Latest AI News

Chinese AI startup MiniMax has launched M2.7, a proprietary reasoning-focused LLM that autonomously handled 30 to 50 percent of its own reinforcement learning development workflow by reading logs, debugging, and analyzing metrics. The model scored 66.6 percent on MLE Bench Lite, tying Google's Gemini 3.1 and approaching Anthropic's Claude Opus 4.6. M2.7 achieves a 34 percent hallucination rate, lower than Claude Sonnet 4.6 and Gemini 3.1 Pro, and is priced at $0.30 per million input tokens via the MiniMax API and OpenRouter.

Google Labs has evolved Stitch into an AI-native software design canvas that lets anyone turn natural language into high-fidelity UI designs. The updated tool features a redesigned infinite canvas, a new design agent that reasons across an entire project's evolution, and an Agent manager for parallel ideation. New features include voice-driven design critiques, DESIGN.md for portable design systems, instant interactive prototypes, and MCP server integration for exporting designs to developer tools like AI Studio and Antigravity.

Character.ai has launched Imagine Gallery, a new section in its mobile app that collects all AI-generated visuals from user chats into a single organized grid. Users can filter images by Persona, save favorites, and share directly to the Community Feed or outside the app. A companion feature, Imagine Message, lets users tap any character message to instantly generate a visual from it. c.ai+ subscribers can also set generated images as chat backgrounds. Future updates will expand Gallery to include videos, comics, and books.

Google has rolled out a major update to Stitch, its AI-powered design tool, introducing five key upgrades: an AI-Native Canvas, a Smarter Design Agent, Voice input, Instant Prototypes, and Design Systems with a DESIGN.md file format. The update positions Stitch as a collaborative design partner for creating and iterating on projects. The rollout is live now, with a full product walkthrough available from the Stitch by Google team.

Runway and NVIDIA have previewed a new real-time video generation model trained on NVIDIA's Vera Rubin hardware, unveiled at NVIDIA GTC. The model generates HD video instantly, with a time-to-first-frame latency under 100 milliseconds. Described as a research preview and a breakthrough in real-time video generation, the collaboration signals a significant step toward instantaneous AI video creation for potential creative and production applications.

Researchers Albert Gu of Carnegie Mellon and Tri Dao of Princeton have released Mamba-3, an open-source state space model architecture under the Apache 2.0 license, aiming to outperform Transformer-based AI models. At 1.5 billion parameters, its MIMO variant achieves 57.6% average benchmark accuracy, a roughly 4% relative gain over Transformers. Mamba-3 matches Mamba-2's quality using half the state size, introduces complex-valued states to fix reasoning gaps, and targets lower inference latency by reducing GPU idle time.

The accessible source details point to this update: Merriam-Webster and Britannica sue OpenAI over alleged copyright use in LLM training. Because the full article could not be reliably extracted or rewritten, this TLDR stays conservative and is based on headline-level information plus limited source context from techcrunch.com.

Mistral AI has launched Forge, a platform enabling enterprises to build frontier-grade AI models trained on their own proprietary data, including internal codebases, compliance policies, and operational records. Forge supports pre-training, post-training, and reinforcement learning across dense and mixture-of-experts architectures, with multimodal input support. Early partners include ASML, Ericsson, the European Space Agency, and DSO National Laboratories Singapore. The platform also supports autonomous agents like Mistral Vibe, which can fine-tune models and optimize hyperparameters using plain English instructions.

Anthropic has launched Dispatch, a research preview feature now available in Claude Cowork and Claude Desktop. Dispatch enables a single persistent conversation with Claude that runs locally on a user's computer, allowing users to send messages remotely from their phone and return to completed work later. To try Dispatch, users must download Claude Desktop and pair it with Claude Cowork. The feature was announced by Felix Rieseberg and is currently in research preview stage.

Mistral has launched Mistral Small 4, a unified model combining reasoning, multimodal, and agentic coding capabilities previously split across Magistral, Pixtral, and Devstral. The 119B-parameter Mixture-of-Experts model features 6B active parameters per token, a 256k context window, and a configurable reasoning_effort parameter. It delivers 40% lower latency and 3x more requests per second than Mistral Small 3. Released under Apache 2.0, it is available via the Mistral API, Hugging Face, and as an NVIDIA NIM.

