VentureBeat

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Meta AI has announced the open-source release of MobileLLM, a set of language models optimized for mobile devices, with model checkpoints and code now accessible on Hugging Face. However, it is presently only available under a Creative Commons 4.0 non-commercial license, meaning enterprises can’t use it on commercial products. Originally described in a research paper published in July 2024 and covered by VentureBeat, MobileLLM is now fully available with open weights, marking a significant milestone for efficient, on-device AI. The release of these open weights makes MobileLLM a more direct, if roundabout, competitor to Apple Intelligence, Apple’s on-device/private cloud hybrid AI solution made up of multiple models, shipping out to users of its iOS 18 operating system in the U.S. and outside the EU this week. However, being restricted to research use and requiring downloading and installation from Hugging Face, it’s likely to remain limited to a computer science and academic audience for now. More efficiency for mobile devices MobileLLM aims to tackle the challenges of deploying AI models on smartphones and other resource-constrained devices. With parameter counts ranging from 125 million to 1 billion, these models are designed to operate within the limited memory and energy capacities typical of mobile hardware. By emphasizing architecture over sheer size, Meta’s research suggests that well-designed compact models can deliver robust AI performance directly on devices. Resolving scaling issues The design philosophy behind MobileLLM deviates from traditional AI scaling laws that emphasize width and large parameter counts. Meta AI’s research instead focuses on deep, thin architectures to maximize performance, improving how abstract concepts are captured by the model. Yann LeCun, Meta’s Chief AI Scientist, highlighted the importance of these depth-focused strategies in enabling advanced AI on everyday hardware. MobileLLM incorporates several innovations aimed at making smaller models more effective: • Depth Over Width: The models employ deep architectures, shown to outperform wider but shallower ones in small-scale scenarios. • Embedding Sharing Techniques: These maximize weight efficiency, crucial for maintaining compact model architecture. • Grouped Query Attention: Inspired by work from Ainslie et al. (2023), this method optimizes attention mechanisms. • Immediate Block-wise Weight Sharing: A novel strategy to reduce latency by minimizing memory movement, helping keep execution efficient on mobile devices. Performance Metrics and Comparisons Despite their compact size, MobileLLM models excel on benchmark tasks. The 125 million and 350 million parameter versions show 2.7% and 4.3% accuracy improvements over previous state-of-the-art (SOTA) models in zero-shot tasks. Remarkably, the 350M version even matches the API calling performance of the much larger Meta Llama-2 7B model. These gains demonstrate that well-architected smaller models can handle complex tasks effectively. Designed for smartphones and the edge MobileLLM’s release aligns with Meta AI’s broader efforts to democratize access to advanced AI technology. With the increasing demand for on-device AI due to cloud costs and privacy concerns, models like MobileLLM are set to play a pivotal role. The models are optimized for devices with memory constraints of 6-12 GB, making them practical for integration into popular smartphones like the iPhone and Google Pixel. Open but non-commercial Meta AI’s decision to open-source MobileLLM reflects the company’s stated commitment to collaboration and transparency. Unfortunately, the licensing terms prohibit commercial usage for now, so only researchers can benefit. By sharing both the model weights and pre-training code, they invite the research community to build on and refine their work. This could accelerate innovation in the field of small language models (SLMs), making high-quality AI accessible without reliance on extensive cloud infrastructure. Developers and researchers interested in testing MobileLLM can now access the models on Hugging Face, fully integrated with the Transformers library. As these compact models evolve, they promise to redefine how advanced AI operates on everyday devices. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A common criticism of big tech is that their platforms treat users as little more than glassy eyeballs to be monetized with targeted ads. This will soon change, but not because tech platforms are moving away from aggressively targeting users. Instead, our ears are about to become the most efficient channel for hammering us with AI-powered influence that is responsive to the world around us. Welcome to the Whisperverse.    