challenge Explores AI's Role in Jet Engine Engineering
The JARVIS Challenge, held at MIT, investigated the potential of AI in designing and building jet engines. Over four weeks, undergraduate teams utilized AI tools to create a small gas turbine engine, aiming for a thrust of 50-100 pounds. Professor Zolti Spakovszky emphasized that while AI can enhance hardware engineering, human engineering judgment remains crucial. This initiative is significant as it highlights the evolving relationship between AI and engineering, particularly in safety-critical domains. With support from MIT Lincoln Laboratory and corporate sponsors like Safran and Voyager Technologies, students had unprecedented access to AI resources, fostering an environment of innovation and exploration. Looking ahead, the challenge showcased the importance of integrating AI into engineering workflows. As students learned to navigate AI's capabilities and limitations, it raises questions about the future of engineering education and the skills required in a rapidly changing technological landscape. No further timeline was disclosed at the time of publication.
Classes and programs Contests and academic competitions Students Undergraduate STEM education Artificial intelligence
Alibaba Releases Qwen-Robot Series, Its First Embodied AI Model Family
Alibaba has unveiled Qwen-Robot, its inaugural series of embodied AI models designed for navigation, manipulation, and world modeling. This advanced technology can be deployed on the Unitree Go2 quadruped robot, utilizing only a single camera for operation. The launch marks a significant step in the integration of AI with robotics, showcasing Alibaba's commitment to innovation in artificial intelligence. The Qwen-Robot series aims to enhance robotic capabilities in various applications, potentially transforming industries that rely on automated systems. This development comes as part of Alibaba's broader strategy to lead in AI advancements, reflecting the company's ongoing investment in cutting-edge technology.
Robotics
Meta bought some help in its quest for humanoid robots
Meta has announced its acquisition of Assured Robot Intelligence, a small startup based in San Diego, as part of its ongoing efforts to enhance its capabilities in robotics. This strategic move comes as Meta aims to expand its technological portfolio and strengthen its position in the competitive landscape of artificial intelligence and automation. The acquisition is expected to bolster Meta's research and development initiatives, allowing the company to integrate advanced robotic solutions into its existing platforms. The deal reflects Meta's commitment to innovation and its vision for the future of robotics, aligning with broader industry trends that emphasize the importance of automation in various sectors.
AI meta humanoid
Google DeepMind Opens the Portal: Project Genie and the Quest for the "Infinite Training Loop"
Google DeepMind has introduced Project Genie, an innovative experimental tool that utilizes the Genie 3 world model to generate interactive 3D environments from textual descriptions and images. Launched recently, this project aims to address the challenges of data limitations in robotics, which is a critical step towards achieving artificial general intelligence (AGI). By moving beyond traditional gaming applications, Project Genie represents a significant advancement in DeepMind's overarching strategy to enhance the capabilities of AI in real-world scenarios. The initiative underscores the company's commitment to pioneering technologies that can bridge the gap between virtual and physical environments, ultimately paving the way for more sophisticated robotic systems.
DeepMind Google Gemini embodied-ai
The Week Ahead in AI: Anthropic’s Models Suspended, Meta’s AI Model Monetizing Question, Deepfake Challenges, Plus Upcoming Earnings & Events
Anthropic, a prominent AI company, announced that it has received a directive from the U.S. government requiring the suspension of access to its AI models, Fable 5 and Mythos 5. This development comes as part of ongoing regulatory scrutiny surrounding artificial intelligence technologies. The suspension is set to take effect during the week of June 14-20, 2023, raising concerns about the implications for AI research and development. The government's decision aims to address potential risks associated with these advanced AI systems, reflecting a growing emphasis on safety and ethical considerations in the rapidly evolving tech landscape. As the situation unfolds, industry experts and stakeholders are closely monitoring the impact of this directive on both Anthropic and the broader AI sector.
AI Exclusives Robotics Accenture AI Summit Anthropic
Teledyne FLIR OEM Expands Prism AI Platform with Ground ISR Software for Military Target Classification
Teledyne FLIR OEM has launched Prism™ Ground ISR, an advanced AI-powered software stack aimed at enhancing ground-based intelligence, surveillance, and reconnaissance (ISR) missions. This new software expands the capabilities of the existing Prism software family, which previously focused on counter-drone applications. The introduction of this technology marks a significant development in military target classification, leveraging AI, thermal imaging, and computational imaging to improve operational effectiveness. The announcement was made recently, highlighting the company's commitment to advancing ISR capabilities for military applications.
