Latest AI News

OpenAI is offering verified university students in the United States and Canada $100 in free Codex credits, equivalent to 2,500 ChatGPT credits. Students enrolled at degree-granting universities can claim the offer by verifying their status through SheerID using a university email. Credits are added to a personal ChatGPT account, work across Free, Go, Plus, and Pro plans, extend Codex usage beyond plan limits, and expire 12 months after the grant date. Only one credit offer is allowed per student.

Blue Origin has filed an FCC application for Project Sunrise, a proposed constellation of up to 51,600 satellites designed to provide orbital data center services for AI computing workloads. The satellites would operate in sun-synchronous orbits between 500 and 1,800 kilometers altitude, using optical intersatellite links and solar power. Blue Origin argues space-based compute is cheaper than terrestrial alternatives due to always-on solar energy and no land or grid costs. The company joins SpaceX and startup Starcloud in the emerging orbital data center race.

Perplexity has launched Perplexity Health, a suite of connectors linking electronic health records from over 1.7 million care providers, Apple Health, and wearables including Fitbit, Ultrahuman, and Withings to deliver personalized health answers. Powered by partners like b.well and Terra API, it draws on medical records, lab results, and wearable data simultaneously. Answers cite peer-reviewed journals and clinical guidelines. Health data is encrypted and never used to train AI models. The feature rolls out to Pro and Max users in the US first.

The White House released a light-touch AI policy blueprint on Friday, urging Congress to codify federal rules that would preempt state AI laws deemed burdensome to innovation. The framework addresses political bias in models, child protections including age-gating requirements, and asks Congress to avoid creating new federal AI agencies. It also calls on lawmakers to codify Trump's ratepayer protection pledge, signed by Amazon, Google, and OpenAI. Senate Commerce Chair Ted Cruz hopes to advance legislation by end of April.

Amazon is reportedly developing an Alexa-centered smartphone code-named "Transformer," more than a decade after the failed Fire Phone. According to Reuters, the device is being built within Amazon's ZeroOne group, led by J Allard, formerly of Microsoft's Zune and Xbox teams. The phone draws inspiration from the minimalist $700 Light Phone and may skip a traditional app store in favor of mini apps. No release timeline or price has been announced.

YouTube has started surveying viewers in March 2026, asking whether videos "feel like AI slop" as part of a crackdown on low-quality AI-generated content. Users are shown a video, title, and thumbnail, then asked to rate it on a five-point scale from "not at all" to "extremely." The survey was first spotted by vidIQ on X. Some users speculate YouTube is collecting the data to train Google's Veo video model, though YouTube has not confirmed what actions follow repeated AI slop flags.

Anthropic's Claude Cowork tool has introduced Projects, a new feature designed to help users organize tasks and context around a specific area of work. Unlike cloud-based storage, files and instructions are saved locally on the user's computer, keeping data on-device. Users can import existing projects with a single click or start a new one from scratch. The feature aims to consolidate work into one focused environment within Cowork, streamlining how users manage ongoing tasks.

DoorDash has launched DoorDash Tasks, a program letting its 8 million Dashers earn money by completing short data-collection activities beyond deliveries. Tasks include photographing restaurant dishes, hotel entrances, and recording speech in other languages to train AI and robotic systems. Since 2024, Dashers have completed over 2 million tasks. A standalone app is being piloted for AI and robotics data collection. Partners span retail, insurance, hospitality, and tech. The service is available in select U.S. locations, excluding California, New York City, Seattle, and Colorado.

Val Kilmer, who died in 2025 after battling throat cancer, will appear via generative AI in the indie film As Deep as the Grave, directed by Coerte Voorhees. Kilmer was originally cast as Father Fintan, a Catholic priest and Native American spiritualist, but was too ill to film. His estate and daughter Mercedes approved the AI recreation, which uses family-provided photos and footage across different life stages. The production followed SAG guidelines and compensated Kilmer's estate. Co-stars include Tom Felton and Abigail Breslin.