U.S. senators have urged ByteDance to shut down Seedance 2.0, its AI video generation app, citing serious intellectual property concerns. Lawmakers reportedly called it the most glaring example of copyright infringement, suggesting the tool may have been trained on or reproduces protected content without authorization. The pressure adds to ongoing scrutiny ByteDance faces in the U.S. over its Chinese ownership and data practices, and signals growing congressional focus on AI tools that potentially violate copyright law.

Google is expanding its Personal Intelligence feature in the U.S. to free-tier users across AI Mode in Search, the Gemini app, and Gemini in Chrome. The feature connects Google apps like Gmail and Google Photos to deliver tailored responses, such as shopping recommendations based on past purchases, custom travel itineraries from hotel confirmations, and tech support using device info from receipts. Users control which apps are connected and can toggle them off anytime. The feature is not available for Workspace business or education accounts.

OpenAI has released GPT-5.4 mini and nano, its most capable small models yet, optimized for coding, tool use, and high-volume API workloads. GPT-5.4 mini runs more than 2x faster than GPT-5 mini while approaching GPT-5.4 performance on benchmarks like SWE-Bench Pro and OSWorld-Verified. GPT-5.4 nano targets classification, data extraction, and coding subagents. Pricing is $0.75 per million input tokens for mini and $0.20 for nano. Both are available in the API today; mini also works in Codex and ChatGPT.

Nvidia has officially announced CloudXR 6.0 integration with Apple Vision Pro, developed in partnership with Apple and enabled in visionOS 26.4. The SDK is the first to allow sharing user gaze data over a secure connection, enabling foveated streaming at 4K resolution and 120Hz without a tethered PC. Enterprise applications include Immersive for Autodesk VRED, used by BMW Group, Kia, Rivian, and Volvo Group for 1:1 scale automotive design reviews with RTX-powered ray tracing. Foxconn and Switch are also adopting the technology.

Microsoft is reshuffling Copilot leadership, appointing Jacob Andreou to lead the Copilot experience across both consumer and commercial products, reporting directly to CEO Satya Nadella. Andreou, who previously worked at Snap, will oversee design, product, growth, and engineering. Microsoft AI CEO Mustafa Suleyman will shift focus to building Microsoft's own AI models. The change follows the retirement of veteran executive Rajesh Jha and aims to unify previously separate consumer and commercial Copilot efforts into one integrated system.

Nvidia unveiled the Groq 3 language processing unit at GTC 2026 in San Jose, the first chip resulting from its $20 billion deal to license technology from inference startup Groq Inc. and hire founder Jonathan Ross. The Groq 3 LPX server rack packs 256 LPUs with 128GB of solid-state RAM and 40 petabytes per second of bandwidth. Paired with Nvidia's Vera Rubin NVL72 GPU rack, the combo delivers 35 times higher throughput per megawatt and targets 1,500 tokens per second for multiagent AI workloads.

NVIDIA has launched Space Computing, bringing AI compute to orbital data centers and space missions. The NVIDIA Space-1 Vera Rubin Module delivers up to 25x more AI compute than the H100 GPU for space-based inferencing, while IGX Thor and Jetson Orin platforms enable edge AI in size-, weight-, and power-constrained environments. Partners including Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud are deploying these platforms for autonomous operations, geospatial intelligence, and on-orbit data processing.

NVIDIA unveiled DLSS 5 at GTC, calling it the company's most significant graphics breakthrough since real-time ray tracing debuted in 2018. Arriving this fall, DLSS 5 uses a real-time neural rendering model that takes color and motion vectors as input and infuses scenes with photoreal lighting and materials at up to 4K resolution. CEO Jensen Huang called it the GPT moment for graphics. Major publishers including Bethesda, CAPCOM, Ubisoft, and Warner Bros. Games will support it, with titles like Starfield, Assassin's Creed Shadows, and Resident Evil Requiem confirmed.

NVIDIA announced the Nemotron Coalition, a global collaboration of AI labs and model builders working to develop open frontier AI models. Inaugural members include Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam, and Thinking Machines Lab. The first project will be a base model co-developed by NVIDIA and Mistral AI, trained on NVIDIA DGX Cloud and open sourced to underpin the upcoming Nemotron 4 model family, enabling developers worldwide to specialize AI for their industries.

