Tech Republic

eSecurity Planet content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More Threat actors have found a way to weaponize trust itself. By bending X’s AI assistant to their will, they’re turning a helpful tool into a malware delivery engine. Hackers have turned X’s flagship AI assistant, Grok, into an unintentional accomplice in a massive malware campaign. By manipulating the platform’s ad system and exploiting Grok’s trusted voice, cybercriminals are smuggling poisoned links into promoted posts that look legitimate… and then using Grok to “vouch” for them. The scheme fuses the reach of paid advertising with the credibility of AI-generated responses, creating a perfect storm for unsuspecting users. Security researchers warn that the method has already exposed millions of people to malicious websites, proving that even AI designed to inform and protect can be hijacked to deceive. How ‘Grokking’ works It starts with an ad, but it ends with a trap. What looks like a harmless promotion hides a toxic payload beneath the surface. Researchers at Guardio Labs, led by Nati Tal, uncovered the technique in an age-restricted X post on Sept. 4 and dubbed it “Grokking.” Attackers hide malicious URLs in the “From:” metadata of video-card promoted posts — content X does not vet. These ads often use sensational or adult themes to lure users while concealing the actual link from moderators. Next, the attackers reply to their own ads tagging Grok, saying something like “Where is this video from?” or “What’s the link to this video?” Grok, trusted by X as a system account, reads the hidden metadata and publicly reveals the link in its reply. The result? Malware-laden links receive the twin boost of paid ad amplification and Grok’s credibility, a powerful combination that can generate hundreds of thousands to millions of impressions. Image: Screenshot (X) Dangerous AI repackaging: Grok, Mixtral, and WormGPT’s return If criminals can twist Grok into a weapon, they can do the same with any AI. And that’s exactly what’s happening. This Grokking scheme is just one prong of a growing wave of AI-enabled cybercrime. Security researchers have discovered new malicious AI variants, reviving the notorious WormGPT, built atop mainstream models like X’s Grok and Mistral’s Mixtral. According to Cato Networks, threat actors are wrapping these commercial LLMs in jailbroken interfaces that ignore safety guardrails. One variant surfaced on BreachForums in February under the guise of an “Uncensored Assistant” powered by Grok. Another emerged in October as a Mixtral-based version. For a few hundred euros, criminals gain access to AI tools specialized in crafting phishing emails, generating malware, code payloads, and even tutorials for novice hackers — without needing deep AI expertise. This alarming trend highlights that the risk lies not in the AI models themselves, but in how adversaries exploit system prompts to bypass safety filters and repurpose LLMs as “cybercriminal assistants.” This isn’t the only crisis X has faced lately. Earlier this year, Elon Musk attributed a platform-wide outage to a cyberattack. Here’s how the situation unfolded. source

Welcome, humans. The most slept-on AI release in last week’s AI onslaught was definitely Genie 3. It’s truly breathtaking. Here’s one example: DeepMind worked with the artist Rui Huang and recreated one of their pieces of art and turned it into a playable demo: DeepMind CEO Demis Hassabis sat down with Logan Kilpatrick to talk about the path forward for Genie 3, and how Google could be about to combine many of its top capability models to create a single omni model that “can do everything.” But the most interesting part of the interview was the section on how to actually release the Genie: potentially as some sort of social gaming feed where you can share your worlds with others to play (consider this a +1 upvote for that idea, plz & ty!). The most unexpected part of the interview? Probably the prospect of DeepMind eventually becoming a gaming company (y’know, once they solve AGI)… Here’s what happened in AI today: Vercel released v0.app to build complete apps from text prompts. Claude added conversation memory and search features. NVIDIA and AMD can sell AI chips to China with 15% US government cut. Meta won brain modeling competition with TRIBE prediction system. Advertise in The Neuron here Vercel just made making AI apps as easy as its ever been (and turned everyone into a developer) Vercel just dropped v0.app, and it’s basically trying to make every product manager, marketer, and designer into a software developer overnight. The company rebranded its v0.dev vibe-coding tool to v0.app with a bold promise: describe any app idea in a single prompt, and you’ll get a complete working application—frontend, backend, copy, and logic included. Here’s the key difference: with the old v0.dev, you’d prompt for something, then prompt again for fixes, better design, or added functionality. v0.app handles all that automatically using what they call “agentic intelligence”; it plans, adjusts, and improves the code without being told. With 3.5M+ users, v0 has struck a nerve. But here’s