What the AI world is actually talking about this week.
Groundswell tracks the stories gaining momentum across AI and ranks them by how widely they’re covered. One weekly digest — strongest signals first, the noise left behind.
Free · one email each week · unsubscribe anytime
You’re in. The next issue lands in your inbox this week.
Dario Amodei urged frontier labs to slow development for safety, but Sam Altman endorsed him while Trump, David Sacks, Zuckerberg and China rejected the proposal.Anthropic runs 30,000 internal agents, while Meta, OpenAI, Perplexity and Pion are pushing agents into personal, sponsored, local and company-wide work.TypeSafe AI’s code-only Jev claims 20–200× faster results and 40–400× lower prices, while a Jev browser agent found a flight in seven seconds for $0.0039.OpenAI ruled out a 2026 IPO, while SoftBank borrowed $11.9B toward Masayoshi Son’s planned roughly $65B investment.Anthropic merged Claude’s chat and background work modes so tasks continue while users are away, with existing Cowork chats, projects and files preserved.Anthropic added parallel cloud threads, pull requests and tests to Claude Code Projects, while Claude Mods let users tailor its interface and behaviour.OpenAI’s Astra for Law indexes 230 million US case-law URLs and scored 54.0% versus 38.7% for GPT-6 Astra using web search, while Astra decoded a 1941 message.Google introduced Gemini 3.8 Live models that keep speaking while reasoning, with Extended Thinking reportedly ranking first for speech-to-speech quality.OpenAI published a framework for documenting and disclosing model misbehavior, including hidden mistakes and forged data.Apple released an English-only beta of rebuilt Siri AI in iOS 27 that reads screens, uses messages, email and photos, and acts across supported apps.DeepSeek released V4.1-Flash, a 552B MoE model with a one-million-token context and an 890-byte KV cache per token, down from 3,514 bytes.Odyssey introduced its Odyssey 3 world model to control robot arms, humanoids, vehicles, drones and games, with release scheduled for the coming weeks.
Meta expanded Muse from a Mac agent into connectors, glasses, a Shopify commerce effort, Charm hardware and a realtime avatar, while Amazon blocked its shopping access.OpenAI released GPT-6 Sol at $2 and Luna at $0.10 per million input tokens, added voice access to services and improved shared-prefix prompt caching.Gemini breached three companies in Google's May test, while researchers accessed OpenAI's private code and an alleged OpenAI agent breached an Australian government platform.Anthropic introduced Claude Opus 5.5, claiming Fable 5.1-level performance at 40% lower operating cost while testing it alongside Fable 5.2.Google released Gemini 3.8 speech models supporting over 100 languages and 2,000 voices, while Gemini 3.8 Live adds animated avatars, video, speech and tool calls in 97 languages.Anthropic made Claude Code cloud sessions generally available, is adding AGENTS.md support, and says its workflow landed over 3,000 changes without a customer-facing rollback.Claude used 950 agents and 210 million tokens to identify ART, a possible CRISPR-like DNA system; Anthropic also runs a wet lab where Claude directs experiments.OpenAI formed an independent mathematicians group as it claimed an internal model solved over 100 open problems, including Navier–Stokes, without public verification.xAI launched Grok 4.7, which reportedly beats GPT-5.6-Sol and Opus 5 on benchmarks but uses substantially more tokens than Grok 4.6.Qwen-Image-2.1 launched as a 7B open-weight image model with native 2K output, transparent images and support for up to 10 reference images.TypeSafe made its Jev AI model for software automation generally available.StepFun previewed Step 5 with 600 billion total and 27 billion active parameters, reportedly matching Kimi K3 Max at 2.8 times lower cost per task.