Within the next few years, an AI-powered voice will burrow into your ears and take up residence inside your head. It will do this by whispering guidance to you throughout your day, reminding you to pick up your dry cleaning as you walk down the street, helping you find your parked car in a stadium lot and prompting you with the name of a coworker you pass in the hall. It may even coach you as you hold conversations with friends and coworkers, or when out on dates, give you interesting things to say that make you seem smarter, funnier and more charming than you really are. These will feel like superpowers. The ‘Whisperverse’ will require highly advanced technology Of course, you won’t be the only one “augmented” with context-aware AI guidance. Everyone else will have similar abilities. This will create an arms race among the public to embrace the latest AI-powered enhancements. It will not feel like a choice, because not having these capabilities will put you at a cognitive disadvantage. This is the future of mobile computing. It will transform the bricks we carry around into body-worn devices that see and hear our surroundings and covertly offer useful information and friendly reminders at every turn. Most of these devices will be deployed as AI-powered glasses because that form-factor gives the best vantage point for cameras to monitor our field of view, although camera-enabled earbuds will be available too. The other benefit of glasses is that they can be enhanced to display visual content, enabling the AI to provide silent assistance as text, images, and realistic immersive elements that are integrated spatially into our world. Also, sensored glasses and earbuds will allow us to respond silently to our AI assistants with simple head nod gestures of agreement or rejection, as we naturally do with other people.   This future is the result of two technologies maturing and merging into one — AI and augmented reality. Their combination will enable AI assistants to ride shotgun in our lives, observing our world and giving us advice that is so useful, we will quickly feel like we can’t live without it. Of course there are serious privacy concerns, not to mention the risk of AI-powered persuasion and manipulation, but what choice will we have? When big tech starts selling superpowers, to not have these abilities will mean being at a disadvantage socially, professionally, intellectually and economically. ‘Augmented mentality’ changing our lives I’ve been writing about our augmented future for more than 30 years, first as a researcher at Stanford, NASA and the U.S. Air Force, and then as a professor and entrepreneur. When I first started working in the field we now call “augmented reality,” that phrase didn’t exist, so I described the concept as “perceptual overlays” and showed for the first time that AR could significantly enhance human abilities. These days, there is a similar lack of words to describe the AI-powered entities that will sit on our shoulders and coach us through our day. I often refer to this emerging branch of computing as “augmented mentality” because it will change how we think, feel and act.    Whatever we end up calling this new field, it is coming soon and it will mediate our lives, assisting us at work, at school or even when grabbing a late-night snack in the privacy of our own kitchen. If you are skeptical, you’ve not been tracking the massive investment and rapid progress made by Meta on this front and the arms race they are stoking with Apple, Google, Samsung and other major players in the mobile market. It is increasingly clear that by 2027, this will become a major battleground in the mobile device industry. The first of these devices is already on the market — the AI-powered Ray-Bans from Meta. Although currently a niche product, I believe it is the single most important mobile device being sold today. That’s because it follows the new paradigm that will soon define mobile computing: Context aware guidance. To enable this, the Meta Ray-Bans have onboard cameras and mics that feed a powerful AI engine and pumps verbal guidance into your ears. At Meta Connect in September, the company showcased new consumer-focused features for these glasses, such as helping users find their parked cars, translating languages in real-time and naturally answering questions about things you see in front of you. ‘Cute’ creatures rather than ‘creepy’ ones Of course, the Meta Ray-Bans are just a first step. The next step is to visually enhance your experience as you navigate your world. Also in September, Meta unveiled their prototype Orion glasses that deliver high quality visual content in a form factor that is finally reasonable to wear in public. The Orion device is not planned for commercial deployment, but it paves the way for consumer versions to follow. So, where is this all headed? By the early 2030’s, I predict the convergence of AI and augmented reality will be sufficiently refined that AI assistants will appear as photorealistic avatars that are embodied within our field of view. No, I don’t believe they will be displayed as human-sized virtual assistants who follow us around all day.  That would be creepy. Instead, I predict they will be rendered as cute little creatures that fly out in front of us, guiding us and informing us within our surroundings. Back in 2020, I wrote a short story (Carbon Dating) for a sci-fi anthology in which I refer to these AI assistants as