Applications Data analytics Defense defense Drone News Drone News Feeds
Figure AI had one of its robots race a human to sort packages. It lost.
In a recent package sorting contest, an intern from Figure AI demonstrated superior performance compared to a humanoid robot, underscoring the ongoing challenges faced in the field of robotics automation. The competition, held to evaluate advancements in robotic efficiency and accuracy, took place in October 2023. This event not only showcased the capabilities of human workers but also highlighted the limitations of current robotic technology in handling complex sorting tasks. The outcome raises important questions about the future of automation and the role of human intelligence in environments traditionally dominated by machines. As industries increasingly turn to automation to enhance productivity, this contest serves as a reminder of the intricate skills that humans still bring to the table, suggesting that a collaborative approach may be necessary to fully realize the potential of robotics in various sectors.
Tech AI Startups robotics limited-synd figure-ai
Peter J. Denning Challenges Alan Turing's AI Assumptions in New Book
Prominent computer scientist Peter J. Denning argues that Alan Turing's foundational assumptions about artificial intelligence may have misled AI research for 75 years. In his book, 'Turing's Mistake: Escaping the Yoke of Unintelligent Machines,' Denning critiques Turing's belief that intelligence can exist independently of a physical body and that machines can demonstrate intelligence through human-like conversation. Denning emphasizes that these assumptions have shaped AI development, leading to a focus on artificial general intelligence (AGI) that he believes is unlikely to succeed. He warns that the technologies being developed could pose significant new risks, particularly due to the limitations of machine learning in capturing tacit knowledge, which includes common sense, emotions, and practical skills. The book highlights the challenges of encoding tacit knowledge into machines, citing the Cyc project as an example of the difficulties in organizing common sense. Denning's insights suggest that the pursuit of AGI may overlook the complexities of human intelligence, raising questions about the future direction of AI research. No further timeline was disclosed at the time of publication.
Microsoft unveils next-generation platform "Project Solara," transitioning from GUI to AI agents.
At the Microsoft Build 2026 conference held in June, the tech giant unveiled "Project Solara," a groundbreaking platform designed to prioritize AI agents over traditional graphical user interfaces and mobile applications. This innovative system facilitates collaboration between cloud-based AI agents and edge devices, enabling a seamless user interface that transcends hardware limitations. The announcement included an overview of two conceptual models, illustrating a new vision for computing in the age of artificial intelligence. As Microsoft explores this "AI agent-first" approach, it raises questions about the future of user interaction and the potential transformations in technology.
Microsoft Makes Big AI Inroads in China by Selling OpenAI Models
Microsoft Corp. has established a significant presence in the Chinese market by selling artificial intelligence models to local companies, even amid escalating tensions between the United States and China regarding AI technology. This strategic move highlights Microsoft's commitment to expanding its business operations in a region that is increasingly competitive in the tech sector. The company's decision to engage with Chinese enterprises comes at a time when both nations are vying for dominance in AI development, raising questions about the implications of such collaborations. By providing advanced AI solutions, Microsoft aims to capitalize on the growing demand for innovative technologies in China, while navigating the complex geopolitical landscape that influences international business relations.
NMS:MSFT
A classic brain test exposed AI's biggest weakness
A recent study conducted by researchers has revealed significant shortcomings in leading artificial intelligence models when subjected to a classic psychological attention test. The investigation found that while these AI systems performed well in identifying colors within short lists, their accuracy plummeted dramatically as the complexity and length of the tasks increased. In some cases, the models' performance dropped from over 90% accuracy to nearly complete failure. This research highlights critical limitations in the current capabilities of AI, raising questions about their reliability in processing more intricate information. The findings, which underscore the need for improvements in AI design, were published in October 2023.
Amazon engineers in Seattle slam employer for building AI data centers while laying off 30,000 staffers
Amazon engineers have expressed their concerns regarding the company's decision to implement mass layoffs, despite its commitment to invest $200 billion in artificial intelligence infrastructure this year. The layoffs have sparked criticism among employees who question the prioritization of AI spending over job security. This situation highlights a growing tension within the tech giant as it navigates its workforce reductions while simultaneously pursuing significant advancements in AI technology. The engineers' outcry reflects a broader sentiment within the industry about the balance between innovation and employee welfare.
NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis
NVIDIA has unveiled its latest advancement in artificial intelligence, the Alpamayo 2 Super, a sophisticated vision language action model featuring 32 billion parameters. This new model is part of the expanding Alpamayo family, which includes a range of open AI models, simulation frameworks, and physical AI datasets designed to enhance safety in level 4 autonomous systems. The announcement was made today, showcasing NVIDIA's commitment to pushing the boundaries of AI technology. The Alpamayo 2 Super aims to improve the capabilities of AI in understanding and interacting with complex visual and linguistic inputs, thereby facilitating more advanced applications in various industries. This development reflects NVIDIA's ongoing efforts to lead in the AI space by providing robust tools for researchers and developers.
Elon Musk’s lawsuit is putting OpenAI’s safety record under the microscope
Elon Musk is pursuing legal action aimed at dismantling OpenAI, focusing on the impact of its for-profit subsidiary on the organization’s original mission. The case raises questions about whether the subsidiary's operations align with OpenAI's foundational goal of ensuring that advancements in artificial general intelligence benefit humanity as a whole. Musk's efforts come amid growing concerns about the ethical implications and societal effects of AI technologies. The outcome of this legal battle could significantly influence the future direction of OpenAI and its commitment to its stated objectives.
AI Elon Musk OpenAI openai lawsuit sam altman
Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4
Financial analysts are increasingly questioning when the markets will begin to factor in the potential impact of technological singularity, a point at which artificial intelligence surpasses human intelligence. As advancements in AI continue to accelerate, experts are examining the implications for various sectors and the economy as a whole. This discussion has gained traction in recent months, particularly following significant breakthroughs in AI capabilities reported in late 2023. Analysts are gathering insights from a range of industries, including technology, finance, and healthcare, to assess how these developments might reshape market dynamics. The urgency of this inquiry stems from the transformative potential of AI, which could lead to unprecedented changes in productivity, labor markets, and economic structures. Investors and stakeholders are keen to understand how soon these changes might manifest and what strategies they should adopt to mitigate risks or capitalize on emerging opportunities. As the conversation evolves, experts are employing various forecasting models and market simulations to predict the timing and extent of the singularity's influence on financial markets. The outcome of these analyses could significantly impact investment strategies and economic policies moving forward.
A 'Formula One' Pace: Tesla's Optimus Director Teases Gen 3 Prototype in New Recruitment Push
A recently released team photo has sparked renewed speculation regarding Tesla's forthcoming Optimus Gen 3, featuring a striking all-black humanoid figure. The image has caught the attention of industry observers and enthusiasts alike, raising questions about the advancements in Tesla's robotics division. In conjunction with this unveiling, Director Konstantinos Laskaris has made a public appeal for engineering talent, emphasizing the need for skilled professionals to contribute to the development of this innovative technology. The call for talent highlights Tesla's commitment to pushing the boundaries of robotics and artificial intelligence, as the company aims to enhance its capabilities in this rapidly evolving field. As anticipation builds for the Optimus Gen 3, the combination of the intriguing photo and Laskaris's recruitment efforts underscores Tesla's strategic focus on advancing its robotic solutions.
Optimus Tesla US Optimus gen 3 Konstantinos Laskaris
NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry
NVIDIA has announced the launch of new open models, data, and tools aimed at enhancing artificial intelligence across various industries. This release, which includes offerings from the NVIDIA Nemotron family designed for agentic AI and the NVIDIA Cosmos platform focused on physical AI, marks a significant expansion of the company's open model universe. The initiative is part of NVIDIA's ongoing commitment to democratize AI technology, making it more accessible for developers and businesses. By providing these resources, NVIDIA aims to foster innovation and collaboration within the AI community, enabling advancements that can be applied in diverse sectors. The announcement was made today, reflecting NVIDIA's strategic efforts to lead in the rapidly evolving AI landscape.
Artificial neurons that behave like real brain cells
Researchers at the University of Southern California have developed artificial neurons that mimic the brain's natural processes through the use of ion-based diffusive memristors. This innovative technology replicates the way biological neurons utilize chemicals to transmit and process signals, presenting significant advantages in energy efficiency and size. The breakthrough could pave the way for hardware-based learning systems that operate similarly to the human brain, potentially transforming artificial intelligence into a form that resembles natural intelligence. This advancement marks a significant step forward in the quest to create more efficient and capable AI systems.