OpenAI released GPT-5.4 on March 5, 2026, its most capable frontier model combining advanced reasoning, elite coding, and native computer use. The GPT-5.4 Thinking variant supports complex reasoning with upfront planning and mid-response steerability, which OpenAI reportedly uses to monitor internal coding agents for misalignment. The model achieves a 75% success rate on OSWorld-Verified, reduces hallucinations by 33%, supports a 1 million token context window, and cuts token usage by 47% via a new Tool Search API feature.

Anthropic conducted what it claims is the largest and most multilingual qualitative study ever, interviewing 80,508 Claude users across 159 countries and 70 languages in December using an AI interviewer tool. Results show users' top hopes include professional excellence (18.8%), personal transformation (13.7%), and life management (13.5%). Key fears include job displacement and loss of human cognitive ability. Notably, hope and alarm coexisted within individuals rather than dividing respondents into opposing camps.

OpenAI is acquiring Astral, the company behind popular open source Python tools uv, Ruff, and ty, to accelerate its Codex AI coding platform. Astral's tools are used by millions of Python developers for dependency management, linting, and type safety. After closing, the Astral team will join the Codex team, with plans to integrate their tooling directly into Codex workflows. Codex has already seen 3x user growth and 5x usage increase this year, with over 2 million weekly active users.

Adobe has launched Firefly Custom Models in public beta, letting creators and brands train AI image generators on their own artwork to preserve consistent styles across projects. The tool can maintain details like stroke weight, color palettes, lighting, and character features. Custom models are private by default and won't feed Adobe's general Firefly training data. Adobe also checks uploaded images for Content Authenticity Initiative credentials, blocking assets whose creators have opted out of AI training.

Uber will invest $1.25 billion in Rivian as part of a robotaxi partnership, with an initial $300 million paid at signing. The companies plan to deploy 10,000 autonomous Rivian R2 vehicles in cities including San Francisco and Miami starting in 2028, expanding to 25 cities by 2031. The fleet will be exclusive to Uber's app. Rivian is also developing custom AI chips and plans to add lidar to R2 vehicles in 2026 to support Level 4 autonomy.

CodeRabbit has launched Plan, a collaborative planning tool that turns vague ideas into agent-ready prompts before any code is written. Formerly called Issue Planner, Plan lets teams start from a concept, text prompt, or image rather than requiring an existing ticket. It generates editable, phased plans with context pulled from codebases, knowledge bases, and tools like Notion, Jira, and Linear. The goal is to reduce code churn, technical debt, and rework caused by poorly written prompts handed to coding agents.

Lovable has expanded beyond full-stack app building to handle file analysis, document generation, and image and video creation. The AI agent can now run Python scripts, install tools, convert file formats, and process data in a secure environment. Users can upload CSVs, PDFs, Excel files, and slide decks, then generate branded invoices, pitch decks, and reports ready to download. Supported formats include PowerPoint, Word, PDF, CSV, Excel, JSON, XML, images, and video, all within a single conversation.

Cursor has launched Composer 2, a frontier-level coding model now available in its AI code editor. The model scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual, significantly outperforming Composer 1.5. It is priced at $0.50/M input and $2.50/M output tokens, with a faster variant at $1.50/M input and $7.50/M output. Composer 2 was trained using continued pretraining and reinforcement learning on long-horizon coding tasks, enabling it to solve challenges requiring hundreds of actions.

Microsoft has unveiled MAI-Image-2, its latest text-to-image model, which has reached the #3 spot on the Arena.ai text-to-image leaderboard. The model emphasizes enhanced photorealism with natural lighting and accurate skin tones, reliable in-image text generation, and rich scene creation. MAI-Image-2 is available now in the MAI Playground, is rolling out on Copilot and Bing Image Creator, and API access is available for select Microsoft customers with broader developer access via Microsoft Foundry coming soon.

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.