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Nvidia announced that its Nvidia AI Blueprint will make it easy for developers in any industry to build AI agents to analyze video and image content. With this technology, Nvidia said any industry can now search and summarize vast volumes of visualdata. Accenture, Dell and Lenovo are among the companies tapping a new Nvidia AI Blueprint to develop visual AI agents that can boost productivity, optimize processes and create safer spaces. Enterprises and public sector organizations around the world are developing AI agents to boost the capabilities of workforces that rely on visual information from a growing number of devices — including cameras, IoT sensors and vehicles. To support their work, a new Nvidia AI Blueprint for video search and summarization will enable developers in virtually any industry to build visual AI agents that analyze video and image content. These agents can answer user questions, generate summaries and enable alerts for specific scenarios. Part of Nvidia Metropolis, a set of developer tools for building vision AI applications, the blueprint is a customizable workflow that combines Nvidia computer vision and generative AI technologies. Global systems integrators and technology solutions providers including Accenture, Dell and Lenovo are bringing the Nvidia AI Blueprint for visual search and summarization to businesses and cities worldwide, jump-starting the next wave of AI applications that can be deployed to boost productivity and safety in factories, warehouses, shops, airports, traffic intersections and more. Announced ahead of the Smart City Expo World Congress, the Nvidia AI Blueprint gives visual computing developers a full suite of optimized software for building and deploying generative AI-powered agents that can ingest and understand massive volumes of live video streams or data archives. Users can customize these visual AI agents with natural language prompts instead of rigid software code, lowering the barrier to deploying virtual assistants across industries and smart city applications. Nvidia AI Blueprint harnesses vision language models Visual AI agents are powered by vision language models (VLMs), a class of generative AI models that combine computer vision and language understanding to interpret the physical world and perform reasoning tasks. The Nvidia AI Blueprint for video search and summarization can be configured with Nvidia NIM microservices for VLMs like Nvidia VILA, LLMs like Meta’s Llama 3.1 405B and AI models for GPU-accelerated question answering and context-aware retrieval-augmented generation. Developers can easily swap in other VLMs, LLMs and graph databases and fine-tune them using the Nvidia NeMo platform for their unique environments and use cases. Adopting the Nvidia AI Blueprint could save developers months of effort on investigating and optimizing generative AI models for smart city applications. Deployed on Nvidia GPUs at the edge, on premises or in the cloud, it can vastly accelerate the process of combing through video archives to identify key moments. In a warehouse environment, an AI agent built with this workflow could alert workers if safety protocols are breached. At busy intersections, an AI agent could identify traffic collisions and generate reports to aid emergency response efforts. And in the field of public infrastructure, maintenance workers could ask AI agents to review aerial footage and identify degrading roads, train tracks or bridges to support proactive maintenance. Beyond smart spaces, visual AI agents could also be used to summarize videos for people with impaired vision, automatically generate recaps of sporting events and help label massive visual datasets to train other AI models. The video search and summarization workflow joins a collection of Nvidia AI Blueprints that make it easy to create AI-powered digital avatars, build virtual assistants for personalized customer service and extract enterprise insights from PDF data. Nvidia AI Blueprints are free for developers to experience and download, and can be deployed in production across accelerated data centers and clouds with Nvidia AI Enterprise, an end-to-end software platform that accelerates data science pipelines and streamlines generative AI development and deployment. AI agents to deliver insights from warehouses to world capitals Enterprise and public sector customers can also harness the full collection of Nvidia AI Blueprints with the help of Nvidia’s partner ecosystem. Global professional services company Accenture has integrated Nvidia AI Blueprints into its Accenture AI Refinery, which is built on Nvidia AI Foundry and enables customers to develop custom AI models trained on enterprise data. Global systems integrators in Southeast Asia — including ITMAX in Malaysia and FPT in Vietnam — are building AI agents based on the video search and summarization Nvidia AI Blueprint for smart city and intelligent transportation applications. Developers can also build and deploy Nvidia AI Blueprints on Nvidia AI platforms with compute, networking and software provided by global server manufacturers. Nvidia AI Blueprints are incorporated in the Dell AI Factory with Nvidia and Lenovo Hybrid AI solutions. Companies like K2K, a smart city application provider in the Nvidia Metropolis ecosystem, will use the new Nvidia AI Blueprint to build AI agents that analyze live traffic cameras in real time. This will enable city officials to ask questions about street activity and receive recommendations on ways to improve operations. The company also is working with city traffic managers in Palermo, Italy, to deploy visual AI agents using NIM microservices and Nvidia AI Blueprints. Nvidia will talk more about this at the Smart Cities Expo World Congress, taking place in Barcelona through Nov. 7. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Disclaimer: I voted for Kamala Harris in the 2024 presidential election and stand by my choice. Republican politician and businessman Donald J. Trump has won the 2024 U.S. presidential election in a strong political comeback, despite various pre-election polls showing him neck-and-neck with his opponent Kamala Harris (the current and now outgoing Vice President, a Democrat). As many who follow the news know all too well, one of his most outspoken allies in this election was none other six-company owner/operator and technology multibillionaire Elon Musk, who committed tens of millions in funding to a political action committee advocating Trump’s re-election. All of Musk’s technology companies stand to benefit from Trump’s return to office Musk owns or operates the following companies, all of which stand to benefit from Trump retaking power: Tesla Motors: Though Trump has pledged not to enforce electric vehicle mandates or tighter emissions standards, Musk’s popular electric vehicle and autonomous vehicle company could benefit from loosened restrictions on vehicle standards overall, especially with regards to autonomy. Already, Tesla stock is up more than 13% today on the election being called for Trump. SpaceX: Musk’s rocketry and spacefaring company has feuded before with the federal government, particularly the Federal Aviation Administration (FAA) which just last month levied $633,009 in fines to SpaceX alleging it failed to “follow its license requirements during two launches in 2023.” Musk would likely seek to use the Trump Administration to recall this fine and remove future licensing requirements preventing what he sees as necessary speed and nimbleness from the agency or a more “hands-off” approach. Starlink: Similarly, Musk’s satellite internet offshoot Starlink, which currently has more than 6,000 satellites beaming internet from orbit, would likely benefit from Trump’s pledges to reduce administrative burdens and red tape from federal regulatory agencies such as the Federal Communications Commission (FCC) and FAA. Neuralink: Musk’s experimental brain implant company has reportedly caused the death, injury and dismemberment of test monkeys but also also been successfully implanted into a paralyzed human patient, allowing them to control a computer with their brain signals. Given it is a medical device, it is overseen by the federal Food and Drug Administration, which has already approved Neuralink implantation in humans and trials. But the Trump victory will likely only further clear the way for Neuralink to ramp up its trials on more human subjects and do so faster. X: Musk’s social network, the renamed Twitter he purchased for $44 billion two years ago, has already been through a process of mass and targeted layoffs, as well as policy and feature updates permitting more freewheeling and extremist speech and content, and led to a more political right-wing oriented userbase and content. This trend is likely to continue and X to gain even more prominence as a mouthpiece for Musk’s, Trump’s, and their allies’ positions. xAI may benefit and move from being a runner-up in the AI race to a leader But most importantly of all, xAI, Musk’s AI startup offshoot of X designed to rival his former company OpenAI, is now likely to become far more of a viable alternative to the U.S. government and military as a contractor and AI technology services provider. Already, the U.S. government has been courted by and is reportedly working with OpenAI, Anthropic, and Meta to integrate generative AI models into various departments. However, now that Musk helped propel Trump to a victory, expect xAI to join in the list of federally approved AI vendors and possibly even preferred AI vendors — though of course, the government is technically supposed to remain vendor-neutral for companies operating within the U.S., signing contracts based on request-for-proposals and the businesses’ fitness for the job. xAI will also likely benefit from repealed Biden-era AI Executive Order Yet as AI influencer Andrew Curran noted on Musk’s X network this morning, another direct outcome of Trump reassuming the White House come January 2025 (when he is to be sworn in) is a strong likelihood — outlined in the Republican Party election platform — of the repeal of outgoing President Joe Biden’s Executive Order (EO) on AI, which Biden issued in October 2023 and requires developers of powerful foundation models to share safety test results and other critical information with the US government and subjects companies training AI models to red-teaming exercises by the federal agency The National Institute of Standards and Technology (NIST). While many in the AI industry and outside of it applauded this order as a means of ensuring safety of AI deployments on American and global society, some analysts suggested it could lead to undermining U.S. AI competitiveness on the global stage, both in the commercial (direct-to-consumer and business-to-business) marketplace and the military arena. As such, with the likely repeal of this EO come January 2025 or early 2025, it could aid xAI and its competitors in shipping new models faster — though as we’ve seen with xAI’s Grok-2 and its permissive image generation feature, that can also lead to a rise in deepfakes and other wild, offensive, but also creative and imaginative AI imagery. Either way, things are looking good for Musk’s companies and xAI in particular – and that may help the company’s models become more enticing to developers and business customers. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More We’ve known it for some time, but now it’s certain: The generative AI race is as much a contest for developers as it is for end-users. Case-in-point: Today, Elon Musk’s xAI, the spinoff startup of the social network X that uses its data to train new large language models (LLMs) such as the Grok family, announced its application programming interface (API) is now open to the public and with it comes $25 free per month in API credits through the end of the year. Given it’s already November, that’s just 2 months worth of free credits, or $50 total. xAI's API is live! – try it out @ https://t.co/BZD8ZyOTTY* 128k token context * Function calling support* Custom system prompt support* Compatible with OpenAI & Anthropic SDKs* $25/mo in free credits till EOYhttps://t.co/CCQAry6d5w https://t.co/MEEU2wkstS — xAI (@xai) November 4, 2024 Musk previously announced the xAI API was open in beta three weeks ago to the date, but apparently uptake was not enough for his liking, hence the added incentive of free dev credits. Is $25 per month with 2 months remaining really that much of a carrot? It doesn’t sound like much coming from the world’s wealthiest man and multi-billionaire, and it’s not really on a per user basis nor in aggregate, but it may be enough to entice some developers to at least check out xAI’s tools and platform for building apps atop of the Grok models. Specifically, xAI’s API is priced at $5 per million input tokens and $15 per million output, compared to $2.50/$10 for OpenAI’s GPT-4o model and at $3/$15 for Anthropic’s Claude 3.5 Sonnet model. Ultimately, that means xAI’s $25 credit won’t get the developer very far — only about two million tokens in and one million out per month. For reference, a million tokens is equivalent to 7-8 novels worth of words. The context limit, or how many tokens can be inputted or outputted in one interaction through the API, is around 128,000, similar to OpenAI’s GPT-4o and below Anthropic’s 200,000 token window, and well below Google Gemini 1.5 Flash’s 1-million context window length. Also, from my brief test of the xAPI, I was only able to access grok-beta and text only, no image generation capabilities such as those found on Grok 2 (powered by Black Forest Labs’ Flux.1 model). New Grok models coming soon According to xAI’s blog post, this is actually “a preview of a new Grok model that is currently in the final stages of development,” and a new Grok “vision model will be available next week.” In addition, xAI notes that the grok-beta supports “function calling,” or the ability for the LLM to take commands from a user and access functions of other connected apps and services, even executing them on the user’s behalf (if the connected app allows such access). Compatible with the competition Furthermore, the xAI account on the social network X posted that the xAI API is “compatible with OpenAI & Anthropic SDKs,” or the software development kits of different web tools used by developers of those platforms, meaning it should be relatively easy to switch out those models for grok-beta or others on the xAI platform. Musk’s xAI recently switched on its “Colossus” supercluster of 100,000 Nvidia H100 GPUs in Memphis, Tennessee, which is being used to train its new models — the largest or one of the largest in the world — so apparently that facility is already hard at work. What do you think? Is it enough to get the developers out in the VentureBeat audience to try building atop xAI? Let me know: [email protected]. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More MIPS released its P8700 CPU based on the RISC-V computing architecture to target driver assistance and autonomous vehicle applications. The San Jose, California-based company, which focuses on developing efficient and configurable intellectual property compute, licenses its designs to other chip makers. Today, it is announcing the general availability launch of the MIPS P8700 Series RISC-V Processor. Designed to meet the low-latency, highly intensive data movement demands of the most advanced automotive applications such as ADAS (advanced driver assistance system) and autonomous vehicles, the P8700 delivers accelerated compute, power efficiency and scalability, said Sameer Wasson, CEO of MIPS, in an interview with VentureBeat. “Automotive is a big segment where we focus. It continues being a very exciting place. Some companies came and some disappeared,” Wasson said. “They lost interest. They came out of COVID and refilled their inventory. But what’s happening in the industry right now is very interesting. I think autonomy is now coming back to that steady growth rate.” He added, “It is one of the biggest driving forces to continue innovating in terms of bringing better solutions. If you think about the solutions today, most of the deployments in vehicles are driven by what used to be vehicle technology. That was basic microcontrollers, simple stuff. They could open and close doors, run internal combustion engines. As autonomy grows, you’re going to see compute needs evolve toward more AI network compute. That allows you to have higher levels of autonomy.” “We have technology and we have a play in making it much more mainstream than it has been,” he said. MIPS has been supplying Mobileye with processors like the P8700 for years. Typical solutions for ADAS and autonomous driving rely on a brute-force approach of embedding a higher number of cores at higher clock rates driving synthetic, albeit