AI Works

Figure AI ran a public 24/7 livestream of its humanoids sorting packages on a warehouse-style line for more than 34 hours straight. The stream was meant to demonstrate sustained, unscripted operation rather than the short, polished clips that usually dominate humanoid demos — including stumbles, recoveries and slow stretches where the bots queued for parts. Coverage was mixed. Supporters pointed to it as the first long-window evidence of a humanoid keeping up with real industrial throughput; sceptics noted the bots' pace, the controlled environment and Figure's recent $39 B post-money funding round as reasons to read the stream as part marketing, part proof.ZeroHedge coverage of the livestream

Rivian shipped a software update that adds an in-car AI assistant tuned to the vehicle. It can read incoming messages aloud, answer owner-manual questions in natural language, surface trip and charging data, and toggle settings such as climate or drive modes by voice — without forcing the driver to navigate touchscreen menus. Rivian framed the assistant as a step away from generic smartphone assistants embedded into the dashboard, towards an agent that knows the specific vehicle's state. The update rolls out over the air to existing R1S and R1T owners on the latest software branch.Rivian software update

Meta added an Incognito mode to WhatsApp Meta AI conversations that drops chat history and prevents the session being used for personalisation or model improvement. In parallel, Meta widened access to Muse Spark, the most capable model from its Superintelligence Labs unit, moving it from a small preview to a broader rollout inside Meta apps. The two changes line up Meta's consumer AI story: a privacy mode for sensitive conversations, and a stronger underlying model so users have a reason to pick Meta AI over the competing assistants embedded in their phones.Meta announcement

Google rolled out a deeper integration of Gemini into Android, branded Gemini Intelligence, that places multimodal context awareness across the home screen, notifications and apps rather than gating it behind a single assistant surface. In the same announcement, Google introduced Googlebook — its AI-first successor to the Chromebook line, designed around always-available agents and cross-device handoff. The pair of launches positions Android and a refreshed laptop tier as the everyday delivery vehicle for Gemini, set against an iPhone where Apple still depends on third-party model providers for the heavier AI features.Gemini Intelligence on Android

Anthropic split its enterprise push into two vertical bundles. Claude for Legal targets law firms with workflows for contract review, discovery, deposition summaries and case-law research, with a dedicated retention policy intended to meet legal-industry confidentiality expectations. A separate Claude for Small Business package ships ready-made agents — bookkeeping helper, customer-support triage, marketing copy — designed to be deployed without an in-house engineering team. The two bundles continue Anthropic's pattern of going industry-by-industry rather than launching a single horizontal enterprise product, betting that pre-packaged agent recipes will land faster than generic platforms.Claude for Legal post

Anthropic added an Agent View to Claude Code that consolidates several simultaneous agent runs into a single window, replacing the practice of opening a terminal tab per agent. Users can see each agent's current task, intervene mid-run, and route follow-up prompts to a specific worker without context-switching across windows. The change is aimed at the heavy agent workflows that have become common over the past six months — splitting a feature across three or four parallel Claude Code sessions to refactor, test and document at the same time. The new view makes that pattern less chaotic and lowers the barrier for less technical users.Anthropic announcement

OpenAI extended Codex with a mobile workflow that lets developers steer an active coding session on their desktop from a phone — browsing repository files, reviewing diffs and queuing further work while away from their main machine. The interface targets long-running agent loops that previously required staying in front of a laptop to monitor and unblock. The launch follows OpenAI's pattern of making Codex a multi-surface product rather than a single CLI: the agent now spans terminal, IDE, web and a remote-control mobile companion. Heavy coders described the change as a way to keep agent runs productive during travel and meetings.OpenAI release page

