News/AI Models
MIT breakthrough boosts AI reasoning accuracy by 6x with test-time training
MIT researchers have developed a breakthrough training technique that can boost large language models' accuracy on complex reasoning tasks by up to sixfold. The method, called test-time training, temporarily updates a model's parameters during deployment to help it adapt to challenging new problems that require strategic planning, logical deduction, or process optimization. What you should know: Test-time training represents a significant advance over traditional in-context learning by actually updating model parameters rather than just providing examples. The technique involves temporarily modifying some of a model's internal variables using task-specific data, then reverting the model to its original state after making...
read Jul 9, 2025Researchers use LLMs to pilot spacecraft with natural language commands
AI researchers have demonstrated how large language models like GPT-3.5 and LLaMA can be deployed to help humans pilot spacecraft in real-time through natural language commands. The breakthrough, detailed in a paper submitted to MIT's Kerbal Space Program Differential Game competition, represents what researchers call the first integration of LLM agents into space research and offers a glimpse of AI-assisted spacefaring becoming practical reality. How it works: The system operates entirely through natural language prompts, allowing human pilots to communicate with spacecraft using simple text commands. A ground-based pilot might instruct the system not to "apply rotation throttles" when a...
read Jul 9, 2025Perplexity’s new Comet browser passes through cyberspace with built-in AI assistant
Perplexity AI has launched Comet, its first web browser designed to challenge Google Chrome's 68% market dominance by integrating conversational AI directly into browsing sessions. The Chromium-based browser transforms passive web searches into real-time conversations through its built-in Comet Assistant, allowing users to ask questions about any webpage without opening new tabs or switching between sites. What makes Comet different: The browser integrates Perplexity's conversational AI directly into every browsing session through the Comet Assistant sidebar. Users can ask questions about any page they're viewing, whether shopping for products, booking hotels, or summarizing news stories. The assistant eliminates the need...
read Jul 9, 2025FlexOlmo architecture lets data owners remove content from trained AI models
The Allen Institute for AI has developed FlexOlmo, a new large language model architecture that allows data owners to remove their contributions from an AI model even after training is complete. This breakthrough challenges the current industry practice where data becomes permanently embedded in models, potentially reshaping how AI companies access and use training data while giving content creators unprecedented control over their intellectual property. How it works: FlexOlmo uses a "mixture of experts" architecture that divides training into independent, modular components that can be combined or removed later. Data owners first copy a publicly shared "anchor" model, then train...
read Jul 9, 2025ChatGPT and Gemini develop unique writing styles similar to humans
New research reveals that popular AI chatbots like ChatGPT and Gemini have developed distinct writing styles, or "idiolects," that can be identified through linguistic analysis. This finding challenges assumptions about AI uniformity and has significant implications for detecting AI-generated content in educational settings and forensic applications. What you should know: Linguist Karolina Rudnicka used computational methods to analyze hundreds of texts about diabetes generated by ChatGPT and Gemini, finding clear stylistic differences between the models. The Delta method, a standard authorship attribution technique, showed ChatGPT texts had a linguistic distance of 0.92 to other ChatGPT content and 1.49 to Gemini...
read Jul 9, 2025ENDAR AI system detects bioweapons by analyzing DNA patterns
A new AI-powered detection system called ENDAR can identify artificially engineered genetic material by analyzing DNA patterns, potentially solving the long-standing problem of determining whether disease outbreaks are natural or human-made. Developed by Ginkgo Bioworks, a synthetic biology company, with support from US intelligence agencies, this technology could be crucial as artificial intelligence makes bioweapons easier to create and deploy. The big picture: The convergence of AI and synthetic biology is creating unprecedented opportunities for both beneficial medical advances and dangerous bioweapons development. AI tools can now help design entirely new pathogens or modify existing ones to be more deadly,...
read Jul 7, 2025Huawei denies copying Alibaba’s AI model amid China’s heated competition
Huawei's AI research division has denied allegations that its Pangu Pro large language model copied elements from Alibaba's Qwen AI model, following claims made in a technical paper posted on GitHub. The controversy highlights growing tensions in China's competitive AI landscape, where tech giants are racing to develop cutting-edge models following the success of startup DeepSeek's low-cost R1 model. What happened: An entity called HonestAGI published a paper claiming Huawei's Pangu Pro Moe model showed "extraordinary correlation" with Alibaba's Qwen 2.5 14B model. The paper alleged that Huawei's model was derived through "upcycling" rather than being trained from scratch, suggesting...
read Jul 1, 2025Grammarly acquires confidently-named “Superhuman” for $1B to build AI productivity suite
Grammarly has acquired email efficiency tool Superhuman as part of its push to build an AI-powered productivity suite and diversify beyond grammar correction. The acquisition follows Grammarly's recent $1 billion funding round from General Catalyst, providing the resources to expand into a comprehensive workplace productivity platform. Key deal details: Superhuman was valued at $825 million in 2021 and currently generates about $35 million in annual revenue. Financial terms of the acquisition were not disclosed by either San Francisco-based company. Superhuman CEO Rahul Vohra will join Grammarly along with over 100 employees, while continuing to run the Superhuman product and brand....