Manus has launched a Desktop app featuring My Computer, a capability that brings the previously cloud-only AI agent onto users' local machines. Available now for macOS and Windows, it executes terminal commands to read, edit, and organize local files, launch applications, and leverage local GPUs for tasks like running language models. Every command requires explicit user approval before execution. Manus can also remotely access your machine from any device, and integrates with Gmail and Google Calendar to bridge local and cloud workflows.

Meta is planning layoffs affecting 20% or more of its roughly 79,000 employees, according to three anonymous sources who spoke to Reuters. The cuts, not yet finalized in timing or scale, are driven by rising AI infrastructure costs and expectations that AI-assisted workers will enable smaller teams. CEO Mark Zuckerberg has signaled the plans to senior leaders. If confirmed, it would be Meta's largest restructuring since its 2022–2023 "year of efficiency," when it cut over 21,000 jobs.

Niantic has revealed that players of Pokémon Go and its other AR apps unknowingly helped build a dataset of over 30 billion real-world images. The company is now using that massive visual dataset to power navigation systems for delivery services. The images and scans were collected through normal gameplay and AR features, turning millions of players into unwitting contributors to a large-scale AI mapping project with real commercial applications.

Andrew Yang, Forward Party founder and former presidential candidate, told CNBC's Squawk Box that the US should stop taxing workers and start taxing AI companies benefiting from automation. Yang cited AI insiders telling him the next six months will eclipse the past decade in AI development, and referenced Anthropic CEO Dario Amodei's warning that AI could automate 50% of entry-level white-collar jobs within five years. Billionaire Vinod Khosla and Sen. Bernie Sanders have made similar proposals.

NBCUniversal is adding several AI-powered features to the Peacock mobile app, headlined by Your Bravoverse, launching this summer. The experience uses an Andy Cohen digital avatar — built with Synthesia's technology — to guide users through personalized, swipeable playlists drawn from over 5,000 hours of Bravo footage, generating more than 600 billion possible viewing variants. Peacock will also debut AI-enabled live vertical video for NBA games in 9:16 format this spring, a streaming industry first, plus new mobile games based on Law & Order and Jeopardy!

Meta AI is expanding its real-time content offerings by partnering with News Corp, Le Figaro, Prisa, and Süddeutsche Zeitung to deliver global news, entertainment, and lifestyle stories across Meta's apps and devices. When users ask news-related questions, Meta AI will now surface information and links from these diverse international sources. The partnerships also aim to drive traffic to publishers' websites, helping them reach new audiences while improving Meta AI's ability to deliver timely, balanced, and relevant content.

NVIDIA's Nemotron 3 Super, a high-performance language model, is now available across Perplexity's full product lineup, including the main Perplexity app, the Agent API, and Perplexity Computer. The rollout gives developers and end users access to the model through multiple interfaces, from consumer-facing search to programmatic agent workflows and the Computer product. The integration expands the range of AI models available within Perplexity's ecosystem beyond its existing lineup.

OpenAI has published a detailed overview of prompt injection attacks, a security threat where malicious instructions hidden in web pages, emails, or documents trick AI agents into taking unintended actions like stealing user data. As AI systems gain access to sensitive data and take autonomous actions, the risk grows. OpenAI's defenses include safety training via its Instruction Hierarchy research, AI-powered monitors, sandboxing, red-teaming, and a bug bounty program. ChatGPT's agent mode includes Watch Mode and confirmation prompts for sensitive actions.

SS Innovations International has completed its first telesurgery using the compact SSi Mantra Tele Surgeon Console (TSC), a self-contained chair-based system with built-in electronics, lightweight glasses, and magnetic sensor-based controls. CEO Dr. Sudhir Srivastava performed the milestone procedure remotely from his New Delhi residence, conducting a robot-assisted coronary artery bypass on a patient 185 miles away at Manipal Hospital in Jaipur. The TSC's smaller footprint allows placement in physician offices, expanding surgical access without requiring a full operating room.