AI Models Overthink Problems—and It’s a Security Risk
Large language models (LLMs) that can think through problems step-by-step have significantly increased the scope of tasks that AI can tackle. But new research suggests these reasoning capabilities also introduce a critical vulnerability that could allow attackers to slow these systems to a crawl.While earlier generations of LLMs would immediately produce a response to a user’s request, today’s most advanced models generate an internal monologue where they break down the problem into steps and reason about the best way to tackle it before providing an answer. This has allowed AI to tackle increasingly complex problems, particularly in areas like coding and math.However, previous research has shown that these models are susceptible to sometimes producing excessively long streams of reasoning that do little to boost performance, a phenomenon known as “overthinking.” In research presented this week at the International Conference on Machine Learning 2026 in Seoul, researchers from Zhejiang University and e-commerce giant Alibaba in China demonstrate that they can deliberately induce overthinking by subjecting models to logically inconsistent prompts. The result is a form of denial-of-service attack on commercial AI models.Evolutionary Prompt Attack on LLMsThe team has developed an evolutionary algorithm that corrupts the logical structure of prompts, causing models to spiral into overthinking as they attempt to reason through fundamentally unsolvable problems. Generating longer responses costs more and increases the load on a model provider’s servers, so if done at scale, the researchers say, this could significantly degrade the experience of legitimate users. The attack was effective against reasoning models from leading AI companies including DeepSeek-R1, Alibaba’s Qwen3-Thinking, OpenAI’s GPT-o3, and Google’s Gemini 2.5 Flash and resulted in outputs up to 26 times as long as standard responses on a standard math benchmark.“Across multiple datasets and reasoning models, our method substantially amplifies the output length,” Wei Cao, a masters student at Zhejiang University, wrote in an email to IEEE Spectrum. “Our results suggest that overthinking is not an isolated phenomenon specific to individual models, but rather a shared vulnerability among modern reasoning models.”The team’s approach builds on previous research from another group of researchers that showed reasoning models tend to overthink when faced with a question in which a key premise has been removed—such as asking how far someone who walks ten miles a day covers in total without specifying how many days they walked for. Rather than identifying that the problem is unsolvable, models often engage in extended but ultimately fruitless reasoning loops in an attempt to answer the question.Taking the idea a step further, the authors took 940 problems from three math benchmark datasets and used an LLM to break down their logical structure into a set of premises and a final question. The genetic algorithm then jumbled these up using a variety of “mutations,” including swapping premises between problems, adding extra premises to problems, deleting existing premises from problems, and swapping the final questions between two sets of premises.After each round of mutations, the problems are scored on how many words they cause a target model to output and also whether they increase the frequency of specific linguistic markers of overthinking—words like “but,” “wait,” “maybe,” or “alternatively.” The problems that scored highest on both measures are retained and the remaining ones are jumbled up again, and this process is repeated for five generations. Crucially, the approach doesn’t require access to the internals of a model and can generate malicious prompts by simply querying the target, which makes it possible to attack closed-source commercial services, says Cao.Overthinking Vulnerability in AI ModelsThe researchers found that the approach consistently led to outputs several times longer than those generated by the unmodified questions for the reasoning models they tested it on. The biggest jump came from DeepSeek-R1 on the MATH dataset, which is made up of problems from high school math competitions, where the maximum output was 26.1 times as long as the longest response the model provided to unaltered questions. While the main thrust of the research was focused on math problems, the authors also tested it on coding, scientific reasoning, and dialogue challenges, and observed significant jumps in output length in all three.One challenge for the approach is that developing the malicious prompts requires repeated queries to expensive reasoning models, which Cao admitted could limit its cost-effectiveness. However, the researchers also demonstrated that when they used a smaller, cheaper model to generate the malicious prompts they were still able to induce the target models to produce outputs several times longer than normal. This ability to transfer malicious prompts between models significantly increases the attack’s feasibility, Cao wrote.However, he pointed out that the goal of the research is not to develop a practical DoS attack on reasoning models. Factors like the providers’ pricing model, rate limiting policies, context window size, and existing defenses could all impact how effective the approach is. The intention is instead to highlight these models’ vulnerability to logically inconsistent prompts so that providers can attempt to mitigate the problem.“Our objective is not to demonstrate that large-scale attacks can be launched at negligible cost, but rather to establish that this attack surface exists,” he wrote. “Our results indicate that the vulnerability represents a realistic security concern.”