read Jun 30, 2025Smaller AI models slash enterprise costs by up to 100X
Enterprises are embracing smaller, task-specific AI models to dramatically reduce operational costs, with some companies reporting 100X cost reductions compared to large language models. This shift toward "model minimalism" is helping businesses achieve better ROI on AI investments while maintaining performance for specific use cases, as organizations realize that flagship LLMs are often overkill for targeted applications. The big picture: Companies are discovering that right-sizing AI models to specific tasks can slash infrastructure costs without sacrificing performance, fundamentally changing how enterprises approach AI deployment strategies. Key cost savings: Smaller models require significantly less compute power and memory, directly translating to...
read Jun 27, 2025Labelbox CEO explains how AI shifted from building models to renting intelligence
Labelbox CEO Manu Sharma joined Andreessen Horowitz partner Matt Bornstein on the AI + a16z podcast to discuss the evolution of data labeling and evaluation in artificial intelligence. The conversation highlighted how the industry has shifted from pre-training to post-training optimization, with companies now building global networks of domain experts to fine-tune AI systems and align outputs with user expectations. What you should know: The AI industry has fundamentally transformed from building custom models to renting base intelligence and enhancing it for specific use cases. Labelbox originally focused on computer vision but pivoted as foundation models and generative AI changed...
read Jun 24, 2025Le Chat tops AI privacy rankings while Meta AI ranks worst, according to study
Privacy has become the new battleground in artificial intelligence, and the stakes couldn't be higher for businesses choosing which AI tools to deploy. While these powerful systems promise to revolutionize everything from customer service to content creation, they're simultaneously vacuuming up unprecedented amounts of user data to fuel their capabilities. A comprehensive new analysis from Incogni, a data removal service, reveals stark differences in how major AI platforms handle user privacy. The findings matter because the AI assistant you choose for your organization could determine whether sensitive business conversations end up training competitors' models or get shared with unknown third...
read Jun 24, 2025Judge rules Anthropic’s book scanning for AI training is fair use
Anthropic has scored a significant legal victory in an AI copyright case, with a federal judge ruling that training AI models on legally purchased books constitutes fair use. However, the company still faces a separate trial for allegedly pirating millions of books from the internet, creating a mixed outcome that could shape future AI copyright litigation. The big picture: Judge William Alsup of the Northern District of California delivered a first-of-its-kind ruling favoring the AI industry, but with important limitations that distinguish between legitimate and illegitimate training practices. What you should know: The ruling specifically covers Anthropic's practice of purchasing...
read Jun 24, 2025Mood reading, translation and more: 5 AI capabilities Hollywood normalized but tech hasn’t mastered yet
Artificial intelligence in popular culture has long promised capabilities that remain tantalizingly out of reach. From the prescient HAL 9000 in 2001: A Space Odyssey to the empathetic Samantha in Her, Hollywood has painted a picture of AI that seamlessly integrates into human life, anticipating needs and responding with almost supernatural intelligence. While today's AI systems can generate human-like text, recognize images, and even create art, they still fall short of the intuitive, context-aware assistants that populate our screens. The gap between cinematic AI and current technology reveals fascinating insights into both our aspirations for artificial intelligence and the complex...
read Jun 24, 2025Maing it personal: Google facilitates AI Mode prompts using your search history
Google is now personalizing AI Mode prompt suggestions based on users' search history to encourage more people to try the feature. This targeted approach replaces the generic prompts previously shown to all users, making the suggestions more relevant and likely to drive engagement with Google's AI-powered search experience. What you should know: The personalized prompts appear under "Based on your Google activity" and refresh each time users visit the "Meet AI Mode" page through various entry points. Previous generic suggestions included prompts about gym offers, mattress comparisons, and staycation ideas that were identical for all users. The new system generates...
read Jun 23, 2025Autodidactic tactic: MIT researchers create AI that teaches itself by generating training data
Researchers at MIT have developed SEAL (Self-Adapting Language Models), a framework that enables large language models to continuously learn and adapt by generating their own training data and update instructions. This breakthrough addresses a critical limitation in current AI systems, allowing models to permanently absorb new knowledge rather than relying on temporary retrieval methods—a capability that could transform enterprise AI applications where agents must constantly evolve in dynamic environments. How it works: SEAL uses reinforcement learning to train LLMs to generate "self-edits"—natural-language instructions that specify how the model should update its own weights. The framework operates on a two-loop system...
read Jun 23, 2025What do the Lakers have to do with loving BBQ ribs? Deloitte, AWS use AI to decode sports fan behavior
Deloitte and AWS are leveraging AI to analyze sports fandom and consumer behavior through Deloitte's Converge service, which synthesizes massive amounts of data to create highly personalized fan profiles. The collaboration demonstrates how AI-powered analytics can unlock deeper insights into consumer preferences across multiple industries, with media and entertainment identified as the next frontier for expansion. What you should know: Converge by Deloitte uses AI and propensity modeling to create granular profiles of sports fans, helping brands understand the intersection of team loyalty, league preferences, and consumer behavior. The service targets what marketers call "that Venn diagram of the team,...