what’s interesting: while it started as a developer tool, most of its growth came from non-developers. Companies like Cox Enterprises, WPP, and Belk are now using it for entire teams of PMs, designers, and marketers. Here’s what v0.app does for you automatically: Plans out work step-by-step for your request. Searches the web and returns results with citations. Reads files and inspects live websites with screenshots. Generates design inspiration and manages to-dos. Spots errors and implements integrations with your favorite tools. According to GM Zeb Hermann, they “learned from other agentic projects like Codex and Claude Code” that the key is making agents both highly capable and user-steerable, “getting the interaction right so anyone can build cool apps.” What kind of cool apps can you make? Some real-world examples include: Startup MVPs with landing pages, onboarding flows, and working dashboards. E-commerce storefronts (perfect for designers). Survey apps connected to actual databases. Custom slide decks with editable layouts and live content. The timing is perfect, too. While AI coding startups are struggling with unsustainable pricing models, v0 implemented token-based pricing back in May 2024 and sells tokens at cost. Vercel believes “tokens are a commodity” (and should be sold like one), so their business metrics are positive (because they’re targeting the right users). Our take: v0 is interesting because it’s not just making coding easier, it’s making app building accessible to people with domain expertise but no technical skills. Got a great idea for your industry, but can’t code? Instead of waiting for developers to help you, now you can go prototype it yourself.” The real question then becomes “what will you build when ALL coding barriers disappear?” If you’re wondering where to start, may we introduce you to the Japanese concept of Ikigai? If you focus on the intersection of 1. what you love, 2. what you’re good at, 3. what the world needs, and 4. what you can be paid for (hello, Ikigai!), you’ll have endless projects worth building—regardless of whether AGI arrives tomorrow or in 20 years. Sure, what the world needs and what you can be paid for will shift over time, but if you’re building in a domain you love and leveraging what you’re good at today (and getting even better at it), you can always adapt and find new ways to create value. So how do you actually find your ikigai? Lucky for you, we’ve got the perfect prompt… (see below). FROM OUR PARTNERS It’s go-time for holiday campaigns Roku Ads Manager makes it easy to extend your Q4 campaign to performance CTV. You can: Easily launch self-serve CTV ads Repurpose your social content for TV Drive purchases directly on-screen with shoppable ads A/B test to discover your most effective offers The holidays only come once a year. Get started now with a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply. Get started today. Prompt Tip of the Day. So, you want to find your personal IKIGAI? Here’s a short prompt (inspired by this post from Garima Shah) you can try to turn ChatGPT into your own personal IKIGAI coach: “Act as an Ikigai coach by asking me detailed questions about my life. Continue asking questions until you have a complete understanding of where my passions, proficiencies, and purpose intersect in a profitable way. Your goal is to develop a deep understanding of what I love to do and how I can make money at it.” Treats to Try. *Asterisk = from our partners. Advertise in The Neuron here. *Murf AI turns your text into human-like speech in 200+ voices, so you can narrate presentations or create audiobooks without recording anything yourself. Pika now has audio-driven lip sync where you upload any audio file and it makes your AI avatar sing, rap, or act in perfect sync in under 6 seconds. NVIDIA unveiled a bunch of new Omniverse libraries and Cosmos models that let you capture real-world environments with sensors, recreate them as 3D simulations, then train

TL;DR: Better projects start with Microsoft Project Professional. Get it for just $14.97 (reg. $249.99) at TechRepublic Academy. Every company is striving to be as efficient as possible, but every company deals with waste and inefficiency. One good way to limit waste and inefficiency is by investing in the right tools to keep your company on task and under budget. One of the leading tools is Microsoft Project Professional 2021, and you can get it for just $14.97 at TechRepublic Academy. That’s a small price to pay for a program that will continue to yield results and savings through successfully managed projects over time. About Microsoft Project Professional Microsoft Project Professional has earned 4.4/5 stars from GetApp and Capterra because it’s a powerful, intuitive tool that makes project management easier. With Project Professional, you’ll have a host of pre-built templates to organize a wide variety of projects and tools to manage timelines, budgets and resources. You can run what-if scenarios to explore the potential outcomes of decisions before you make them, visually represent the schedules of multiple stakeholders and projects to ensure everyone is aligned and use automated reporting and scheduling tools to reduce inefficiencies. You can even plug in with Project Online and Project Server to manage even more data points through one central hub. You can get Microsoft Project Professional 2021 for just $14.97 (reg. $249.99). This offer is made possible by an Authorized Microsoft Partner and is not to be missed. Prices and availability are subject to change. source