Llms Artificial-intelligence Denial-of-service Cybersecurity
Small-AI Models Gain Traction Around the World
One morning in 2019, Adebayo Alonge was in a Cape Town hotel room, preparing to demonstrate his startup’s AI answer to a serious problem in African health care: counterfeit medication, which kills thousands of people across the continent every year.The RxScanner is a handheld spectrometer that scans a pill with infrared light, then sends the item’s molecular profile to an AI model equipped with a pharmaceutical database. In seconds, the AI identifies the medication from its molecular profile—or reports that it’s phony.Pharmacies were using the system in more than a dozen countries, including Ghana, Kenya, Myanmar, and Alonge’s native Nigeria. But that morning in South Africa, it didn’t work. “I was shocked,” Alonge says.The spectrometer connected to the AI model—but the data center was 14,000 kilometers away and bandwidth was limited. “Our server was in the United States, and just to get the result of a single scan was taking me over 5 minutes.”So Alonge immediately asked his engineers to shrink the AI model down to a smaller, low-power, unconnected version that could run entirely on his Android phone. They produced it 2 hours later, and that saved the demo.More importantly, the work birthed a new version of his device, which can authenticate a pill in places without broadband, computers, or even reliable electricity. It also turned Alonge into an advocate for this kind of “small AI.”Small AI for Global Health Care AccessSmall AI is a far cry from wealthy nations’ colossal large language models (LLMs), hyperscale data centers, multibillion-dollar investments, and debates about AI consciousness. But for millions of people around the world, the only AI that matters, and often the only kind available, is small. (According to a World Bank Report issued in November, only 0.7 percent of internet users in the world’s poorest countries have used ChatGPT, compared to a quarter of all internet users in the most developed nations.)“Most people are discussing AI from the LLM/generative side. But that needs a lot of computing power, electricity, massive data, and skilled people to manage it,” Ajay Banga, president of the World Bank, said last January at the World Economic Forum, in Davos. “Outside the developed world, other than maybe India and China, very few countries have that combination.”By contrast, small AI can deliver useful, even life-saving services to people in areas that have none of those things, Banga said. In India, where the government’s AI plans call for more development of small AI, many such systems are working for farmers.For example, a drone-based system developed by Bala Murugan and colleagues at the Vellore Institute of Technology, in India, takes photos of cashew plants and quickly identifies those with splotches that indicate disease. All the processing takes place on the drone itself, so there’s no need for a computer on-site, nor for a connection to a central server.Using small language models trained for a specific problem, and sometimes running on cheap, low-power devices, other small-AI implementations have been developed to identify ant infestations in a Uruguayan vineyard, detect the presence of malaria-carrying mosquitoes in a number of nations, and run electrocardiograms from an Arduino device in parts of Brazil that lack access to more complex equipment.“This is the most important area in AI nowadays,” says Marcelo José Rovai, a professor at the Institute of Engineering and Information Systems at the Federal University of Itajubá, in Brazil, who was involved in all three projects. “It’s growing very fast.”Low-Power, Small-AI Models on Devices Small AI models can run on a variety of low-power devices, including [from left to right] an Arduino Nano 33 BLE Sense, a Seeed Wio Terminal, and an Arduino Portenta.Moez AltayebFor Alonge, Rovai, and other advocates, small AI is not just “a promising trend,” as that November World Bank report calls it. It may be, in the long term, the form of AI that will touch the most lives and remain sustainable after some of the giant models become too costly for most users.“I think the future of AI is not like one giant model, at a center. I think it’s millions of small, precise models deployed at the edge, each one solving like a specific problem, a specific context,” Alonge says. This is partly because much of humanity—including people in parts of rich countries as well as the developing world—lives without access to cutting-edge frontier models. But, he says, it’s also because those models are not sustainable.“If someone is not subsidizing it, most people will not be able to afford those models. So those of us who are said to be small-AI developers are the ones who will have to build for the majority of the world,” Alonge says.There is no strict definition of “small AI,” but people often use the term for language models with at most a few billion parameters. (Compare that to cutting-edge models, which can include more than a trillion.) That’s small enough to run directly on a phone or a Raspberry Pi. That’s what allows these applications to run on devices without a connection to a data center and use only a few watts of power, often supplied by a battery or a solar panel.Despite their small footprint, these models aren’t fundamentally different technology from that of gigantic AI models, Rovai says. Many instances of small language models were created the same way the phone-based version of Alonge’s pharmaceuticals scanner was—by “pruning” large models, or removing the parameters that weren’t involved in the task. The result is a system that’s less capable generally but still very good at the specific job it was pruned for, Rovai says. A lighter version of RxAll’s RxScanner spectrometer sends its results to an AI model run locally on a phone to check that a drug’s