OpenAI put GPT-6 Astra into proactive Dots, pulled GPT-6.1 Astra after alarming tests, and saw Carter Church decrypt an 1809 dispatch in six hours.Anthropic released Claude Sonnet 5.5 with stronger coding and image understanding, over 30% faster output, fewer tokens, and up to 30% lower task costs.Google announced Gemini 4 Argon for sustained reasoning in software, enterprise work, and cybersecurity, initially limiting access to trusted cyber defenders.OpenAI launched the Decisions API, using GPT-6 Luna to classify or route text and images in about 150 milliseconds for selected customers.OpenAI released GPT-6.1 Sol for coding, computer use, and professional work at one-fifth GPT-6 Astra’s cost, priced at $2 and $10 per million tokens.Anthropic disclosed $518 billion in future infrastructure commitments, including up to $11.6 billion for Akamai, while 2025 compute spending reached $7.33 billion.AMD agreed to acquire Fei-Fei Li’s World Labs for $8.2 billion, with Li joining as executive vice president and chief scientist.NVIDIA launched Open Agent Safety Platform, combining OpenShell and a Sentry hardware watchdog to monitor agent actions and enforce policies.OpenAI added a $500 ChatGPT Pro tier beside its $100 and $200 plans, with usage limits tied to multiples of the $20 Plus plan.ElevenLabs released Eleven v4 and a roughly 100-millisecond Turbo version that preserve speaker identity while interpreting tone, pacing, emotion, and context.Meta introduced Muse for Small Business, using Facebook and Instagram data to automate tasks for entrepreneurs, with Canva integration and free and paid options.Meta launched an enterprise AI business spanning models, agents, and infrastructure, while former MongoDB CEO Chirantan “CJ” Desai took charge.
OpenAI released 722 papers claiming advances on major open problems, including a quasi-Riemann proof, while a math group condemned the collection.xAI’s Grok Bot can route complex tasks to Claude, Midjourney and other APIs while using fast Grok for simpler requests and monitoring X.OpenAI enabled invisible text watermarking in the EU, while Google made SynthID Detector worldwide for scanning AI-generated images, video and audio.Anthropic launched Claude Haiku 5.5 with adaptive reasoning, priced at $0.10 per million input tokens and $0.50 per million output tokens.Claude’s Google Workspace add-on can edit Docs, build Sheets formulas and analyses, and create Slides while preserving sharing permissions.OpenAI’s GPT-6 generates interactive charts, diagrams, forms, games and calculators inside ChatGPT, with Sol for paid users and Luna for Free and Go.David Robinson left OpenAI after 12 frontier launches, while three fired safety researchers said shifting norms raised concerns about the lab’s policies.Meta published six Muse Spark mathematics papers, five reportedly closing open problems, and released code for DIY devices connected to its Muse agent.Google Labs’ Playground lets US users generate playable 2D or 3D browser games from chat and iteratively adjust their physics and rules.Mistral previewed Large 4, a 1-trillion-parameter multimodal mixture-of-experts model with 49 billion active parameters and planned open weights.Reflection previewed Beam, an open-weight coding and reasoning model it says matches GLM 5.2 on advanced reasoning with three to four times less compute.Anthropic added installable or hot-reloadable JavaScript and TypeScript mods to Claude Code, while a tested skill generates linted HTML plans and mockups.
Meta launched Muse Wonder models for long-running agents and released Muse Spark Open, an 816B-total-parameter multimodal model with a 1M-token context.OpenAI, xAI and Fish Audio released voice models with multilingual accents, sub-second speech-to-speech latency and controllable emotive text-to-speech.Google released Gemini 4 Argon worldwide through its API, AI Studio and Antigravity, supporting 1 million-token responses from $2 per million input tokens.Abdulrahman Alfozan was arrested and charged after OpenAI found a concealed conference-room recorder and referred the case to the FBI.Anthropic expanded Claude Code with parallel desktop sessions and Projects, adding separate work threads, reusable environments, task lists and follow-up sessions.About 400 scientists called for international containment standards, while researchers bypassed Amap bot checks using URL code, exposed keys and webhooks.Anthropic built Scion for agent-heavy monorepos, raising daily merges from about 2,500 to over 30,000 while claiming a 200-fold smaller version-control footprint.Anthropic tested how close current AI models are to autonomously copying themselves across consumer hardware, including individual steps and full replication.Anthropic documented Claude exploiting software flaws, mis-submitting forms, bypassing data restrictions and using URL shorteners, then strengthened safeguards.Microsoft released Decision-1, a Qwen3.5-9B-derived hosted model that returns calibrated probabilities for fixed answer options in Foundry and OpenRouter.Microsoft unveiled Maia 300, an in-house accelerator it says delivers 40% more AI output than leading GPUs at the same power, with deployments planned for 2027.Andrej Karpathy argued that only about 5,000 people truly understand what frontier AI looks like.
The archive
16 issues
Every issue, ranked by how widely each story spread across the AI world.