Latest AI News

NVIDIA Blackwell GPUs with Confidential Computing are now powering confidential inference in Apple's Private Cloud Compute, which is expanding beyond Apple's data centers to Google Cloud. Announced at WWDC, NVIDIA is collaborating with Apple and Google to support next-generation Apple Intelligence features using Apple Foundation Models, including models leveraging Gemini family technologies. NVIDIA's Confidential Computing provides hardware-rooted trust, encrypted communication, and remote attestation, ensuring no one, including system builders, can access users' data or conversations during processing.

Anthropic has launched two public beta features for Claude Managed Agents on the Claude Platform. Scheduled deployments let agents run on a cron schedule, automatically starting new sessions for recurring tasks like nightly data syncs or weekly compliance scans, with no scheduler to build or host. Environment variable vaults securely store API keys for CLI tools, attaching credentials at the network boundary so agents never see raw keys. Supported CLIs include Browserbase, KERNEL, Notion, Ramp, and Sentry. Rakuten, Actively AI, and Ando are already using scheduled deployments.

Google has launched Gemini 3.5 Live Translate, a new audio model delivering near real-time speech-to-speech translation across 70+ languages. Unlike turn-by-turn systems, it generates speech continuously, staying just seconds behind the speaker while preserving intonation, pacing, and pitch. The model is rolling out in Google AI Studio for developers, Google Meet for enterprise users in private preview, and the Google Translate app on Android and iOS. All generated audio is watermarked with SynthID to help prevent misinformation.

Midjourney founder David Holz has begun sending out invites for the company's first hardware product launch. Holz noted a limited number of spots remain and encouraged anyone who believes they should have received an invite but hasn't to reach out directly. No details about the hardware itself were disclosed in the announcement. This marks a significant expansion for Midjourney beyond its AI image generation software into physical hardware products.

OpenAI has submitted a confidential draft S-1 registration statement to the SEC, a step toward a potential IPO. The company preemptively announced the filing, expecting it to leak. OpenAI stated it has not decided on timing and noted a public offering may still be a while away, as some planned actions are easier to execute as a private company. The filing preserves the option to go public sooner if that proves best, calling it a complicated set of tradeoffs.

Anthropic is releasing a Swift package that lets Apple developers use Apple's Foundation Models framework to call Claude for complex tasks. The integration works on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. Developers can use Apple's on-device models for fast local tasks like summarization, then hand off to Claude for multi-step reasoning, code generation, or web search. The package handles streaming, tool calls, and structured responses, with typed Swift outputs from @Generable annotations feeding cleanly into Claude API calls.

OpenAI co-founders Sam Altman and Jakub Pachocki have outlined a three-part plan for the company's third phase: building an automated AI researcher by March 2028, accelerating broad economic growth, and giving every person on Earth access to a personal AGI. The plan emphasizes distributed power over concentration, calling for international coordination on AI safety. OpenAI also warns that fully automating human decision-making is both dangerous and undesirable, stressing that human judgment must remain central.

Anthropic has launched two new features for Claude connector developers. Published connectors in the directory now have a performance dashboard showing active users, total tool calls, directory rank, health scores, error rates, and latency breakdowns per tool. Usage can also be compared across Claude, Claude Code, and Cowork. Additionally, developers can now submit MCP servers to the directory directly in-app. The observability dashboard is in public beta and requires a Team or Enterprise account with Admin or Owner access.

OpenAI has launched the Economic Research Exchange, a program connecting external researchers with OpenAI's tools and datasets to study AI's economic effects on workers, firms, and institutions. Selected researchers will conduct structured, project-based collaborations with defined milestones and data governance safeguards. OpenAI seeks proposals in fields like labor economics, productivity, and inequality. Applications are open now and close July 5, 2026, with selected researchers notified by July 31. The initiative builds on OpenAI's existing Signals measurement efforts.

At WWDC26, Apple unveiled the next generation of Apple Intelligence and Siri AI, an entirely new version of Siri described as profoundly more intelligent and capable. Siri AI is deeply integrated across iPhone, iPad, Mac, Apple Watch, and Apple Vision Pro, with screen awareness, personal context understanding, web search, and cross-app actions. A dedicated Siri app syncs conversation history via iCloud. Updates also include iOS 27, macOS 27, new parental controls, Safari Notify Me, and photorealistic Image Playground.

Amazon has added an AI-powered custom merchandise design feature to Alexa for Shopping, available now to all U.S. customers via the Amazon Shopping app or Amazon.com. Users describe an idea in a text prompt, and the tool generates a design in seconds for products including T-shirts, hoodies, tumblers, and water bottles. Designs can be shared with friends and family, who can each order their own. Production is handled through Amazon's Merch on Demand service with Prime-eligible shipping.

The American Federation of Musicians has sued Universal Music Group and Warner Music Group in the US District Court for the Southern District of New York, alleging the labels licensed member recordings to AI music generators Suno and Udio without compensating or crediting session musicians. The AFM argues the deals triggered a "new use" provision in its collective bargaining agreement. The union seeks monetary damages and disclosure of which recordings were used for AI training. Both labels denied wrongdoing, calling the lawsuit unproductive amid ongoing negotiations.

Google has upgraded NotebookLM with agentic chat capabilities, code execution, and expanded output formats, powered by Gemini 2.5. Each notebook now includes a secure cloud computer with over 100 curated software skills for deeper analysis. New output formats include PDFs, Excel, PowerPoint, CSV, JSON, and data visualizations. The upgraded system achieved a 65% win rate over the prior version, with a 78.2% win rate in advanced web research. Updates are rolling out to Google AI Ultra and Workspace AI customers globally.