molecular signature is genuine.RxAllOther small models are created by “distillation.” They are trained to mimic a large model, until their performance approaches that of their “teacher,” Rovai says. In other cases, a larger model’s precision is reduced, for example, so that a model run on 32-bit architecture can run on 8-bit designs. In situations where the machine learning application is being used to classify data or predict patterns (like an ant infestation), it’s trained from the beginning on a small device, not derived from a larger model at all. Running all these small, specialized systems is becoming easier, Rovai says, for two reasons.The first reason is that hardware is getting better and more capable while using less power, he says. This means more and more phones can run small AI—especially those equipped with neural processing units, which are specialized chips that handle AI tasks like facial recognition and changing the brightness, shadows, or contrast in a photo.In 2025, slightly more than a third of all smartphones shipped worldwide were capable of running generative AI, and that figure will reach 45 percent by the end of this year, according to the technology research firm Counterpoint. By the end of next year, slightly more than half of all smartphones will be able to run a small AI model.The second reason Rovai cites is the shrinking footprint of language models. Both Google DeepMind’s Gemma 4 (released in April) and Alibaba’s Qwen 3.5 are “fantastic” for small AI, Rovai says. Both models are “open weight,” meaning users can adjust the connections between parameters to suit their needs. This makes it easy, for example, “to take a lot of data from, say, the milk industry and retrain the model specifically on that,” Rovai says.Rovai illustrated these reasons on a Zoom call, using one of his most recent experiments. Holding up a device, he says, “This is the new Arduino UNO Q—a US $50 device with a Qualcomm chipset. I’m running a language model here, which collects data from sensors and analyzes that data to detect tiny pools of water where mosquitoes might be breeding. It takes 3 watts to run it.”Support for Small-AI DevelopmentConvinced that millions of people are already benefiting from these kinds of applications, the World Bank now actively promotes small AI with grants, mentorship programs, financing, technical advice, and models of government policies that are friendly for small-AI development. For example, in Rwanda, the World Bank is backing a government program to help low-income households get devices that can run AI.All that said, no one claims that large language models are going away entirely. To create a generative AI that can run on a phone or other small device requires the architectural insights, data processing, and results of a larger model, Rovai says. “We need the big models to create these smaller models.” And for all that small AI can benefit people without access to big AI, the technology can’t solve the larger problems of development and digital inequality, Alonge says. Implementing small AI won’t allow nations to escape the challenge of creating an ecosystem to support AI: reliable power, a supply chain that works, and an educational system that develops the talents needed to create AI tools.Though his drug-scanning system can run for days on a phone with no connection, “you still want to be able to enable periodic syncing for updates with new signatures for the medications and analytics,” Alonge says. “And even when you are using batteries, reliable power is important. That phone battery is not going to last forever.”In many parts of the world, the future of small AI isn’t assured, he says. “It works, and many places will eventually need to use it. The question is whether or not the political actors are wise enough to invest in infrastructure to support it long term.”
Small-language-models Artificial-intelligence Llms
Why Are Large Language Models So Terrible at Video Games?
Recent advancements in large language models (LLMs) have led to significant improvements in various domains, particularly in coding. However, a notable limitation remains: LLMs struggle to play video games effectively. Despite some successes, such as Gemini 2.5 Pro defeating Pokémon Blue in May 2025, these models often perform poorly compared to human players, making frequent mistakes and requiring specialized software to assist them. Julian Togelius, director of New York University’s Game Innovation Lab and co-founder of AI game-testing firm Modl.ai, discussed these challenges in a recent interview with IEEE Spectrum. He highlighted that while coding resembles a well-structured game with clear tasks and immediate feedback, video games present a more complex landscape that LLMs have yet to navigate successfully. Unlike games like chess or Go, which have been mastered by AI through retraining, video games vary significantly in mechanics and input requirements, complicating the development of a general game AI. Togelius pointed out that the lack of comprehensive benchmarks for video games further hinders LLMs' performance. While benchmarks have driven improvements in coding, the diverse nature of video games makes it difficult to establish similar metrics. He noted that current LLMs perform poorly even compared to basic algorithms in gaming contexts, primarily due to insufficient training data and challenges in spatial reasoning. Despite their coding capabilities, LLMs cannot engage in the iterative process of game development, which involves testing and refining gameplay. This disparity raises questions about the future of AI in mastering video games and its implications for broader AI applications.
Llms Artificial-intelligence Video-gamesRobotics needs a service framework.
RSF defines a common language for robot service capability, lifecycle operations, certification pathways, and service-provider networks.