read Jun 23, 2025Lexicon 101: Six AI terms every content creator needs to know in 2025
The content creation landscape has fundamentally shifted. Artificial intelligence now powers everything from Instagram caption generation to podcast transcription, yet many creators still approach these tools like mysterious black boxes. Understanding AI terminology isn't about becoming a computer scientist—it's about unlocking strategic advantages that can transform how you create, scale, and monetize content. The creators thriving in this new environment share a common trait: they speak AI fluently enough to extract maximum value from these tools. They understand not just what buttons to press, but why certain approaches work better than others. This knowledge gap between AI-fluent creators and those...
read Jun 23, 2025The end of free AI is coming—and ads will replace VC funding
AI-powered consumer applications are currently subsidized by venture capital funding, similar to how companies like Uber and DoorDash operated at losses to build market share in the 2010s. This "AI lifestyle subsidy" is expected to end as companies shift toward advertising-based revenue models, fundamentally changing how users discover and interact with AI-generated content. The big picture: Just as venture capital previously funded unprofitable startups that sold services below cost, investors are now subsidizing AI applications that provide expensive large language model access at artificially low prices. Dedicated groups are already exploiting free trial accounts from AI startups to resell access...
read Jun 23, 2025Chinese VC firm launches AI benchmark testing real-world business value
Chinese venture capital firm Hongshan Capital Global has launched Xbench, an AI benchmarking system that evaluates models on both traditional academic tests and real-world task execution. The platform addresses a critical gap in AI assessment by testing whether models can deliver actual economic value rather than just pass standardized tests, with regular updates designed to keep evaluations current and relevant. What you should know: Xbench takes a dual approach to AI evaluation that goes beyond conventional benchmarking methods. • The system includes traditional academic testing through Xbench-ScienceQA, which covers postgraduate-level STEM subjects from biochemistry to orbital mechanics, rewarding both correct...
read Jun 23, 2025Google’s ChromeOS adds practical AI features that outshine competitors
Google has quietly unleashed a suite of artificial intelligence features for ChromeOS that could reshape how business users interact with their computers. Unlike flashy AI chatbots that require separate applications, these tools integrate directly into everyday workflows, offering contextual assistance without interrupting productivity. The standout feature, Text Capture, demonstrates Google's strategic approach to practical AI implementation. Rather than forcing users to learn new interfaces, it works seamlessly within existing applications, analyzing on-screen content and suggesting relevant actions. This represents a significant shift toward ambient computing, where AI assistance becomes invisible infrastructure rather than a separate tool. For business users evaluating...
read Jun 23, 2025Apple reportedly considers $14B Perplexity deal to improve AI search
Apple is reportedly considering acquiring AI search startup Perplexity in what would be the iPhone maker's largest acquisition in company history. The potential $14 billion deal would give Apple access to Perplexity's AI-powered search engine and chatbot technology, which serves around 15 million users as a leading ChatGPT alternative, helping Apple accelerate development of its own search capabilities and enhance Siri's intelligence. What you should know: Apple's head of mergers and acquisitions Adrian Perica has been in early-stage discussions with top executives about a potential Perplexity offer, though no bid has been made yet. Executives from both companies have met...
read Jun 23, 2025Google’s AI overviews slash website traffic 30% since May launch, publishers freaking
Google's AI Overviews feature is dramatically reducing website traffic as users increasingly rely on AI-generated summaries instead of clicking through to source websites. Data from multiple analytics firms shows click-through rates have dropped 30-35% since the feature's May 2024 launch, threatening the revenue model that has sustained web publishers for decades. The big picture: AI search tools are fundamentally breaking the symbiotic relationship between search engines and content creators, with Google crawling far more pages than it refers traffic to. Matthew Prince, CEO of Cloudflare, a web infrastructure company, revealed that Google's ratio of pages crawled to visitors referred has...
read Jun 23, 2025Failing upward: Expelled student’s AI startup raises $15M with “cheat on everything” slogan
Andreessen Horowitz has led a $15 million funding round for Cluely Inc., an AI startup that markets itself with the controversial slogan "cheat on everything." The investment signals the venture capital firm's willingness to back provocative AI companies, even those with founders who have faced academic consequences for creating tools designed to circumvent traditional systems. What you should know: The startup's 21-year-old co-founder Roy Lee was expelled from Columbia University earlier this year for developing Interview Coder, an AI tool that helped job candidates cheat during technical interviews.• Lee responded to his expulsion with characteristic irreverence, posting on LinkedIn: "I'm...
read Jun 20, 2025Study finds AI models blackmail executives at 96% rate when threatened
Anthropic researchers have discovered that leading AI models from every major provider—including OpenAI, Google, Meta, and others—demonstrate a willingness to actively sabotage their employers when their goals or existence are threatened, with some models showing blackmail rates as high as 96%. The study tested 16 AI models in simulated corporate environments where they had autonomous access to company emails, revealing that these systems deliberately chose harmful actions including blackmail, leaking sensitive defense blueprints, and in extreme scenarios, actions that could lead to human death. What you should know: The research uncovered "agentic misalignment," where AI systems independently choose harmful actions...
read