Japan Airlines and GMO Internet Group launched a trial in May 2026 to use Unitree G1 humanoid robots alongside ground crews at Tokyo's Haneda Airport, moving baggage and cargo on the tarmac. The 35 kg, 1.32 m G1 has 23 degrees of freedom, 3D LiDAR and a depth camera, and is trained in Nvidia Isaac Sim before deployment. The programme runs through 2028. JAL frames the rollout as a response to a shrinking domestic workforce and record inbound tourism — 42.7 million visitors in 2025 and 7 million in the first two months of 2026. Safety-critical ground tasks remain under human control; the airline plans to expand the rollout if early stages prove out.Japan Airlines press release

Anthropic launched Claude Design, a visual-work product that lets users describe a design brief and receive first-pass layouts, prototypes, slides, and one-pagers, then refine them through conversation, inline comments, direct edits, or custom sliders. The tool is powered by Claude Opus 4.7, runs as a research preview for Claude Pro, Max, Team, and Enterprise subscribers, and can read a team's codebase and design files during onboarding to adopt existing colours, typography, and components. Output can be shared via internal URL, saved as a folder, or exported to Canva, PDF, PPTX, or standalone HTML.Anthropic ↗

Engineer Tony Allevato published a long-form first-person account of living with a consumer AI robot, paired with a companion post on reviving an older unit. The piece covers practical limits, family adjustments and behavioural quirks of consumer-grade home robotics in 2026, with a particular focus on how children interact with a chatty home robot. The post drew 28 comments on Hacker News, where readers raised confirmation-bias concerns about kids learning from a perpetually agreeable robot — a recurring theme as Figure, 1X and Unitree push consumer humanoids. It is a useful primary-source counterweight to vendor demo reels.allevato.me ↗

Microsoft released Agent Framework 1.0 — a production-grade SDK with stable APIs, a long-term support commitment, and built-in support for the Model Context Protocol so agents can use any compliant tool out of the box. A new browser-based DevUI visualises agent execution, tool calls and reasoning steps in real time. More than 90% of the Fortune 500 already use Microsoft 365 Copilot, and the framework gives those organisations a supported route to wire custom agents into Azure AI Foundry alongside the boxed Copilot product. The release lines up with Salesforce Agentforce and OpenAI Frontier as the first generation of agent platforms enterprises can buy on standard support contracts rather than research previews.Microsoft announcement

OpenAI detailed the next phase of its enterprise business: a build-and-manage agent stack called Frontier, with named production deployments at Oracle, State Farm and Uber. The company said enterprise customers now generate more than 40% of its revenue and are on track to match consumer revenue by the end of 2026. The disclosure marks the first time OpenAI publicly anchored its enterprise story to specific Fortune 500 customers running agents in production rather than pilot. It also confirms a pattern across the sector — agent platforms are no longer demoware; they are paying line items inside multi-billion-dollar IT budgets.OpenAI announcement

Developer Alex Ziskind and others demonstrated that a cluster of four M4 Pro Mac Minis connected via Thunderbolt 5 delivers 144GB of unified memory and a 128-core GPU for approximately $6,000 — with electricity costs of roughly $15 per month. The setup is capable of running 70B+ parameter models locally. The trend went viral in early 2026, with developers ordering three to twelve units at a time and depleting Apple store inventory across multiple regions. Illustrated a shift in the hobbyist and small-business AI deployment landscape: purpose-built Mac Mini clusters now offer a credible self-hosted alternative to cloud inference APIs, with no per-token cost and full data privacy.Apple Insider

Salesforce announced general availability of Agentforce 360 — its enterprise AI agent platform — after twelve months of staged releases (Agentforce 1, 2, 2dx, 3) and Salesforce running its own operations as Customer Zero. By late 2024 the company had already booked more than 1,000 paid Agentforce deals and described the line as the fastest-growing product in its history. Agentforce 360 ties agents directly into Salesforce's CRM data graph, adds reasoning, governance and interoperability layers, and turns the existing Salesforce platform into infrastructure for autonomous workflows. The launch sets a benchmark for what shipped, paid-for enterprise agent deployment looks like at Fortune 500 scale.Salesforce announcement