Google announced at WWDC that Apple developers can now call cloud-hosted Gemini models directly through Apple's native Foundation Models framework, available starting with iOS 27, macOS 27, and related platforms. The integration uses Firebase AI Logic via the Firebase Apple SDK, letting developers swap between on-device Apple models and cloud-hosted Gemini with minimal code changes. Gemini is also now integrated into Xcode, offering agentic coding assistance for reviewing code, fixing bugs, and building features without switching tools.

Anthropic is partnering with synthetic, computational, and analytical chemists to improve Claude's chemistry capabilities. In a new white paper, Anthropic chemist David Kamber tested Claude Opus 4.7, Opus 4.6, and Sonnet 4.6 against dedicated NMR software ChemDraw and MestReNova across 20 novel compounds. Opus 4.7 matched or outperformed both tools on hydrogen shift prediction, averaging ±0.079 ppm error, and tied MestReNova on carbon prediction. Claude also successfully performed inverse structure elucidation from NMR spectra alone, a task existing software leaves entirely to chemists.

OpenAI appears to have introduced a feature allowing ChatGPT users to send emails directly from within writing blocks in the ChatGPT interface, according to a post from the official ChatGPT account on X. However, the original post text was not available for review, so specific details about supported platforms, setup requirements, or limitations cannot be confirmed. The title suggests this is a notable workflow integration, but a fully accurate summary requires access to the complete source material.

New York's state legislature passed a one-year moratorium on new large data centers, potentially the first statewide ban of its kind in the US. The bill targets facilities with peak demand of at least 20 megawatts and directs the state's environmental agency to assess electricity, water, and land use impacts. Companies must also fund public hearings three months before approval. Democratic Governor Kathy Hochul, who has until December to decide, has not indicated whether she will sign or veto the bill.

Google has released Quantization-Aware Training (QAT) checkpoints for its Gemma 4 model family, optimizing them for on-device use on mobile phones, laptops, and consumer GPUs. Unlike standard post-training quantization, QAT integrates compression into the training process to preserve model quality. A custom mobile quantization schema reduces the Gemma 4 E2B model's memory footprint to under 1GB. Weights are available on Hugging Face in GGUF and compressed tensor formats, with support for llama.cpp, Ollama, LM Studio, vLLM, and MLX.

OpenAI has launched Lockdown Mode, an optional security setting for ChatGPT designed to protect users handling sensitive data from prompt injection attacks, where malicious instructions hidden on webpages trick AI systems into leaking information. The mode limits certain features, disabling Deep Research and Agent Mode entirely, and restricting internet image loading and file downloads. It is available to all accounts including free-tier users via Settings under Safety and Security. OpenAI is also rolling out an active session manager to monitor account access.

Reve has launched Reve 2.0, an image generation model that replaces text-based intermediate representations with structured layouts. Each layout defines every element's location, size, description, and optional attributes like color or image references, acting as a backbone separating semantic intent from pixel rendering. Built on a unified Large Layout Model trained using Qwen open-source LLMs and a pipeline of billions of images, Reve 2.0 claims top arena rankings among sub-$1 trillion companies, trained on 10x fewer GPUs than competitors.

OpenAI is rolling out Dreaming V3, an upgraded memory system for ChatGPT, now available to Plus and Pro users in the US. Unlike the original saved memories launched in April 2024, Dreaming uses a background process to automatically synthesize memories from chat history, keeping context fresh and accurate over time. The update addresses staleness, correctness, and scalability issues. Users can review a memory summary page and edit what ChatGPT knows about them. Free and Go users will gain access in coming weeks.

NVIDIA has released Nemotron 3 Ultra, a 550B-parameter Mixture-of-Experts model with 55B active parameters designed for long-running AI agents. It delivers 5x higher throughput than comparable open models and reduces agentic task costs by up to 30% by using fewer tokens per turn. Key innovations include a hybrid Mamba-Transformer architecture, NVFP4 precision, and Multi-Teacher On-Policy Distillation training. NVIDIA also launched Nemotron 3.5 Content Safety, a 4B guardrail model covering 23 safety categories, and Nemotron 3.5 ASR for sub-100ms voice agents.

OpenAI has updated GPT-Rosalind, its purpose-built model series for enterprise life sciences research, combining GPT-5.5's agentic coding and tool-use capabilities with stronger intelligence in drug-discovery domains including medicinal chemistry and genomics. The update improves performance across biology analysis, design, and experimental workflows. OpenAI also introduced LifeSciBench, an expert-judged benchmark covering six workflow areas, on which GPT-Rosalind outperforms GPT-5.5 and Grok. The model is now available in research preview to eligible organizations globally via a trusted-access deployment structure.

Krea has launched Krea 2 Turbo, a new AI image generation model that produces high-quality images in approximately two seconds. The model is compatible with style references, moodboards, and LoRAs, making it adaptable for a range of creative workflows. Krea 2 Turbo is available to try for free at krea.ai, positioning it as a fast and accessible option for designers and creators who need rapid AI-assisted image generation results.

Runway has launched Aleph 2.0, a new AI video editing model now available inside its newly released Edit Studio. Aleph 2.0 is designed for precise, targeted edits that modify only the specific elements a user wants changed while leaving the rest of the shot completely untouched. The tool aims to give video creators greater control and accuracy, directly addressing one of the most persistent challenges in AI-assisted video editing: unintended changes to surrounding footage.