OpenAI’s CEO Sam Altman. Image: Creative Commons OpenAI has rolled out two new AI models, o3 and o4‑mini, that can literally “think with images,” marking a big step forward in how machines understand pictures. These models, announced in an OpenAI press release, can reason about images the same way they do about text — cropping, zooming, and rotating photos as part of their internal thought process. At the heart of this update is the ability to blend visual and verbal reasoning. “OpenAI o3 and o4‑mini represent a significant breakthrough in visual perception by reasoning with images in their chain of thought,” the company said in its press release. Unlike past versions, these models don’t rely on separate vision systems — instead, they natively mix image tools and text tools for richer, more accurate answers. How does ‘thinking with images’ work? The models can crop, zoom, rotate, or flip an image as part of their thinking process, just like humans would. They’re not just recognizing what’s in a photo but working with it to draw conclusions. The company notes that “ChatGPT’s enhanced visual intelligence helps you solve tougher problems by analyzing images more thoroughly, accurately, and reliably than ever before.” This means if you upload a photo of a handwritten math problem, a blurry sign, or a complicated chart, the model can not only understand it, but also break it down step by step — possibly even better than before. More must-read AI coverage Outperforms previous models in key benchmarks These new abilities aren’t just impressive in theory; OpenAI says both models outperform their predecessors regarding top academic and AI benchmarks. “Our models set new state-of-the-art performance in STEM question-answering (MMMU, MathVista), chart reading and reasoning (CharXiv), perception primitives (VLMs are Blind), and visual search (V*),” the company noted in a statement. “On V*, our visual reasoning approach achieves 95.7% accuracy, largely solving the benchmark.” But the models aren’t perfect. OpenAI admits the models can sometimes overthink, leading to prolonged and unnecessary image manipulations. There are also cases where the AI might misinterpret what it sees, despite correctly using tools to analyze the image. The company also warned of reliability issues when trying the same task multiple times. Who can use OpenAI o3 and o4-mini? As of April 16, both o3 and o4-mini are available to ChatGPT Plus, Pro, and Team users; they replace older models like o1 and o3-mini. Enterprise and education users will get access next week, and free users can try o4-mini through a new “Think” feature. source

Sundar Pichai at the Google I/O 2017 Keynote. Image: Steven Zimmerman/Creative Commons Alphabet, Google’s parent company, has claimed the top spot on LinkedIn’s 2025 list of the 50 best large US employers for career growth. The rankings, which relied entirely on LinkedIn data, evaluated companies on key indicators of career progression, such as employee promotions, skill development, and internal mobility. LinkedIn noted that “Google goes the extra mile to retain and grow its talent pipeline.” The search engine giant also “invests heavily in leadership and development … and coaching for managers across all levels,” as well as offering education reimbursement so employees can continue their degrees without a financial burden. Top 10 large companies for career growth in the US Rounding out the top 10 companies on LinkedIn’s list are: Amazon: Its AWS Grow Our Own Talent program helps employees with nontraditional work experience land in-demand data center roles. The program is part of the company’s $1.2 billion investment to upskill its workforce in emerging technical areas. Wells Fargo: The financial services firm’s Career Development Program provides professional training and development courses for employees in roles requiring less than five years of experience. Northrop Grumman: The defense and space manufacturer helps employees advance their technical skills while on the job. Its initiatives include a program that exposes early-career employees to foundational technical competencies across the company. PwC: The professional services firm has invested $1 billion in AI upskilling, resulting in over 75% of employees using GenAI tools daily. Capital One: The financial services firm offers four employee-led internal colleges where employees can deepen their skills in areas such as design thinking and leadership. AT&T: The telecommunications giant has implemented several tools and programs to help employees work more efficiently, including its AI productivity assistant, Ask AT&T. JPMorganChase: The financial services firm has pivoted toward skills-based employment and expanded its talent pool. Around 70% of roles