unrealistic and unrealized performance. The P8700 with its multi-threaded and power-efficient architecture allows MIPS customers to implement fewer CPU cores and much lower thermal design power (TDP) than the current market solutions, thereby allowing OEMs to develop ADAS solutions in an affordable and highly scalable manner. It also mitigates the system bottlenecks of data movement inefficiency by providing highly efficient, optimized and lower power latency sensitive solution specifically tailored for interrupt laden multi-sensor platforms. “If you look at the RISC-V space overall, I think these spaces are ready for disruption, with a chance for new architectures coming in,” Wasson said. “Otherwise, EVs will be much more expensive than they need to be.” For Level 2 or higher ADAS systems with AI Autonomous software stack, the MIPS P8700 can also offload core processing elements that cannot be easily quantized in deep learning and reduced by sparsity-based convolution processing functions, resulting in a greater than 30% better AI Stack software utilization and efficiency. “The automotive market demands CPUs which can process a large amount of data from multiple sensors in real-time and feed the AI Accelerators to process in an efficient manner,” said Wasson. “The MIPS Multi-threading and other architectural hooks tailored for automotive applications, make it a compelling core for data intensive processing tasks. This will enable automotive OEMs to have high performance compute systems which consume less power and better utilize of AI Accelerators.” The MIPS P8700 core, featuring multi-core/multi-cluster and multi-threaded CPU IP based on the RISC-V ISA, is now progressing toward series production with multiple major OEMs. Key customers like Mobileye (Nasdaq: MBLY) have embraced this approach for future products for self-driving vehicles and highly automated driving systems. “MIPS has been a key collaborator in our success with the EyeQ™ systems-on-chip for ADAS and autonomous vehicles,” said Elchanan Rushinek, executive vice president of engineering for Mobileye, in a statement. “The launch of the MIPS P8700 RISC-V core will help drive our continued development for global automakers, enabling greater performance and excellent efficiency in cost and power usage.”  The P8700 Series is a high-performance out-of-order processor that implements the RISC-V RV64GC architecture, including new CPU and system-level features designed for performance, power, area form factors and additional proven features built on legacy MIPS micro-architecture deployed in more than 30+ car models today across the global OEM market. Mobileye chip for vehicles with P8700 CPU from MIPs. Engineered to deliver industry-leading compute density, MIPS’ latest processor harnesses three key architectural features, including MIPS out-of-order multi-threading, which enables execution of multipleinstructions from multiple threads (harts) every clock cycle, providing higher utilization and CPU efficiency. It also has coherent multi-core, multi-cluster, where the P8700 Series scales up to 6 coherent P8700 cores in a cluster with each cluster supporting direct attach accelerators. And it has functional safety designed to meet the ASIL-B(D) functional safety standard (ISO26262) by incorporating several fault detection capabilities such as end-to-end parity protection on address and data buses, parity protection on software visible registers, fault bus for reporting faults to the system, andmore. The MIPS P8700 processor is now available to the broader market, with key partnerships already in place. Shipments with OEM launches are expected shortly. MIPS has been around for three decades and billions of its chips have shipped to date. In the past, Wasson said vendors were using the wrong computer architecture, which was built for entertainment and screen applications, rather than hardcore AI problems. “What we are trying to do is go focus on building compute for ADAS and higher levels of autonomy, from the ground up,” he said. Vasanth Waran, worldwide head of business development at MIPS, said in an interview with VentureBeat that other architectures have been pushing performance forward through brute force, adding more complexity and scaling, but not necessarily coming up with affordable designs. “If you want to bring it to a larger market, you want autonomy to be affordable, and you want it to scale,” Waran said. “There needs to be a more pure approach, given the lack of a better word, and that’s what motivated us. The 8700 from the

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Tech leaders said they are ready to work with the new Trump administration, stating that American leadership in AI and the government’s focus on tech policies must be ensured.  Throughout the campaign, Donald Trump and his running mate, JD Vance, presented a tech industry-friendly approach and courted personalities like Elon Musk to shore up support from the sector. AI companies, like Musk’s xAI, could greatly benefit from this more tech-focused administration, especially if the Biden administration’s flagship AI executive order is repealed.  OpenAI CEO Sam Altman congratulated Trump, adding, “It is critically important that the US maintains its lead in developing AI with democratic values.” Greg Brockman, OpenAI president, echoed the same sentiment, pointing out that he believes it is with technology and AI that the country can “continue to lead the world and protect democratic values.” Perplexity CEO Aravind Srinivas also took to social media to offer his congratulations.  “USA is the land of dreams, opportunity and competition. Look forward to working with the new government to improve how people search for information online with AI,” he said.  Srinivas also touted Perplexity’s election information hub. According to Srinivas, around 10% of Perplexity usage on November 5 revolved around the elections.  Sundar Pichai, CEO of Google and its parent company, Alphabet, said the US is undergoing a “golden age of innovation.” Apple CEO Tim Cook, who is starting to roll out more AI features on its devices, also promised to work with the administration. LinkedIn CEO Reid Hoffman, an outspoken supporter of Kamala Harris, expressed the need to “get to the hard work of bridging divisions and ensuring that all Americans can enjoy safe, secure, and prosperous futures.” Change in policies The Biden administration has been vocal in seeking to support AI innovation with balancing privacy protections, culminating in the AI executive order in October last year. Since then, the government began looking into the potential dangers of open-weight models and asked companies like OpenAI and Anthropic to submit their unreleased AI models for safety evaluations.  Harris, who ran against Trump instead of President Joe Biden, represented the US in international gatherings on AI safety and regulation.  Tech companies faced scrutiny during the Biden administration as the government put forward several anti-trust cases. The Department of Justice, after winning its monopoly case against Google, put forward a potential plan to break up the tech giant.  Game company Epic won against Google, accusing the search giant of monopoly. Epic’s lawsuit against Apple, however, failed. The DOJ filed a separate antitrust case against Apple in March.  A more tech-friendly administration may mean a less litigious DOJ or Federal Trade Commission and fewer antitrust lawsuits, though Trump previously sued tech companies in his first term.  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI transformed its popular ChatGPT service into a powerful search engine today, marking the company’s boldest move yet to compete with Google. The upgrade lets users ask questions in plain English and get real-time information about news, sports, stocks, and weather — features that until now required a separate search engine. “We believe finding answers should be as natural as having a conversation,” an OpenAI spokesperson told VentureBeat. The company will roll out the feature first to paying subscribers, with plans to expand to free users in coming months. ChatGPT Search: How OpenAI’s new AI-powered web search actually works Unlike traditional search engines (i.e. Google and Bing) that return a list of links, ChatGPT now processes questions in natural language and delivers curated answers with clear source attribution. Users can click through to original sources or ask follow-up questions to dig deeper into topics. The technology builds on OpenAI’s SearchGPT experiment from July, which tested the search features with 10,000 users. That limited release helped the company refine how its AI processes web information and attributes sources. The system runs on a specialized version of GPT-4o, OpenAI’s most advanced AI model. The company trained it on massive amounts of web data and fine-tuned it to understand context across longer conversations. Major news publishers partner with OpenAI to power next-generation search results Major news organizations including the Associated Press, Axel Springer, and Vox Media have partnered with OpenAI to provide content. The deals aim to address long-standing concerns about AI systems using publishers’ work without permission or payment. “ChatGPT search promises to better highlight and attribute information from trustworthy news sources, benefiting audiences while expanding the reach of publishers like ourselves who produce premium journalism,” said Pam Wasserstein, President of Vox Media, in a statement. Publishers can opt out of having their content used for AI training while still appearing in search results. Inside OpenAI’s $5 billion bet on custom chips and AI infrastructure The launch comes as OpenAI races to build its own technology infrastructure. The company recently announced deals with AMD, Broadcom, and TSMC to develop custom AI chips by 2026 — a move to reduce its reliance on Nvidia’s expensive processors. These investments don’t come cheap. Microsoft, OpenAI’s biggest backer with nearly $14 billion invested, said this week the partnership will cut into its quarterly profits by $1.5 billion. OpenAI itself expects to spend $5 billion this year on computing costs. This massive investment in custom silicon and infrastructure signals a crucial shift in OpenAI’s strategy. While most AI companies remain dependent on Nvidia’s chips and cloud providers’ data centers, OpenAI is making an ambitious play for technological independence. It’s a risky bet that could either drain the company’s resources or give it an insurmountable advantage in the AI arms race. By controlling its own chip destiny, OpenAI could potentially cut its computing costs in half by 2026. More importantly, custom chips optimized specifically for GPT models could enable capabilities that aren’t possible with general-purpose AI processors. This vertical integration — from chips to models to consumer products— mirrors the playbook that helped Apple dominate smartphones. The new search features will appear on ChatGPT’s website and mobile apps. Enterprise customers and educational users will get access in the next few weeks, followed by a gradual rollout to OpenAI’s millions of free users. For now, Google remains the dominant force in search. But as AI technology improves and more users grow comfortable with conversational interfaces, the competition for how we find information online appears poised for its biggest shake-up in decades. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Enterprises looking to deploy multiple AI agents often need to implement a framework to manage them.  To this end, Microsoft researchers recently unveiled a new multi-agent infrastructure called Magnetic-One that allows a single AI model to power various helper agents that work together to complete complex, multi-step tasks in different scenarios. Microsoft calls Magnetic-One a generalist agentic system that can “fully realize the long-held vision of agentic systems that can enhance our productivity and transform our lives.” The framework is open-source and available to researchers and developers, including for commercial purposes, under a custom Microsoft License. In conjunction with the release of Magnetic-One, Microsoft also released an open-source agent evaluation tool called AutoGenBench to test agentic systems, built atop its previously released Autogen framework for multi-agent communication and cooperation. The idea behind generalist agentic systems is to figure out how autonomous agents can solve tasks that require several steps to finish that are often found in the day to day running of an organization or even an individual’s daily life.  From the examples Microsoft provided, it looks like the company hopes Magnetic-One fulfills almost mundane tasks. Researchers pointed Magnetic-One to tasks like describing trends in the S&P 500, finding and exporting missing citations, and even ordering a shawarma.  How Magnetic-One works Magnetic-One relies on an Orchestrator agent that directs four other agents. The Orchestrator not only manages the agents, directing them to do specific tasks, but also redirects them if there are errors. The framework is composed of four types of agents other than the Orchestrator: Websurfer agents can command Chromium-based web browsers and navigate to websites or perform web searches. It can also click and type, similar to Anthropic’s recently released Computer Use, and summarize content.  FIleSurfer agents read local files list directories and go through folders. Coder agents write codes, analyze information from other agents and create new artifacts. ComputerTerminal provides a console where the Coder agent’s programs can be executed.  The Orchestrator directs these agents and tracks their progress. It starts by planning how to tackle the task. It creates what Microsoft researchers call a task ledger that tracks the workflow. As the task continues, the Orchestrator builds a progress ledger “where it self-reflects on task progress and checks whether the task is completed.” The Orchestrator can assign an agent to complete each task or update the task ledger. The Orchestrator can create a new plan if the agents remain stuck.  “Together, Magentic-One’s agents provide the Orchestrator with the tools and capabilities that it needs to solve a broad variety of open-ended problems, as well as the ability to autonomously adapt to, and act in, dynamic and ever-changing web and file-system environments,” the researchers wrote in the paper.  While Microsoft developed Magnetic-One using OpenAI’s GPT-4o — OpenAI is after, all a Microsoft investment — it is LLM-agnostic, though the researchers “recommend a strong reasoning model for the Orchestrator agent such as GPT-4o.”  Magnetic-One supports multiple models behind the agents, for example, developers can deploy a reasoning LLM for the Orchestrator agent and a mix of other LLMs or small language models to the different agents. Microsoft’s researchers experimented with a different Magnetic-One configuration “using OpenAI 01-preview for the outer loop of the Orchestrator and for the Coder, while other agents continue to use GPT-4o.” The next step in agentic frameworks Agentic systems are becoming more popular as more options to deploy agents, from off-the-shelf libraries of agents to customizable organization-specific agents, have arisen. Microsoft announced its own set of AI agents for the Dynamics 365 platform in October.  Tech companies are now beginning to compete on AI orchestration frameworks, particularly systems that manage agentic workflows. OpenAI released its Swarm framework, which gives developers a simple yet flexible way to allow agents to guide agentic collaboration. CrewAI’s multi-agent builder also offers a way to manage agents. Meanwhile, most enterprises have relied on LangChain to help build agentic frameworks.  However, AI agent deployment in the enterprise is still in its early stages, so figuring out the best multi-agent framework will continue to be an ongoing experiment. Most AI agents still play in their playground instead of talking to agents from other systems. As more enterprises begin using AI agents, managing that sprawl and ensuring AI agents seamlessly hand off work to each other to complete tasks is more crucial.  source