Latest AI News

CapCut has launched Video Studio, a new timeline-free video creation tool available on CapCut Web. The canvas-based AI production workspace supports Dreamina Seedance 2.0 and is designed for creators at all skill levels. Unlike traditional video editors that rely on timelines, Video Studio offers a more flexible, canvas-based approach to assembling and producing video content, aiming to make AI-assisted storytelling more accessible.

Senate Democrats are pushing legislation to codify Anthropic's restrictions on autonomous weapons and mass surveillance into law. Sen. Adam Schiff (D-CA) is drafting a bill requiring human oversight in life-or-death AI decisions, while Sen. Elissa Slotkin (D-MI) introduced the AI Guardrails Act, which limits the Pentagon's use of AI for domestic surveillance, autonomous lethal weapons, and nuclear detonation. The effort follows the Trump administration blacklisting Anthropic as a supply-chain risk after it refused military terms that competitor OpenAI accepted.

ARC Prize has launched ARC-AGI-3, described as the only unsaturated agentic intelligence benchmark in the world. Unlike most benchmarks that test existing model knowledge, ARC-AGI-3 evaluates how AI systems learn. The results reveal a stark gap: humans score 100% while current AI scores under 1%. ARC Prize says this disparity demonstrates that artificial general intelligence has not yet been achieved. The benchmark is designed to remain meaningful as AI capabilities advance.

Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez are introducing the Artificial Intelligence Data Center Moratorium Act, which would halt all new AI data center construction nationwide until Congress passes federal AI legislation establishing worker and consumer protections, environmental safeguards, and civil rights defenses. The ban could remain in place for years, as Congress is far from passing any such legislation. The bill represents one of the most aggressive AI policy positions taken this Congress.

President Trump signed an executive order on January 23, 2025, establishing the President's Council of Advisors on Science and Technology (PCAST), revoking a Biden-era version from 2021. The council, comprising up to 24 members, will be co-chaired by the Assistant to the President for Science and Technology and the Special Advisor for AI and Crypto. PCAST will advise on AI, quantum computing, and biotechnology policy, with administrative support provided by the Department of Energy.

Google has launched Lyria 3 Pro, an advanced version of its AI music generation model that can create tracks up to 3 minutes long with structural awareness, supporting elements like intros, verses, choruses, and bridges. The model is now available across Vertex AI, Google AI Studio, the Gemini API, Google Vids, the Gemini app for paid subscribers, and ProducerAI. All outputs are watermarked with SynthID, and the model avoids mimicking named artists.

Claude has expanded mobile access to its work tool integrations, allowing users to interact with Figma, Canva, and Amplitude directly from their phones. Users can now explore Figma designs, create Canva slides, and check Amplitude dashboards through the Claude mobile app. The update brings previously desktop-available integrations to mobile, making it easier for professionals to access design, presentation, and analytics tools on the go.

"Figure's humanoid robot F.03 has made history as the first humanoid robot to visit the White House, according to a post from Figure founder and CEO Brett Adcock. The milestone marks a symbolic moment for the humanoid robotics industry, with Figure positioning itself at the forefront of bringing advanced bipedal robots into high-profile public and governmental settings. Figure has been developing its humanoid platform with ambitions to deploy robots in commercial and industrial environments."

Lovart has launched a new Move Object tool that allows users to reposition any element within an image without needing masks, layers, or manual re-rolling. Using either a rectangular or lasso selection tool, users can click on any object, drag it to a new location, and optionally enter a text prompt to apply further modifications. The workflow produces a single, clean, consistent image output. The Move Object tool is now live and available on the Lovart platform.

Google Research has unveiled TurboQuant, a theoretically grounded quantization algorithm for compressing large language models and vector search engines with zero accuracy loss. Developed by researchers Amir Zandieh and Vahab Mirrokni, TurboQuant combines two companion methods, PolarQuant and Quantized Johnson-Lindenstrauss, to eliminate memory overhead common in traditional quantization. It compresses key-value caches to just 3 bits without fine-tuning, achieves up to 8x speedup over bit unquantized keys on H100 GPUs, and reduces KV memory by at least 6x.