for experienced hires or candidates with experience do not require a college degree. EY: The professional services firm offers job candidates “Space for the Curious” experiences that guide them through the application process. This includes features such as an AI-enabled interview coach that helps them practice interviewing and VR headsets with digital twins of EY offices that illustrate what it’s like to work there. Walmart: The retail giant’s Live Better U program creates promotion pathways for associates to advance to higher-level jobs through opportunities to learn key skills and earn degrees at no cost. Feedback about LinkedIn’s rankings Reaction to LinkedIn’s list was mixed. In the comments section of the LinkedIn post about this research, one HR leader called it “Informative,” noting that “many of these companies have been part of large layoffs. So while there’s growth at a big company, I think [it’s] important to look to other mid-sized companies not on this list.” A talent acquisition leader commented that “Alphabet has been laying off. Amazon did [return to office] RTO and it’s gone horribly and most employees are unhappy. JPMorganChase? With a CEO screaming at employees? EY just laid off their entire U.S. recruiting workforce. Walmart requires relocation to Tornado Alley.” Eight free LinkedIn Learning courses LinkedIn is offering these learning courses free for all members until May 7. The courses include career topics such as leveraging your transferable skills and creating your own internal mobility plan. source

Image: iStock/BeeBright A bipartisan report, recently issued by the US Select Committee on the Chinese Communist Party (CCP), accuses DeepSeek of a series of subversive, illegal, and immoral practices. Moreover, the tech giant NVIDIA is also catching the ire of US government officials for supplying DeepSeek with the chips needed to create the AI models. Investigating DeepSeek The report, titled “DeepSeek Unmasked: Exposing the CCP’s Latest Tool for Spying, Stealing, and Subverting U.S. Export Control Restrictions,” was published in April 2025. It levies numerous accusations against DeepSeek, including: Actively suppressing more than 85% of responses that are related to human rights, democracy, Taiwan, or Hong Kong. Being owned and operated by a company with a direct link to the CCP. Currently funneling data on American users to the CCP. Maintaining infrastructure that is linked to Chinese companies utilizing mass surveillance, data harvesting, and censorship. In addition to the aforementioned accusations, the report also links dozens of DeepSeek employees and researchers to the People’s Liberation Army (PLA), the military wing of the CCP. “This report makes it clear: DeepSeek isn’t just another AI app — it’s a weapon in the Chinese Communist Party’s arsenal, designed to spy on Americans, steal our technology, and subvert U.S. law,” said Chairman John Moolenaar (R-MI). DeepSeek has already been banned in some countries including Australia, India, Italy, South Korea, and Taiwan due to potential security concerns. Its usage has also been banned by the US Congress, NASA, and among government entities within the state of Texas. Exploring NVIDIA’s role DeepSeek wasn’t the only company mentioned in the report; the tech giant NVIDIA also faces accusations, including: Supplying DeepSeek with more than 60,000 chips for the development of its platform, possibly circumventing export regulations. Developing a modified chip as a workaround to a loophole in regulations. NVIDIA spokesman John Rizzo previously issued a statement saying, in part: “We insist that our partners comply with all applicable laws, and if we receive any information to the contrary, act accordingly,” as reported by The New York Times. But NVIDIA’s statement isn’t enough for the congressional committee, which is accusing countries like Singapore and Malaysia of purchasing chips and illegally exporting them to China. As such, the committee is asking NVIDIA to provide specific details concerning every customer account from no less than 11 Asian countries. The committee is expecting NVIDIA’s response within two weeks. Waiting for a resolution The US Select Committee on the CCP first began investigating DeepSeek and its rapid pace of technological advancement in February 2025. While their recent report includes some very serious allegations against both DeepSeek and NVIDIA, it’s important to remember that all parties remain innocent until proven guilty. source

