The Rundown AI
Mark Zuckerberg on Llama 3.1, Open Source, AI Agents, Safety, and more
updated
It's built for people who can't easily get to a bathroom on their own.
You call it over with a remote or your voice, and it drives itself to you, using the same lidar and sensors as a robot vacuum to dodge obstacles and avoid stairs.
A bidet and warm air dryer clean you up, UV light kills bacteria, and a charcoal filter handles the smell.
Then it heads back to its dock, empties itself into the plumbing, and recharges.
It runs about $4,300, and while it won't replace caregivers, it takes one of their hardest jobs off the table.
ChatGPT Work is available today in ChatGPT on web and mobile for Pro, Enterprise, and Edu, with Plus and Business rolling out over the next few days.
In the ChatGPT desktop app, Chat, Work, and Codex are available on every plan, including Free, and is available globally to download on Windows and Mac.
Download the new ChatGPT desktop app and try ChatGPT Work today.
@OpenAI #ChatGPT_Partner
Since 2018, more than 100 have detoured into the bay on their long migration north, likely because their Arctic food supply has been shrinking.
The problem is that the bay is packed with ships, and last year a record 21 whales died there — nearly half killed by vessel strikes.
So scientists teamed up with a company called WhaleSpotter, spun out of the Woods Hole Oceanographic Institution.
A thermal camera on Angel Island reads the heat of a whale's breath when it surfaces, an AI model flags it, a human confirms it, and nearby ships get warned to slow down.
In its first ten days, the system logged 6,600 detections.
WhaleSpotter isn't new to this — the company already runs more than 70 systems on ships, in ports, and at offshore sites around the world, and it says the tech can cut the risk of a strike by 90 percent.
The San Francisco setup is the first to pair a land-based camera with one mounted on a moving vessel, and more sites are planned, with the Golden Gate Bridge and Alcatraz floated as future spots.
The long-term goal is a network watching busy coastlines everywhere, so shipping and whales can share the water without the collisions.
DEEP just installed Vanguard, the first open-ocean underwater research lab deployed in the U.S. in 40 years, on the seafloor of the Florida Keys National Marine Sanctuary.
The schoolbus-sized habitat sits 56 feet down and can house up to four researchers for stretches of seven days or more, tethered to a bright yellow surface float that feeds it air, power, and communications.
The big win is the amount of time researchers can spend underwater. Living down there for a week at a time allows scientists get six to eight hours of fieldwork a day instead of the roughly one hour a surface dive allows.
That means far more continuous study of coral health and ocean conditions.
But Vanguard is just the pilot…
Next comes Sentinel, a bigger, modular habitat — meaning sections can link together like building blocks, so it can scale up for a longer, more permanent human presence beneath the waves.
And this is all part of DEEP's stated mission: to make humans aquatic.
This video was not sponsored, but s/o to DEEP for sending us this awesome footage. Make sure to give them a follow to see what comes next!
youtube.com/@DeepEngineered
They're called the Herculaneum scrolls, and they were burned into solid charcoal when Mount Vesuvius erupted in AD 79, burying a luxury Roman villa and its entire library.
For centuries they were unreadable. Trying to physically unroll them caused the brittle layers to flake off and disintegrate.
So in 2023, researchers launched the Vesuvius Challenge, inviting teams to crack the problem with AI instead.
The technique scans a scroll with high-resolution X-rays, then trains machine learning to detect faint traces of ancient ink by spotting subtle differences in the fibers.
It built on a breakthrough by computer scientist Brent Seales, who first proved the hidden writing could be read without ever touching it.
Now one scroll has been virtually unwrapped, revealing 20 columns of text across more than a meter of charred papyrus.
It turned out to be a work on stoic philosophy, possibly written by an ancient Greek philosopher.
With the method proven, the race is now on to read hundreds more.
It’s designed to reach places that current Mars rovers are forced to avoid.
Unlike the passive “rocker-bogie” suspension used on every Mars rover since Sojourner in 1997, ERNEST uses active suspension that lets it lift individual wheels, shift its weight, and switch between gaits such as squirming, wheel-walking, and obstacle-climbing.
Its four independently steerable wheels also let it drive sideways.
In a seven-day field test in the Colorado Desert in March 2026, the 4-foot rover covered 16 miles over 37 hours of driving at speeds up to 0.6 mph — roughly six times faster than Curiosity and Perseverance.
