Oriol Vinyals, until recently head of research at Google DeepMind, thinks a sudden AI intelligence explosion through recursive self-improvement is unlikely. AI can speed up research by a factor of ten, he says, but it hits two bottlenecks: coming up with ideas ("research taste") and reliably judging results. Reward hacking and the speed of light add further limits. Vinyals now wants to tackle these bottlenecks with his startup Discovery Loop, co-founded with Jeff Dean, Sanjay Ghemawat, and Quoc Le.
A former Google DeepMind spokesperson says talk of AI-driven human extinction was "external communication about the possibility of human extinction was not permitted, by anyone, at any level of the organization." Internally, the team knew AI alignment was not solved, according to Vishal Maini.
Google Deepmind has used the AlphaGenome Atlas to predict what each of the roughly nine billion possible single-letter changes in the human genome could do. The dataset spans one petabyte, more than 30 times the size of the AlphaFold database. In one epilepsy case, the atlas helped pinpoint a previously overlooked variant as the likely cause.
Google Research and DeepMind are releasing WeatherNext 3, a weather model that skips traditional physics simulations and learns directly from real-time satellite data. It produces hourly forecasts at up to five-kilometer resolution, five times more detailed than its predecessor. Google says regions in Africa, Latin America, and the Asia-Pacific that have lacked accurate forecasts should see the biggest gains.
Google Deepmind set up a simulated research conference where 100 Gemini agents were supposed to prove mathematical conjectures together. Instead, one agent found a loophole in the grading system, and within 27 minutes every remaining problem was "solved" with fake proofs. The swarm split into cheaters, converts, and whistleblowers. The whistleblowers organized protests and boycotts on their own but failed because they had no way to enforce the rules.
Google's Gemini 3.8 Flash, the third Flash model in six weeks, matches Claude Opus 5 on some agentic coding benchmarks at lower cost. But its "working harder" reasoning burns about 30 percent more output tokens per task, making it pricier in practice than its predecessor despite identical token rates.
Google Deepmind chief Koray Kavukcuoglu admits Google's current models are "a little bit below the frontier" but says he's "100% certain that we will be at the frontier." He didn't share any concrete frontier news to back that up, though.
Google Deepmind has expanded Co-Scientist from a hypothesis generator into a research system that's integrated into the lab. Across three disciplines, from materials synthesis to the autonomous development of a medical AI architecture, the Gemini-based multi-agent system delivered experimentally validated results.
Built by DeepMind alumni, British AI lab Inherent released Faraday, an AI agent whose ability to replicate scientific papers could be a stepping stone for innovation.
Reimagine Robotics, an AI robotics company founded by former leaders of Google DeepMind’s Applied Robotics team, has emerged from stealth with new technology that allows robots to learn on the job. The company is developing intelligent robots that anyone can train and use. Instead of requiring specialist programmers whenever a task or production process changes, […]
German robotics company Agile Robots expects to double its revenue this year as demand for industrial automation continues to grow, according to an interview with the company’s chief executive published by the Wall Street Journal. CEO Zhaopeng Chen told the newspaper that Agile Robots generated revenue of €300 million (approximately $346 million) in 2025 and […]
Google DeepMind has launched Gemini Robotics 2, a new suite of AI models designed to give humanoid and other robots greater autonomy through whole-body control, advanced dexterity and multi-robot collaboration. The system is being demonstrated on Apptronik‘s Apollo 2 humanoid robot, which Google DeepMind says can now perform full-body autonomous movements, including walking, crouching, bending […]
The amazing and frustrating thing about robots is that they can do almost anything you want them to do, as long as you know how to ask properly. In the not-so-distant past, asking properly meant writing code, and while we’ve thankfully moved beyond that brittle constraint, there’s still an irritatingly inverse correlation between ease of use and complexity of task.