OpenAI is shutting down its Sora AI video app just months after its fall 2025 launch, and Disney is pulling out of the $1 billion investment deal it signed with OpenAI last December. Disney had agreed to invest and license some of its characters for use in Sora, with the goal of integrating the technology into Disney+. OpenAI says AI video will continue within ChatGPT, but the standalone Sora app is being discontinued. The closure leaves Google as the dominant player in AI video generation.

Sora, the AI video generation tool, is shutting down both its app and API. The team announced the closure in a post addressed to its user community, thanking creators who used the platform and built community around it. The announcement acknowledged the news would be disappointing to users. The team promised to share additional details soon, including specific timelines for the app and API shutdown and further information on next steps.

Nvidia CEO Jensen Huang declared on the Lex Fridman podcast that he believes AGI has already been achieved, responding to Fridman's definition of AGI as an AI system capable of starting and running a successful tech company worth over $1 billion. Huang said, "I think it's now. I think we've achieved AGI." He then partially walked back the claim, noting that AI agents frequently lose user interest after a few months and that the odds of 100,000 agents building a company like Nvidia are zero percent.

Ai2 has released MolmoWeb, an open-weight visual web agent available in 4B and 8B parameter sizes that operates entirely from browser screenshots without parsing HTML or accessibility trees. Alongside the model, Ai2 released MolmoWebMix, a dataset containing 30,000 human task trajectories across 1,plus websites, 590,000 subtask demonstrations, and 2.2 million screenshot question-answer pairs — described as the largest public collection of human web-task execution ever assembled. MolmoWeb outperforms open-weight rivals on four live-website benchmarks including WebVoyager and DeepShop.

Sakana AI has launched Sakana Chat, a web chat service powered by its new Namazu alpha model series, which adapts open-weight frontier LLMs for Japanese users via post-training. The three models — Namazu-DeepSeek-V3.Terminus, Llama-3.Namazu-405B, and Namazu-gpt-oss-120B — correct biases and censorship tendencies in base models, reducing DeepSeek's 72% question refusal rate to nearly zero. The models include web search and maintain benchmark parity with their base models on reasoning, coding, and Japanese language tasks.

Smallest.ai has launched Lightning V3, a text-to-speech model built for conversational voice agents. Lightning V3.1 supports 15 languages including English, Hindi, Tamil, and Spanish, with automatic language detection and mid-sentence switching. Lightning V3.2 adds instruction-following controls for emotional register, pitch, volume, and whispering. The model achieves a 3.89 MOS score and 5.38% word error rate, and beats OpenAI's gpt-4o-mini-tts on naturalness in roughly 76% of comparisons. Voice cloning works from just 5 to 15 seconds of audio.

Anthropic has launched Auto Mode for Claude Code, a permissions setting that lets Claude autonomously handle tool-call decisions without constant approval prompts. A classifier reviews each action before it runs, blocking destructive operations like mass file deletion, sensitive data exfiltration, or malicious code execution. It serves as a safer middle ground between Claude Code's conservative default mode and the risky --dangerously-skip-permissions flag. Auto Mode is available now as a research preview for Team plan users, supporting Claude Sonnet 4.6 and Opus 4.6, with Enterprise and API rollout coming soon.

Meta has partnered with Arm to co-develop multiple generations of custom data center CPUs designed for large-scale AI deployments. The first product, the Arm AGI CPU, is Arm's first data center CPU built specifically for the AI era, delivering faster performance per rack more efficiently than legacy CPUs. Meta serves as lead partner and co-developer, and will release board and rack designs under the Open Compute Project. The chip will also work alongside Meta's custom MTIA silicon.

The OpenAI Foundation, chaired by Bret Taylor, plans to invest at least $1 billion over the next year across life sciences, jobs and economic impact, AI resilience, and community programs. Focus areas include AI research for Alzheimer's disease, biosecurity, child safety, and open health datasets. OpenAI co-founder Wojciech Zaremba will lead AI Resilience efforts, while Jacob Trefethen, formerly of Coefficient Giving, will head Life Sciences. The investment counts toward a previously announced $25 billion commitment to curing diseases.