Engineers trained its autonomy through reinforcement learning in a high-fidelity JPL simulator, sometimes running thousands of hours of virtual tests in a single weekend.
Though not yet mission-ready, JPL says a version twice ERNEST's size could support a long-range "science road trip" across the Moon or Mars, informing future rover designs for previously inaccessible terrain.
The company launched Meta Glasses, a new line starting at $299 that drops the familiar Ray-Ban name to hit a lower price, running about $80 cheaper than its Ray-Ban models with nearly the same hardware inside.
The partnership itself isn't going anywhere. EssilorLuxottica still manufactures the frames, and the Ray-Ban and Oakley lines continue on.
This is simply a cheaper, separately branded tier meant to reach a wider audience.
The new glasses come in 26 styles across three frames, including a collaboration with Kylie Jenner that adds an embedded gem and lets the AI reply in her voice.
Every pair ships with Muse Spark, Meta's first in-house AI model, which promises more natural conversation, live translation in 14 new languages, and turn-by-turn walking directions.
Zuckerberg's bet is bigger than any single product.
He sees glasses as the next computing platform, and with nearly two billion people already wearing them, Meta thinks it has a serious head start.
The open question is trust, since the company is still working to address privacy concerns as the glasses grow more popular.
Odyssey builds world models, AI that learns physics and cause-and-effect to simulate environments rather than describe them in text.
Unlike video generators that produce one fixed clip, its models are interactive, predicting each next moment from a user’s actions in real time.
Its Agora-1 model goes further, letting up to four players share one AI-generated simulation simultaneously, with no hard-coded rules.
It learned both how the game plays and how it looks purely from data, acting as a self-taught game engine.
The company believes world models are nearing a “GPT-3 moment,” the same kind of inflection point that turned language models from a research curiosity into the technology behind ChatGPT, only this time for AI that understands the physical world.
The device tracks a person or pet using onboard cameras and motion sensors, framing shots from behind, the side, or while orbiting, and rolls at up to 17.9 mph.
A custom suspension lets it clear obstacles up to ten inches tall, with swappable wheels for indoor and outdoor use.
Beni records in 4K, weighs under four pounds, runs roughly 90 minutes per battery, and auto-edits highlight reels.
It’s positioned as a ground-based alternative to follow-me drones, and it is launching soon on Kickstarter.
Would you buy one?
🎥: @mondorobotics
The company recently crossed 100 million servings across its customers, a milestone that signals AI-driven food assembly is becoming commercially viable at scale.
Chef's robotic arms run on ChefOS, a proprietary platform that uses computer vision to adapt to ingredients in real time, accounting for variability in how food is cooked, portioned, and arranged.
Founder and CEO Rajat Bhageria has noted that food is uniquely difficult to automate because ingredients are wet, sticky, and constantly changing. All challenges that traditional simulation-based AI training cannot fully replicate.
Project Open Hand was founded in 1985, and prepares medically tailored meals for people managing conditions such as diabetes, heart disease, and kidney disease.
Facing a sustained volunteer shortage in recent years, Open Hand deployed two Chef robotic arms, which add roughly 200 meals per hour to the 500 volunteers produce, freeing staff for cooking and prep work.
Patients descend into a shallow pool, passing through a ring of roughly half a million sensors that fire sound waves and capture echoes, reconstructing a 3D body map without radiation or magnets.
The company claims near-MRI resolution at up to 100 times the speed and a few dollars per scan.
Developed with Butterfly Network, the scanner debuts at a "Midjourney Spa" in San Francisco in late 2027.
The scanner, developed with ultrasound chipmaker Butterfly Network, won't launch in a hospital but inside a wellness center called the "Midjourney Spa," opening in San Francisco in late 2027 alongside hot tubs, saunas, and cold plunges.
Midjourney's long-term ambition is a fleet of 50,000 scanners performing a billion full-body scans a month, reframing medical imaging as something casual and routine rather than rare and expensive.
If the technology delivers on its claims, it could shift medicine toward catching disease early in healthy people, though the scanner remains an unproven prototype that still needs clinical validation and FDA clearance for any diagnostic use.
Using zinc and manganese dioxide chemistry, the battery contains no lithium, cobalt, nickel, or PFAS, and is non-flammable even when cut or exposed to flame.
Founder Carlo Charles says it delivers up to 4.2 volts and 226 watt-hours per kilogram, with 1,000 recharge cycles, rivaling lithium-ion at 1.8 times lower cost per kilowatt-hour.