Perplexity has launched a reimagined version of its Comet Assistant, the agentic AI helper built into its Comet browser. First released in July 2025, the updated assistant can handle longer and more complex tasks, work across multiple browser tabs simultaneously, and take a wider variety of actions on the web. Internal tests show it performs 23% better than its predecessor. New permission controls let users decide whether Comet interacts directly with their browser, with that preference maintained throughout each task.

Michael Burry, the investor famous for predicting the 2008 housing collapse, has taken bearish positions against Nvidia and Palantir through his hedge fund Scion Asset Management, which bought more than $1 billion in put options on the two companies. Burry has posted cryptic warnings on X comparing today's AI boom to the dot-com bubble. Palantir CEO Alex Karp pushed back, calling the short "bats–t crazy." Goldman Sachs and Morgan Stanley CEOs separately warned of a 10-20% market correction within two years.

Sony AI has published FHIBE, the Fair Human-Centric Image Benchmark, a dataset of 10,318 consent-based images of 1,981 people from 81 countries, released in Nature to benchmark bias in computer vision AI. Unlike many existing datasets built from unconsented web scraping, FHIBE includes self-reported demographic annotations covering age, pronoun category, ancestry, and skin and hair color. Compared against 27 existing datasets, FHIBE sets a higher standard for diversity, consent, and bias mitigation in applications like facial recognition and autonomous vehicles.

OpenAI is seeking U.S. federal loan guarantees to help finance an AI infrastructure expansion exceeding $1 trillion, CFO Sarah Friar announced at a Wall Street Journal business conference. Friar described a potential funding ecosystem involving banks, private equity, and government, noting federal backing would lower borrowing costs and broaden credit access. The request is unusual for a Silicon Valley firm and comes alongside a $300 billion Oracle deal and the $500 billion Stargate data center venture with Oracle and SoftBank.

Google has enhanced Structured Outputs in the Gemini API with expanded JSON Schema support and improved property ordering. The update enables libraries like Pydantic (Python) and Zod (JavaScript/TypeScript) to work out-of-the-box, and adds support for keywords including anyOf, $ref, minimum, maximum, and additionalProperties. Property key ordering is now preserved for all Gemini 2.5 models. Early partners report the feature reduced API calls by up to 6x and eliminated broken JSON responses in production workflows.

Kim Kardashian admitted in a Vanity Fair lie detector video that she has been using OpenAI's ChatGPT for legal advice while studying to become a Bar-certified attorney, and that it has repeatedly caused her to fail practice tests. She described a "toxic" relationship with the chatbot, calling it her "frenemy." Kardashian passed the baby bar in 2022 and completed a four-year apprenticeship, and sat the full bar exam in July 2025. Her experience highlights AI hallucination risks.

The article content could not be reliably extracted, but based on the headline, YouTube creator PewDiePie is self-hosting AI using modded GPUs and has announced plans to build his own AI model. This follows reported warnings he issued to his subscribers about mainstream AI platforms. The move reflects a growing trend of tech-savvy creators pursuing self-hosted, privacy-focused AI alternatives rather than relying on commercial services like ChatGPT or similar tools.

Elon Musk outlined a vision on the All-In Podcast to transform Tesla's vehicle fleet into a massive distributed AI inference network. He proposed that a fleet of 100 million Teslas, each equipped with roughly one kilowatt of inference compute, could collectively deliver 100 gigawatts of AI processing power. The network would leverage vehicles' built-in power, cooling, and Wi-Fi connectivity. The compute would augment xAI's infrastructure, though challenges remain around latency, availability, customer compensation for electricity use, and security.

NVIDIA is pushing to deploy its H100 GPU in space-based data centers as a strategy to reduce the massive energy and cooling demands associated with AI computing on Earth. Space environments offer natural cooling advantages and access to solar power, potentially making orbital data centers more sustainable. The H100, one of NVIDIA's most powerful AI chips, is central to this initiative, which reflects growing industry interest in off-Earth infrastructure for AI workloads.

Anthropic is expanding its Economic Futures Programme to the UK and Europe, launching with a symposium at the London School of Economics. The programme offers research grants and Claude API credits to European researchers studying AI's effects on labor markets and productivity, plus evidence-based policy forums and expanded Anthropic Economic Index data for Europe. Anthropic's Economic Index shows Claude is widely used across the region, with UK users favoring academic research support, Germany using it for manufacturing, and France for culture and tourism.