OpenAI has launched a product discovery program for merchants in ChatGPT, allowing businesses to share product feeds so their items appear in AI-powered shopping results. Merchants can integrate catalogs via SFTP, APIs, or supported commerce platforms without rebuilding existing systems. Purchases are completed on merchant-owned sites with no fees charged by OpenAI. The program is built on the Agentic Commerce Protocol and is currently live in the U.S., with Etsy and Shopify merchants among early participants.

OpenAI is expanding its Agentic Commerce Protocol (ACP) to power richer product discovery in ChatGPT, launching more visual and immersive shopping experiences. Users can now browse products visually, compare options side-by-side with pricing and reviews, and upload images to find similar items. Major retailers including Target, Sephora, Nordstrom, Walmart, and Best Buy have integrated with ACP. Walmart is launching an in-ChatGPT app with account linking and loyalty payments. Updates roll out to all free, Plus, and Pro users this week.

KDnuggets author Kanwal Mehreen recommends 10 X accounts for cutting through AI hype to find reliable LLM updates. The list includes DAIR.AI for paper threads, Andrej Karpathy for deep learning intuition, Sebastian Raschka for practical implementation, alphaXiv for arXiv paper discovery, The Rundown AI for product launches, AK for new model releases, Ahmad Osman for local LLM infrastructure, Matt Wolfe for tool roundups, Simon Willison for hands-on building, and Ethan Mollick for real-world AI impact on work and education.

Brett Adcock, founder of Figure AI and Archer Aviation, has launched Hark, a new artificial intelligence lab focused on building what he describes as the most advanced personal intelligence in the world. The company operated in stealth for eight months while assembling a combined AI and hardware team. Hark plans to develop both advanced software intelligence and new hardware products, signaling Adcock's continued ambition at the intersection of AI and physical technology.

Instagram has launched an AI Transition feature for Stories that uses AI-powered morphing to stitch multiple still images into a seamless video clip. Users select two or more photos from their image gallery, and the tool generates an animated sequence that posts as a Story. The feature requires agreeing to Meta AI terms. Generated Stories can also be downloaded and re-uploaded as Reels. Instagram's Help Center notes the feature is currently rolling out to some users.

OpenAI is offering private-equity firms preferred equity stakes with a guaranteed 17.5% minimum return to attract partners for enterprise AI joint ventures, outbidding rival Anthropic, which offered no such returns. OpenAI is courting firms including TPG, Bain Capital, Advent International, and Brookfield to raise about $4 billion at a $10 billion pre-money valuation. The joint ventures aim to deploy AI tools across PE-owned companies, easing upfront costs ahead of potential IPOs. Thoma Bravo declined to participate, citing concerns over long-term profit profiles.

Cisco released its free LLM Security Leaderboard, ranking models on resistance to single- and multi-turn adversarial attacks. Anthropic dominated, claiming 8 of the top 10 spots, with Claude Opus 4.5 ranking first, followed by Sonnet 4.5 and Haiku 4.5. OpenAI's GPT-5.2 and GPT 5 Nano placed 7th and 9th. Mistral's Magistral Small 2509 and Ministral 3 14b Instruct ranked last, with DeepSeek, Cohere, Qwen, and xAI models also appearing in the bottom 10.

OpenAI is in advanced talks to buy electricity from Helion Energy, a fusion startup backed by CEO Sam Altman, who has recused himself from deal discussions and stepped down as Helion's board chair. The agreement would give OpenAI an initial 12.5% of Helion's production, targeting 5 gigawatts by 2030 and scaling to 50 gigawatts by 2035. Helion, which also has a power deal with Microsoft, says it is near scientific breakeven but no private fusion company has yet achieved that milestone.

Luma Labs has released Uni-1, a new AI model designed to think and generate pixels at the same time, rather than separating reasoning from image generation. Unlike traditional models that handle these processes sequentially, Uni-1 integrates both into a unified approach, which Luma describes as less artificial and more intelligent. The release marks a notable shift in how generative AI models handle visual output alongside reasoning.

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.