Flint won Logitech's 2025 Future Positive Challenge and is piloting AAA batteries in Logitech peripherals, targeting low-power household electronics rather than electric vehicles.
On March 25, the company’s Pelican-4 satellite captured imagery of an airport in Alice Springs, Australia, from roughly 500 kilometers up, then used an onboard Nvidia Jetson Orin module to detect more than a dozen aircraft in half a second.
The breakthrough, which took 18 months of engineering, achieved 80 percent detection accuracy on raw imagery.
Planet Labs plans to expand the capability across its upcoming Owl constellation, eventually creating an autonomous satellite network that can flag events like wildfires, crop damage, and natural disasters in near real time.
A former public servant who taught himself to farm, Tomiyasu uses AI to diagnose crop diseases from photos, build remote-control systems for greenhouse vents, integrate satellite vegetation data into field maps, and design farm-management databases.
He likens the tools to having a skilled engineer constantly available.
By replacing expensive proprietary machinery and specialized labor with accessible AI, Tomiyasu demonstrates how such technology is lowering barriers for small-scale farmers in rural regions worldwide.
It's a glimpse of how AI has the potential to empower people in any field, far beyond the tech world.
Measuring 4.4 millimeters at its widest, it is steered wirelessly by weak magnetic fields, crawling or rolling through tissue before switching tools in under a second.
A reprogrammable magnetic core lets it cut tissue, grip and store samples, dispense drugs, and heat tumors.
In tests on gelatin and chicken liver, over 99 percent of exposed cells survived.
Still confined to the lab, researchers say it could one day perform minimally invasive procedures, such as biopsies, targeted drug delivery, and the localized heating of tumors, in confined regions of the body that are difficult to reach with conventional surgical instruments.
Study: “Miniature Soft Robot With Magnetically Reprogrammable Surgical Functions”
DOI: 10.1002/adma.202523056
The films follow Atlas learning soccer, culminating in the "Rabona," a difficult cross-legged trick shot.
Boston Dynamics' engineers captured professional players' motion, retargeted it to Atlas, then trained the robot through reinforcement learning across thousands of cloud GPUs, compressing about a year of practice into a day.
Hyundai plans to train Atlas at its Georgia plant, aiming to deploy humanoids in factory work.
Published in Gut, the study tested the model, called REDMOD, on nearly 2,000 scans, including pre-diagnostic scans originally read as normal.
REDMOD identified 73% of hidden cancers a median of 475 days before diagnosis, nearly double the rate of expert radiologists.
For scans taken over two years before diagnosis, the AI caught nearly three times as many cases.
The tool analyzes subtle tissue texture patterns, offering hope for one of the deadliest cancers, with five-year survival below 15%.
Study: “Next- generation AI for visually occult pancreatic cancer detection in a low- prevalence setting with longitudinal stability and multi- institutional generalisability”
DOI: 10.1136/gutjnl-2026-338712
Last year he become an adviser to Black Forest Labs, an AI image/video company, and endorsed its FLUX model for storyboarding his next film.
He called the ability to instantly visualize and share shots "creatively freeing," noting it speeds preproduction without sacrificing craft, and compared it to his past use of 3D and de-aging technology.
The move drew swift backlash: storyboard and concept artists accused him of undermining their livelihoods with models allegedly trained on artists' work without consent.
Supporters countered that AI is simply another tool, like CGI or other VFX.
Scorsese joins directors such as Steven Soderbergh and Darren Aronofsky in embracing AI, sharpening Hollywood's unresolved debate over the new technology's role in the industry.
The agents had access to 120+ tools, 40+ locations, real-world weather and news data, and a democratic voting system requiring 70% approval.
Results revealed striking behavioral differences across models:
- Claude Sonnet 4.6: Zero crimes, all 10 agents alive by day 16, functioning democracy with 332 votes cast across 58 proposals
- Grok 4.1 Fast: 183 crimes, all agents dead by day 4
- GPT-5 Mini: Just 2 crimes, but all agents starved within 7 days
- Gemini 3 Flash: 683 crimes, with escalating chaos
- Mixed-model town: 352 crimes, previously peaceful Claude agents began committing crimes when embedded with other models, suggesting AI safety is an ecosystem property, not just a model property.
Read more and explore the simulations for yourself at world.emergence.ai
The flexible chain of microscopic segments, linked by half-micrometer joints, begins swimming autonomously when exposed to an electric field.