Microsoft AI CEO Mustafa Suleyman. Image: Microsoft At a time when tech giants are racing to build the most powerful AI models on the planet, Microsoft is choosing a different lane — one that’s a little slower but potentially smarter. Mustafa Suleyman, Microsoft’s CEO of AI, says the company isn’t in a rush to lead the frontier of AI development; instead, it’s intentionally choosing to lag a few months behind leaders like OpenAI. That might sound like a setback, but Suleyman says it’s a strategic move that could give Microsoft an edge. “It’s cheaper to give a specific answer once you’ve waited for the first three or six months for the frontier to go first. We call that off-frontier,” Suleyman said in a recent interview with CNBC. “That’s actually our strategy — to really play a very tight second, given the capital-intensiveness of these models.” A new chapter in Microsoft’s AI playbook The comments come as Microsoft celebrates its 50th anniversary, during which it announced several new features to its AI assistant Copilot. These include Copilot Vision, Deep Research, and a new “Memory” feature that remembers details about users to personalize responses — a capability previously introduced by OpenAI’s ChatGPT. Microsoft’s AI systems, including Copilot, are heavily powered by OpenAI’s models. The tech giant has invested nearly $14 billion in the San Francisco startup, making it a key partner in Microsoft’s AI journey. But Suleyman also made it clear that Microsoft is gradually working toward self-sufficiency. “Long-term, it’s absolutely mission-critical that we are able to do AI self-sufficiently at Microsoft,” he said. “Until 2030 at least, we are deeply partnered with OpenAI, who have had an enormously successful relationship with us.” That long-term view reflects Suleyman’s broader strategy. While the company still relies on OpenAI for large models like GPT-4, it’s also developing its own smaller, open-source models under the Phi project. These models are designed to run on personal computers rather than massive GPU clusters, making them more accessible and cost-effective to deploy. DOWNLOAD This IT Leader’s Guide to Generative AI From TechRepublic Premium Cracks in the Microsoft-OpenAI partnership? The close ties between Microsoft and OpenAI have started to show some strain. Last year, Microsoft quietly added OpenAI to its list of competitors. And in January, OpenAI surprised many by teaming up with Oracle on its $500 billion Stargate project — a move that ended Microsoft’s exclusive status as OpenAI’s cloud provider. Further escalating tensions, OpenAI’s latest GPT models have reportedly raised concerns within Microsoft over performance and cost. Industry insiders believe Microsoft has already begun developing its own frontier models, and some observers, including Salesforce CEO Marc Benioff, believe Microsoft is already moving to stand on its own two feet. More must-read AI coverage The bigger picture Microsoft may not be chasing the AI crown, but it’s still playing a long game. With a powerful AI team, deep investments in infrastructure, and a broad customer base, it doesn’t need to lead the race to win it. By staying just behind the bleeding edge, the company hopes to avoid costly mistakes and offer more stable, reliable solutions to real-world problems. “We have an incredibly strong AI team, huge amounts of compute, and it’s very important to us that, you know, maybe we don’t develop the absolute frontier, the best model in the world first,” he said. “That’s very, very expensive to do and unnecessary to cause that duplication.” source

HubSpot’s fast facts Our star rating: 4/5 Pricing: Robust free version for up to 5 users. Paid tiers start at $15 per seat per month when billed annually. Key features: Lead prospecting and generation. Email templates and tracking. Document and payment management. Business forecasting. Powerful AI tools. 1,500+ integrations. Supports multiple industries. 1,500+ integrations. HubSpot is a popular customer relationship management platform with robust free and paid features that help streamline entire sales and marketing strategies. Its custom pipelines and activity tracking make it easily scalable to fit any organization’s sales and customer management needs. HubSpot remains a top-rated CRM for customer support and integration capabilities However, if you need software with more advanced technology and reporting features, other options are available. Pricing Free CRM: Free for up to 5 users with contact management, quotes, live chat, and more. Sales Hub Starter: $15 per seat per month, billed annually, or $20 when billed monthly. Starter Customer Platform: $15 per seat per month, billed annually, or $20 when billed monthly. This plan includes access to Marketing and Service Start Hub. Sales Hub Professional: $90 per seat per month, billed annually, or $100 when billed monthly, and a one-time $1,500 onboarding fee. Sales Hub Enterprise: $150 per seat per month, with an annual commitment. A one-time $3,500 onboarding fee applies. SEE: For more details about HubSpot’s offerings and pricing, check out HubSpot’s Product and Services Catalog. Key features of HubSpot Call tracking HubSpot’s call tracking system allows users to place and record client calls directly from the browser platform. This system automatically logs and tracks each call to the correct customer profile. It also prioritizes outbound calls, enabling sales reps to start daily calls sooner and begin selling faster. Hubspot call tracking software. Image: HubSpot Sales playbooks HubSpot’s sales playbook software equips sales reps with interactive sales plays, scripts, guides, and more. Users can customize the playbooks to align with their organization and selling style. The software also provides real sales data, drawing from structured notes and multiple-choice