Reindent

Google's Gemini Omni 1.2 Flash lets you edit a video the way you edit a chat. One sentence per turn takes the same clip from a striped tent to an oak forest, a stable, a market fair and a stone castle, and the two characters keep the exact same performance. Clips now run up to 15 seconds, and it holds your character, product and style references from the first frame to the last. Every clip here is Google's own demo. Footage: Google DeepMind and Google AI (@GoogleAI)

Bambu Lab's new H3C prints a shoe with two soft materials in one job. It carries nine PrintTools at once, each with its own filament path and heater, so the next one is already hot while the current one prints. One job can go from rigid plastic to soft TPU, or from one color to the next, in as little as 6 seconds in Bambu's own PLA tests. It even printed a whole watch in one go: a 0.2 mm nozzle for the dial, a 0.4 for the case, and soft TPU for the band. Footage: Bambu Lab (@BambulabGlobal) and Creative Bloq (@CreativeBloq)

This hurricane rips up nearly 4,000 real blocks in a Minecraft-style world, trees and buildings included, says its creator, stink-O. The funnel pulls the ground up into the clouds. When it passes, the plains go quiet again. Footage: stink-O (@Stinkysunstep)

I Just Wanted to Fish, a fishing game made with heavy help from Claude, is out on Steam as a free demo. One developer, Fork Dev. Claude was used in most of the game, and all the models and animations came out of Blender MCP. The game and level design stayed with the dev. Four weeks ago it was a prototype that used 100% of Opus 5.5's weekly usage. Footage: Fork Dev (@ForkedPush)

Anthropic found a Linux kernel bug with Claude, and Claude wrote a working remote root exploit for servers with a multipath TCP listener. CVE-2026-98432. In Linux since 4.4, rated 9.8. No attacks reported. Fixed upstream, but Ubuntu and Red Hat don't have a fixed kernel yet. Patch and reboot when your distro ships it. Until then, turn off MPTCP if you don't use it. Sources: Anthropic, Atlas (atlasum), Daniel Lemire (@lemire)

An Xuan turned 30 days of Claude Code into a running chip factory. It replays real data: 2,656 sessions, 86,131 tool calls and 27B tokens. Every session is a wafer lot, every tool call a process step, every error a bad wafer. Built with Claude Opus 5.5, An Xuan says. Footage: An Xuan (@anxuanng)

Charlie Chapman just added firefighting to a game that turns the whole Earth into an open world. Park the fire truck, raise the ladder, spray water into a burning office building. The streets are real open map data, streamed into Charlie's own C++ engine. Charlie says the game's coding is almost entirely Claude Opus 5.5. Prompted from a phone on lunch breaks, tested in the evenings. Footage: Charlie Chapman (@_chuckyc)

Vercel Labs just released agent-desktop. Coding agents can now use your desktop apps, each with its own cursor. Runs on macOS, Windows and Linux, in the background while you keep working. Here Claude Code, Cursor and fx work Notes, Calculator and Reminders on one Mac at the same time. Agents working side by side on one computer is how we run our own company, in Jauvex Pro. Footage: Chris Tate (@ctatedev), Lahfir (@mdlahfir), Vercel Labs Jauvex is not affiliated with or endorsed by Vercel.

Claude subscriptions are back in OpenCode. Pick Claude in the model menu and it works. Claude uses OpenCode's tools, follows your permission rules, and you can switch models mid-session. Daniel Smolsky from OpenCode says they got Anthropic's permission. Footage: Daniel Smolsky (@DanielSmolsky), OpenCode (@opencode)

NVIDIA just showed humanoid robots running, walking and working inside GPU physics simulation. Robot arms sort parts, humanoids load washing machines and pack boxes, and whole warehouses are simulated down to the forklifts. All computer graphics and simulation, not generative AI. Footage: NVIDIA (@nvidia)

HyperFrames just added computer use to HyperFrames Studio. Your agent can click, type and record in any app on your computer. Ask it to record a flow in your app and it opens the app, records the screen, then rebuilds the frames in HTML. Real screens, not imagined ones. Giving agents real work like this is how we run our own company, in Jauvex Pro. Footage: HyperFrames (@HyperFrames_) Jauvex is not affiliated with or endorsed by HyperFrames.

Binance just paused withdrawals after a security incident. It says user funds remain safe. Chainalysis estimates about $5.9B was stolen, at the time of theft. Tether froze more than 1.567B USDT, and Circle more than 800M USDC. CZ says that as a shareholder, he expects user funds to be fully covered. Statements: @binance, @chainalysis, @paoloardoino, @jerallaire, @cz_binance

You can now book a humanoid robot to clean your home, on a recurring subscription. Tau Robotics just opened it up: up to 3 hours per cleaning. Here it wipes the sink, the counter and the toilet. First take, uncut, at 4x speed. A human supervises each robot and can take over. Selected households, from a waitlist. Footage: Alexander Koch (@alexkoch_ai), Tau Robotics

Herdr just launched a desktop app you change by asking your agent. Every part of it is a file your agent can edit, and it's open source. It runs on a new engine, herdr-core, where agents start and talk to other agents. Close or update the app and your terminals keep running. Agents working with agents is how we run our own company, in Jauvex Pro. Footage: Herdr (@herdrdev) Jauvex is not affiliated with or endorsed by Herdr.

Clicky just learned to draw. Ask about what's on your screen, and it sketches the explanation right beside it. No chat. It turned a GPT lecture into questions and answers, drew how pushing air down lifts an F-35, and labeled the camera cable on a Raspberry Pi. It's out now, and anyone can try it. Footage: Farza (@FarzaTV)

OpenAI just added API credits to eligible ChatGPT Pro and Business plans. Pro gets $100, $200 or $500 a month, by tier. Business seats add $20 or $100 each, pooled up to $500. The first grant is prorated, and credits expire each cycle. You claim them in ChatGPT billing. Footage: OpenAI Developers (@OpenAIDevs)

This mammoth isn't on the desk, and it still reacts when you touch it. Simulon just launched Reality-Aware Rendering: place 3D creatures with your phone and act out the scene. When you stop recording, it renders the take around what you did and said. Bottom: what was really there. Top: the final. Available now on iPhone, Android coming soon. Footage: Divesh Naidoo (@diveshnaidoo), founder of @Simulon

Apple just brought Touch to the Mac, for the first time. On the new MacBook Pro you reach out and move a 3D model with your finger, right on the screen. Scrub a video timeline, play instruments, and it all works with the keyboard and trackpad. Pre-orders are open now. Footage: Apple, via Tim Cook (@tim_cook)

Liquid AI just launched LFM3: six open-weight models around 2B parameters, and the omni one sees, listens and drives apps. Asked in German what a red logo says, it answers: Bic Camera. Asked for a flight, it calls search flights, types Boston and San Francisco, and picks the date, the flight and a seat. We run our own company on a team of AI agents, in Jauvex Pro. Footage: Liquid AI (@liquidai) Jauvex is not affiliated with or endorsed by Liquid AI.

Supabase just launched Compute: your agents and long-running jobs now run right next to your Postgres data, in the same project. Run supabase compute new, pick a runtime and a size, then deploy with supabase compute push. Private alpha for now: evaluation only, us-west-2, with a waitlist. We run our own company on a team of AI agents, in Jauvex Pro. Footage: Supabase (@supabase) Jauvex is not affiliated with or endorsed by Supabase.