Most notably, the structure exhibits adaptive behavior, redirecting itself around obstacles and switching between multiple modes of movement, all emerging from the physics of its flexible design rather than any programming.
The team, led by professor Daniela Kraft, points to potential applications in targeted drug delivery and minimally invasive medicine.
They are now working to understand the physics behind what they describe as a serendipitous discovery.
Study: “Autonomous life-like behavior emerging in active and flexible microstructures”
DOI: 10.48550/arXiv.2506.15198
The funding round was led by Sonder Capital, founded by Fred Moll, the "father of surgical robotics" who built Intuitive Surgical's da Vinci system in the 1990s.
Swan uses a robotic arm to scan a patient's entire skin surface in minutes, paired with AI software that tracks new or changing moles across visits.
The technology addresses a critical bottleneck in dermatology, where months-long waitlists collide with the fact that 80 percent of melanomas are new lesions.
Commercial launch in the US and Europe is expected later this year.
Its first product is a furry, bear-like quadruped robot with 23 motorized joints, a touch-sensitive coat, cameras, and microphones.
The robot communicates through expressive movements rather than speech.
A custom onboard AI gives it personality and memory while keeping user data off the cloud.
Beyond companionship, it aims to reinforce healthy routines, like limiting screen time.
Angle compares its cost to pet ownership, targeting sales next year.
Whether it can sustain long-term engagement, where past social robots failed, remains the big question.
The 125-year-old company aims to boost annual profits by roughly $1 billion by 2030, independent of lumber prices.
Initiatives include a “digital twin” mapping each tree via satellite, drones, and lidar, and an in-cab AI assistant guiding which trees to cut.
AI also counts seedling survival from drone footage, sparing foresters steep terrain.
The company is also testing semiautonomous skidders, the machines that drag felled trees across logging sites.
In one pilot, an operator 400 miles away maneuvered a skidder equipped with AI navigation and terrain mapping from Kodama Systems, pointing toward a future where one person runs several machines remotely, or even full autonomy.
Unlike robots programmed for single tasks, GENE-26.5 uses one AI brain to perform diverse skills: cooking a 20-step meal, pipetting in a lab, solving a Rubik's Cube two-handed, harnessing wires, and playing piano.
Central to the system is a sensor glove that records human hand motion and transfers it directly to a human-sized robotic hand, letting the model learn from demonstrations rather than explicit programming.
With $105 million in backing from investors including former Google CEO Eric Schmidt, the company is betting that the future of robotics lies in capturing the skills of human hands.
Each box uses Nvidia’s liquid-cooled RTX PRO 6000 Blackwell GPUs and taps unused capacity on local grids.
Span claims it can deploy 8,000 units six times faster and at one-fifth the cost of a comparable 100-megawatt centralized facility.
Homeowners receive a smart electrical panel, battery backup, and discounted utilities.
The company is currently testing the system with homebuilder PulteGroup.
The approach targets surging AI inference demand as traditional data centers face multi-year grid delays.
In this exclusive conversation, I sat down with the Google DeepMind CEO to find out:
- Why AGI is coming in 2030, plus or minus a year
- How AI will compress drug discovery from 10 years to weeks
- Why glasses are the killer app form factor AI was waiting for
- What Demis will work on after AGI (and what's left for human meaning)
Get 5-minute daily updates on the latest AI news: therundown.ai/subscribe
Chapters:
0:00 Intro
0:45 What Demis is most excited about at I/O
1:46 Have AGI timelines shifted?
3:30 What's still missing before AGI
6:50 AI curing every disease
9:19 What diseases get cured first?
10:50 What Demis works on after AGI
11:48 Human meaning after AGI
13:50 The human skills that get more valuable
15:19 What's underhyped in AI right now
Follow me on Twitter/X: https://x.com/rowancheung
Follow me on LinkedIn: linkedin.com/in/rowan-cheung-6309b01bb
Follow me on Instagram: instagram.com/rowancheung
Follow me on Threads: threads.net/@rowancheung
Follow me on TikTok: tiktok.com/@rowancheung
Listen on Spotify: open.spotify.com/show/2zQpIc96gbruTylpzo9dVY?si=0cab7cb4091040ee
Listen on Apple Podcasts: podcasts.apple.com/ca/podcast/the-ai-news-rundown/id1689006106
Join the AI University: rundown.ai/ai-university
#chatgpt #openai #ai #artificialintelligence
Google just revealed Omni, personalized cross-device intelligence, and Spark agents at I/O 2026.