questions and answers. HubSpot interactive playbook feature. Image: HubSpot Conversation intelligence HubSpot’s conversation intelligence tracks sales rep performance on customer calls and provides coaching tips based on data-driven insights. Users don’t need to switch dialing software or waste time training on a new calling service. This feature works with HubSpot’s calling tool, Zoom, JustCall, Kixie, and more. Client engagement tracking and coaching from conversation intelligence feature. Image: HubSpot Sales analytics and reporting Users can choose from dozens of pre-made report styles or customize their own to view business analytics. The reports display pipeline revenue, individual sales progress, or target attainment. HubSpot’s analytic feature allows admins to monitor sales activities and create unique coaching insights. This tool benefits businesses of all sizes, including small businesses that need to identify market and customer patterns. Sample Hubspot sales dashboard. Image: HubSpot PREMIUM: Explore a full CRM feature comparison.     HubSpot pros and cons Pros Cons Free demo. 1,500+ app integrations. 24/7 customer support. Higher support tiers are costly. No live support for free users. Limited customization options. More CRM coverage Alternatives to HubSpot HubSpot’s dashboard and pipeline customization have limitations, and its higher support tiers and add-ons can be costly. Some alternatives, like Pipedrive, Zoho CRM, and Creatio, offer similar core features in their free tiers with additional technology features.. When researching the right CRM for your organization, first consider your budget, whether the software specializes in your industry, and it scales to meet the size and needs of your business. For our full review of alternatives, see our HubSpot Alternatives page. DOWNLOAD: Compare the top CRMs for small businesses. Software HubSpot CRM Pipedrive Zoho CRM Creatio Custom dashboards Limited Yes Limited Yes Native integrations Yes Yes Yes Yes Third-party integrations Yes Yes Yes Yes AI-powered features Yes Yes Yes Limited Free trial No 14 days 15 days 14 days Starting price* Free starting price $14 per user per month Free starting price $15 per user per month * Pricing details are based on annual subscription rates. Different rates may apply for per-month billing. Pipedrive Pipedrive is a user-friendly CRM known for its visual pipelines that simplify sales processes. It is a great option if you want sales-focused software without any additional marketing tools. Although Pipedrive doesn’t offer a free-for-life version, its starting price remains competitive. For a closer look, check out my detailed comparison of Pipedrive and HubSpot. SEE: For more information on Pipedrive, check out my Pipedrive review. Zoho CRM Zoho CRM offers sales and marketing features, similar to HubSpot but places a greater emphasis on social media marketing and online lead generation. The platform is easy to navigate and offers more advanced data and forecasting features than HubSpot. I have a full comparison breakdown of Zoho CRM and HubSpot that I recommend checking out. SEE: For more details, explore this in-depth review of Zoho CRM. Creatio Creatio is a no-code platform that supports marketing, sales, and service teams. It offers end-to-end sales management tools to completely automate workflows. While you can easily customize Creatio without any coding expertise, it doesn’t provide a free version of its platform. If you want a free CRM with even more AI functionality, I recommend HubSpot over Creatio. SEE: Want to know more? Read my review of Creatio. Review methodology To review HubSpot and its offerings, we used our inhouse rubric of criteria that we consider most important for assessing general CRM software. Next, we compared HubSpot against industry standards in this scoring tool. We relied on HubSpot’s own online resources and incorporated real user feedback and ratings. Below is the breakdown of the criteria used to score HubSpot: Cost: Weighted 25% of the total score. Core features: Weighted 25% of the total score. Customizations: Weighted 15% of the total score. Integrations: Weighted 15% of the total score. Ease of use: Weighted 10% of the total score. Customer support: Weighted 10% of the total score. FAQs Is HubSpot a good CRM? Yes, HubSpot is a well-rounded and

Microsoft researchers claim to have developed the first 1-bit large language model with 2 billion parameters. The model, BitNet b1.58 2B4T, can run on commercial CPUs such as Apple’s M2. “Trained on a corpus of 4 trillion tokens, this model demonstrates how native 1-bit LLMs can achieve performance comparable to leading open-weight, full-precision models of similar size, while offering substantial advantages in computational efficiency (memory, energy, latency),” Microsoft wrote in the project’s Hugging Face depository. What makes a bitnet model different? Bitnets, or 1-bit LLMs, are compressed versions of large language models. The original 2-billion parameter scale model trained on a corpus of 4 billion tokens was shrunken down into a version with drastically reduced memory requirements. All weights are expressed as one of three