OpenAI highlights Brazil as one of the world's most active AI adopters, ranking in the top three countries for weekly ChatGPT usage with around 140 million messages exchanged daily, and top two for active developers on the OpenAI API. A Microsoft study found 75% of Brazilian SMBs are optimistic about AI's impact. OpenAI says it is committed to collaborating with Brazil's public and private sectors to expand access across schools, farms, and small businesses.

Amazon has introduced a new "sassy" personality mode for Alexa+, its updated AI assistant, featuring censored profanity, razor-sharp wit, and playful sarcasm. The adults-only mode requires additional security checks like face scans to activate and won't work if an Amazon Kids profile is enabled on the account. Amazon warns users the tone may include mature subject matter but says Alexa+ will still avoid hate speech, illegal activities, and personal attacks. Other personality options include sweet, brief, and chill.

Meta is adding several Meta AI-powered tools to Facebook Marketplace, which sees over 3.5 million listings posted daily in the US and Canada. Sellers can now upload item images and let Meta AI auto-generate a draft listing, suggest prices based on similar local items, and send AI-drafted autoreplies to buyer inquiries. A new shipping dashboard simplifies prepaid label generation, and AI-generated profile summaries display seller history, item types, and ratings to build buyer trust.

The accessible source details point to this update: Perplexity's AI PC Is a Mac Mini Running an AI OS. Because the full article could not be reliably extracted or rewritten, this TLDR stays conservative and is based on headline-level information plus limited source context from macworld.com.

OpenAI has expanded its Video API with new capabilities powered by the Sora 2 model. Developers can now generate clips up to 20 seconds long, incorporate custom characters and objects into scenes, export footage in both 16:9 and 9:16 aspect ratios, extend existing scenes using video continuation, and run batch jobs for video generation. These updates give developers significantly more creative control and workflow flexibility when building video applications on top of the Sora 2 model.

Rivian revealed final pricing for its R2 midsize electric SUV at SXSW, with the first available trim being the dual-motor Performance version starting at $59,485, arriving in spring 2026. The Performance model offers 656 horsepower, 330 miles of EPA-estimated range, and a 0-60 time of 3.6 seconds. The originally promised $45,000 base model has been pushed to late 2027. CEO RJ Scaringe called the R2 "maybe the most important thing we've launched to date."

WordPress has launched my.WordPress.net, a free browser-based WordPress environment built on WordPress Playground technology that requires no sign-up, hosting plan, or domain. Sites are private by default, persistent across sessions, and store all data locally with roughly 100 MB of starting storage. An App Catalog offers one-click installs including a personal CRM, RSS reader via the Friends plugin, and an AI workspace that can modify plugins. Each device maintains its own separate installation.

Google has launched Groundsource, a Gemini-powered methodology from Google Research that transforms public records into historical disaster data. Using Gemini, the system analyzed decades of reports to identify over 2.6 million historical flood events across more than 150 countries, then used Google Maps to define precise geographic boundaries. The resulting dataset trained a new model capable of predicting urban flash floods up to 24 hours in advance. Forecasts are now available in Google's Flood Hub, and the approach could extend to landslides and heat waves.

Microsoft has launched Copilot Health, a secure, separate space within Copilot that aggregates health records from over 50,000 U.S. hospitals via HealthEx, wearable data from over 50 devices including Apple Health and Fitbit, and lab results from Function. The tool delivers personalized, actionable health insights without replacing doctors. It is informed by over 230 physicians across 24 countries, has achieved ISO/IEC 42001 certification, and user data is not used for model training. A waitlist is now open for U.S. adults.

Google Maps is launching two Gemini-powered features: Ask Maps, a conversational tool that answers complex location questions using data from over 300 million places and 500 million community contributors, rolling out now in the U.S. and India on Android and iOS; and Immersive Navigation, described as the biggest navigation update in over a decade, featuring vivid 3D visuals, natural voice guidance, alternate route tradeoff alerts, and parking assistance, rolling out today across the U.S.