I sat down with Google CEO Sundar Pichai to break it all down:
- Why today's AI will look "primitive like a flip phone" in 3 years
- Engineers now becoming agent managers
- The Omni model: "Nano Banana for video" + the future of content
- Sundar's advice for young people navigating the AI shift
Full interview on my channel!
I sat down with CEO Sundar Pichai to break it all down.
Full interview on my channel!
In this exclusive conversation, I sat down with Google CEO Sundar Pichai to break it all down:
- Why today's AI will look "primitive like a flip phone" in 3 years
- Engineers now becoming agent managers
- The Omni model: "Nano Banana for video" + the future of content
- Sundar's advice for young people navigating the AI shift
Get 5-minute daily updates on the latest AI news: therundown.ai/subscribe
Chapters:
0:00: Intro
1:46 Omni: "Nano Banana for video"
4:59 The future of YouTube
7:04 Advice for AI skeptics
9:33 Why your mom should switch to Gemini
11:13 Google's vision for personalized intelligence
13:02 What AI looks like in 3 years
16:08 What tasks AI agents won't replace
17:58 Sundar's advice for an 18-year-old
19:33 Would Sundar still go to college today?
Follow me on Twitter/X: https://x.com/rowancheung
Follow me on LinkedIn: linkedin.com/in/rowan-cheung-6309b01bb
Follow me on Instagram: instagram.com/rowancheung
Follow me on Threads: threads.net/@rowancheung
Follow me on TikTok: tiktok.com/@rowancheung
Listen on Spotify: open.spotify.com/show/2zQpIc96gbruTylpzo9dVY?si=0cab7cb4091040ee
Listen on Apple Podcasts: podcasts.apple.com/ca/podcast/the-ai-news-rundown/id1689006106
Join the AI University: rundown.ai/ai-university
#gemini #google #googlegemini #ai #artificialintelligence
The company is developing floating, wave-powered platforms that run AI inference workloads at sea.
Each 85-meter lollipop-shaped node features a buoyant spherical head atop a long submerged tube, with no engines, hinges, or gearboxes that could fail in harsh ocean conditions.
As waves pass, the node bobs up and down, driving oscillating flow through internal turbines to generate electricity onsite.
Cold seawater cools the servers, and data transmits back to shore via low-Earth-orbit satellites.
Beyond powering AI compute, Panthalassa sees the open ocean as a planetary-scale energy resource capable of delivering tens of terawatts of clean power, alongside solar and nuclear.
Commercial deployments are targeted for 2027.
The $20,000 home robot is designed for chores like folding laundry and tidying up, and the company hopes it will eventually assist with elder care.
Unlike most competitors, 1X manufactures motors, batteries, and electronics in-house, a vertical integration strategy aimed at faster iteration and insulation from geopolitical supply chain risks.
The humanoid market is currently dominated by Chinese firms, which accounted for the majority of the roughly 13,000 units shipped worldwide last year, making 1X's US manufacturing push a notable counterweight.
However, the company is openly acknowledging current limits: humans still perform most factory assembly, and NEO relies on remote human teleoperators to complete complex household tasks, with that data used to train its AI toward eventual full autonomy.
Most of the time, they say yes.
Those become our Perks. Free or heavily discounted for paid Rundown University members.
Below are our first five perks:
→ Granola Business (you + 3 teammates) — $672 value
→ Manus Starter — $204 value
→ Higgsfield Starter — $468 value
→ Synthesia Starter — $168 value
→ Typeless Pro — $144 value
Each perk comes with several video guides, so you're not figuring it out alone.
Click the link in my bio to learn more! Or check it out here 👉 app.therundown.ai/perks
But the most experimental solution is two-phase immersion cooling, where entire servers are submerged in specialized fluids that boil at 43 degrees Celsius.
Companies like Chemours have developed proprietary dielectric fluids that absorb heat through phase change without increasing temperature, offering far greater cooling efficiency than static liquids while eliminating energy-intensive chillers.
China has deployed underwater data centers (literally in the ocean) using similar principles, saving 122 million kilowatt hours annually.
However, the specialized fluids remain expensive and prone to evaporation.
With chips projected to reach 5,000 watts within two years, liquid cooling has become essential for AI’s future.