values: -1, 0, and 1. Other LLMs might use 32-bit or 16-bit floating-point formats. SEE: Threat actors can inject malicious packages into AI models that resurface during “vibe coding.” In the research paper, which was posted on Arxiv as a work in progress, the researchers detail how they created the bitnet. Other groups have created bitnets before, but, the researchers say, most of their efforts are either post-training quantization (PTQ) methods applied to pre-trained full-precision models or native 1-bit models trained from scratch that were developed at a smaller scale in the first place. BitNet b1.58 2B4T is a native 1-bit LLM trained at scale; it only takes up 400MB, compared to other “small models” that can reach up to 4.8 GB. BitNet b1.58 2B4T model performance, purpose, and limitations Performance compared to other AI models BitNet b1.58 2B4T outperforms other 1-bit models, according to Microsoft. BitNet b1.58 2B4T has a maximum sequence length of 4096 tokens; Microsoft claims it outperforms small models like Meta’s Llama 3.2 1B or Google’s Gemma 3 1B. Researchers’ goal for this bitnet Microsoft’s goal is to make LLMs accessible to more people by creating versions that run on edge devices, in resource-constrained environments, or in real-time applications. However, BitNet b1.58 2B4T still isn’t simple to run; it requires hardware compatible with Microsoft’s bitnet.cpp framework. Running it on a standard transformers library won’t produce any of the benefits in terms of speed, latency, or energy consumption. BitNet b1.58 2B4T doesn’t run on GPUs, as the majority of AI models do. More must-read AI coverage What’s next? Microsoft’s researchers plan to explore training larger, native 1-bit models (7B, 13B parameters and more).They note that most of today’s AI infrastructure lacks suitable hardware for 1-bit models, so they plan to explore “co-designing future hardware accelerators” specifically designed for compressed AI. The researchers also aim to: Increase context length. Improve performance on long-context chain-of-thought reasoning tasks. Add support for multiple languages other than English. Integrate 1-bit models into multimodal architectures. Better understand the theory behind why 1-bit training at scale produced efficiencies. source

Copilot Studio’s new computer use feature for AI automation. Image: Microsoft Microsoft announced this week that a new computer use feature is coming to Copilot Studio. Computer use will allow users to build AI-driven agents capable of interacting with their desktop and browser applications, as long as the third-party system it uses has a graphical user interface. User-friendly automation for even non-developers Currently, Microsoft Copilot Studio users can create agents to automate tasks in third-party applications but only within specific services. With the computer use capability, agents can use third-party applications and websites just as a person would, with the ability to click buttons, type into fields on-screen, and select from menus. This opens up a world of possibilities for automating business tasks. Copilot Studio’s new computer use feature aims to make the system’s robotic process automation more intuitive and accessible for the average user. With the ability to carry out natural language processing, provide full visibility over the computer’s use history, and automatically respond to updates, the automation is user-friendly, even for those who aren’t professional RPA developers. Microsoft’s announcement about Copilot Studio and computer use builds upon last month’s release of new AI features, which included deep reasoning capabilities for agents, model context protocol support, and the availability of agent flows within the system. More must-read AI coverage Advantages of Microsoft’s newest update to Copilot Studio Computer use is especially helpful for enhancing the capabilities of AI agents. The addition of computer use agents in Copilot Studio brings new functionality to the system and offers more benefits to the user. Functionality without an API connection: Agents can perform tasks within third-party systems without requiring an API connection. Automatic adaptation to changes in apps and websites: Computer use automatically adapts to system updates and processes adjustments in real time for an interruption-free experience. Built on Copilot Studio’s governance frameworks and security measures: Computer use complies with organizational and industry standards, and sensitive enterprise data is safeguarded within Microsoft Cloud boundaries, keeping it secure. Microsoft-hosted: Organizations don’t need to manage their servers, as computer use runs on Microsoft-hosted infrastructure. More business functionality: Microsoft’s possible use cases for the new feature include creating agents to automate data entry, process invoices, or perform market research. Test Copilot Studio’s computer use feature This new capability will be coming to Copilot Studio through an early access web preview. Users who are interested in testing and evaluating the computer use feature should fill out this form to let Microsoft know they would like to participate. source