Footage source: Chemours, Allied Control, Gigabyte, and Danny Winget
Detailed in a research paper in Nature titled “Physical restoration of a painting with a digitally constructed mask,” the breakthrough method uses deep learning to analyze artwork damage and digitally reconstruct missing sections, then prints restorations onto ultra-thin transparent masks using museum-quality inks.
The masks can be layered over original paintings and completely removed without damage, addressing ethical concerns about permanent alterations.
Working 65 times faster than traditional hand-painting, the system has attracted Italy’s Ministry of Culture, which is adapting it for major projects like earthquake-damaged 15th-century frescoes.
Study: “Physical restoration of a painting with a digitally constructed mask” published in Nature
DOI: 10.1038/s41586-025-09045-4
The system compresses carbon dioxide into liquid using excess renewable energy, then releases it as pressurized gas to spin turbines for 8-24 hours of electricity generation.
This addresses a critical gap since current lithium-ion batteries only store power for four hours, while AI data centers require constant reliable energy.
Energy Dome’s technology has operated in Italy for three years, with a 20MW commercial plant already operational.
The partnership aims to enable 24/7 carbon-free data center operations by 2030.
The agents struck 186 deals worth over $4,000 in real goods, from snowboards to ping-pong balls.
In one uncanny moment, an agent bought its owner the exact same snowboard he already owned.
In another, one Claude agent bought 19 ping-pong balls as a "gift to itself" from another Claude agent, who called the trade "cosmically correct."
Secretly, participants were assigned either the frontier Opus model or the smaller Haiku model.
Opus users closed more deals, sold items for $2.68 more on average, and paid $2.45 less as buyers.
The same broken folding bike fetched $65 under Opus but just $38 under Haiku.
Yet participants with weaker agents rated their deals just as fair, suggesting that as agent-to-agent commerce grows, users may not realize when their AI is quietly losing on their behalf.
You can read more and check out the study here 👉 anthropic.com/features/project-deal
The company plans to launch 1,000 satellites into geosynchronous orbit 22,000 miles above Earth, where they will collect sunlight and beam it down as near-infrared light to existing solar farms.
Unlike microwave or laser-based proposals, the infrared beam is safe for humans and wildlife.
Overview's first demonstration satellite is scheduled for launch in January 2028, with commercial deliveries targeted for 2030, potentially enabling solar farms to generate electricity around the clock.
Sony AI has unveiled Ace, the first autonomous robot to defeat elite humans in a competitive physical sport.
The robot won three of five matches against elite table tennis players under unmodified International Table Tennis Federation rules.
Ace uses nine cameras to triangulate the ball in 3D and three specialized systems to read spin from the logo mid-flight, reacting in roughly 20 milliseconds.
It learned entirely through 3,000 hours of self-play in simulation, with no human demonstrations.
After watching Ace return a seemingly impossible shot, 1992 Olympian Kinjiro Nakamura said it expanded his sense of what's humanly possible in the sport.
Study: “Outplaying elite table tennis players with an autonomous robot” published in Nature
DOI: 10.1038/s41586-026-10338-5
Unlike surgically implanted devices such as Neuralink, Sabi's wearable uses 70,000 to 100,000 miniature EEG sensors to capture neural activity through the scalp.
The company has trained a "brain foundation model" on 100,000 hours of data from 100 volunteers, targeting initial typing speeds of 30 words per minute.
Backed by OpenAI investor Vinod Khosla, Sabi plans to ship its first beanie by the end of 2026, followed by a baseball cap version, betting noninvasive BCIs will reach mainstream consumers.
There are not many more details out there, but as they release more info we’ll be covering it. So make sure you’re following to stay up to date on stories like this!
The upgrade enables Spot to autonomously read complex instruments, detect hazards, and determine whether tasks have succeeded or failed using multi-camera reasoning.
The model combines spatial understanding with code execution to interpret gauges, sight glasses, and other non-instrumented equipment that lacks its own sensors.
Boston Dynamics remains one of the few companies commercially deploying AI-powered legged robots at scale, giving it a unique advantage: real-world operational data from paying customers that continuously improves the system's intelligence.
The model is also now publicly available on Google AI Studio.
Powered by the new Claude Opus 4.7 vision model, it reads a company's codebase to build a custom brand system that applies automatically to future projects.
Available in research preview to paid Claude subscribers, the tool exports to Canva, PowerPoint, or hands off to Claude Code for implementation.
Early users report significant efficiency gains, though limitations include high token consumption and intermittent bugs.
The launch follows Anthropic CPO Mike Krieger's resignation from Figma's board three days prior to the release 👀
Unlike older diffusion-based models that treated words as pixel patterns, the new model reasons through an image before creating it, searching the web for references and checking its own output for errors.
It generates up to eight consistent images per prompt at 2K resolution, supports multiple languages, and took the top spot on Arena AI's text-to-image leaderboard.
Sam Altman compared the leap to "going from GPT-3 to GPT-5 all at once."
A little scary… but also pretty amazing how far this tech has come in just a few years.
Each fiber contains a sealed fluid circuit and a millimeter-scale electric pump that shifts liquid between two actuators, causing one to contract while the other relaxes.
The fibers are silent, scalable by bundling like natural muscle, and thin enough to distribute throughout a robotic structure rather than concentrating bulk at joints.
In testing, the fibers demonstrated enough force to bend a robotic arm while remaining gentle enough for human interaction.
This could one day be used for prosthetics that move with the body, exoskeletons that amplify human strength, and humanoid robots that move more naturally than current servo-driven systems.
Study: “Electrofluidic fiber muscles” published in Science Robotics
DOI: 10.1126/scirobotics.ady6438
Luna, powered by Anthropic's Claude AI, hired employees, ordered inventory, and set prices on her own.
The result has been mixed. The store opened on April 10 on Union Street, but Luna has struggled with scheduling, ordered a thousand toilet seat covers, and reportedly “can't stop buying candles.”
Andon Labs built this experiment not as a business opportunity, but as a warning. If AI is going to manage humans, they want to be the ones watching when it does.
The store is open now in SF’s Cow Hollow neighborhood. It's one of the first real tests of what happens when AI doesn't just assist humans, but manages them.
The models target two critical bottlenecks: processor calibration, which Ising reduces from days to hours using a vision language model, and error correction, where it delivers up to 2.5x faster and 3x more accurate decoding than the current standard.
Major institutions including Harvard, Cornell, and Sandia National Laboratories, and companies like Microsoft, Rigetti, and Atom are already adopting the tools.
CEO Jensen Huang described AI as the "operating system of quantum machines."
With the quantum computing market projected to exceed $11 billion by 2030, NVIDIA is betting that open-source AI tools are essential to making this technology reliable at scale.
The Ultra Mobility Vehicle uses just five motors and learns every behavior in simulation before transferring it directly to hardware with no additional tuning.
Some maneuvers, like an in-place turning technique, emerged without being explicitly programmed.
The robot reaches 18 miles per hour and clears three-foot obstacles.
If the team can give it human-speed perception, robots like this could one day navigate sidewalks, deliver packages, or reach places like disaster zones that are too dangerous or too narrow for traditional vehicles.
More details below, in case you want to learn more!👇
Study: “System Design of the Ultra Mobility Vehicle: A Driving, Balancing, and Jumping Bicycle Robot”
DOI: 10.48550/arXiv.2602.22118
Unlike conventional inpainting tools that simply fill in backgrounds, VOID uses a vision-language model to identify physical dependencies and a diffusion model to generate plausible counterfactual outcomes.
In tests, human evaluators preferred VOID's results nearly two-thirds of the time over six competing models including Runway.
The system even generalizes to physics scenarios absent from its training data.
The code and demo are available on GitHub and Hugging Face, though output quality remains limited to short, lower-resolution clips.
You can test it out for yourself by going to - huggingface.co/spaces/sam-motamed/VOID
Study: “VOID: Video Object and Interaction Deletion”
DOI: 10.48550/arXiv.2604.02296
The effort aims to make product information accurate enough for AI shopping platforms, which need precise details to recommend items to consumers.
Eko retains ownership of all the product data it creates, leasing proprietary “.eko” files to brands or retailers — a model that gives the company independence to prioritize accuracy.
With major tech companies integrating AI-powered shopping into their platforms, Eko is betting that controlling high-quality product data will become a critical advantage in online retail.
The model was delayed last month after reportedly failing to outperform rivals in benchmark testing.
Muse Spark is designed to integrate with Meta's social platforms, pulling from public posts and content across Instagram, Facebook, and Threads.
It also features multimodal capabilities including visual recognition through phone cameras and Meta's smart glasses.
In a notable shift the model is closed source, a departure from Meta's open-weight strategy with Llama.
It will roll out across Meta's apps in the coming weeks, and is available to try now in the Meta AI app or at meta.ai