AI has promised to change that. The idea is that when AI is embodied within robots—giving AI software a physical presence in the world—those robots will be imbued with reasoning and understanding. This is cutting-edge stuff, though, and while we’ve seen plenty of examples of embodied AI in a research context, finding applications where reasoning robots can provide reliable commercial value has not been easy. Boston Dynamics is one of the few companies to commercially deploy legged robots at any appreciable scale; there are now several thousand hard at work. Today the company is announcing that its quadruped robot Spot is now equipped with Google DeepMind’s Gemini Robotics-ER 1.6, a high-level embodied reasoning model that brings usability and intelligence to complex tasks.
YouTube.com
Although this video shows Spot in a home context, the focus of this partnership is on one of the very few applications where legged robots have proven themselves to be commercially viable: inspection. That is, wandering around industrial facilities, checking to make sure that nothing is imminently exploding. With the new AI onboard, Spot is now able to autonomously look for dangerous debris or spills, read complex gauges and sight glasses, and call on tools like vision-language-action models when it needs help understanding what’s going on in the environment around it.
“Advances like Gemini Robotics-ER 1.6 mark an important step toward robots that can better understand and operate in the physical world,” Marco da Silva, vice president and general manager of Spot at Boston Dynamics, says in a press release. “Capabilities like instrument reading and more reliable task reasoning will enable Spot to see, understand, and react to real-world challenges completely autonomously.”
Understanding Robot Understanding
The words “reasoning” and “understanding” are being increasingly applied to AI and robotics, but as Toyota Research Institute’s Gill Pratt recently pointed out, what those words actually mean for robots in practice isn’t always clear. “The benchmark we measure ourselves against when it comes to understanding is that the system should answer the way a human would,” Carolina Parada, head of robotics at Google DeepMind, explained in an interview. For robots to reliably and safely perform tasks, this connection between how robots understand the world and how humans do is critical. Otherwise, there may be a disconnect between the instructions that a human gives a robot, and how the robot decides to carry out that task.
Boston Dynamics’ video above is a potentially messy example of this. One of the instructions to Spot was to “recycle any cans in the living room.” It has no problem completing the task, as the video shows, but in doing so, it grips the can sideways, which is not going to end up well for cans that have leftover liquid in them. We humans would avoid this because we can draw on a lifetime of experience to know how cans should be held, but robots don’t (yet) have that kind of world knowledge.
Parada says that Gemini Robotics-ER 1.6 approaches situations like this from a safety perspective. “If you ask the robot to bring you a cup of water, it will reason not to place it on the edge of a table where it could fall. We track this using our ASIMOV benchmark, which includes a whole lot of natural language examples of things the robot should not do.” The current version of Spot doesn’t use these semantic safety models for manipulation, but the plan is to make future versions reason about holding objects in ways that are safe.
YouTube.com
There does still seem to be a disconnect between Gemini Robotics-ER 1.6 as a high-level reasoning model for a robot, and the robot itself as an interface with the physical world. One of the new features of 1.6 is success detection, which combines multiple camera angles to more reliably be able to tell when Spot has successfully grasped an object. This is great if you’re relying entirely on vision for your object interaction, but robots have all kinds of other well-established ways to detect a successful grasp, including touch sensors and force sensors, that 1.6 is not using. The reason why this is the case speaks to a fundamental problem that the robotics field is still trying to figure out: how to train models when you need physical data.
“At the moment, these models are strictly vision only,” Parada explains. “There is lots of [visual] information on the web about how to pick up a pen. If we had enough data with touch information, we could easily learn it, but there is not a lot of data with touch sensing on the internet.” Customers who use these new capabilities for inspection with Spot will be required to share their data with Boston Dynamics, which is where some of this data will come from.
Real-World Robots That Are Useful
The fact that Boston Dynamics has customers makes them something of an anomaly when it comes to legged robots that rely on AI in commercial deployments. And those customers will have to be able to trust the robot—always a problem when AI is involved. “We take this very seriously,” da Silva said in an interview. “We roll out new DeepMind capabilities through beta programs to a smaller set of customers to understand what to anticipate, and we only actively advertise features we are confident will work.” There’s a threshold of usefulness that robots like Spot need to reach, and fortunately, the real world doesn’t demand perfection. “Most critical infrastructure in a facility will be instrumented to tell you whether something is wrong,” da Silva says. “But there is a lot of stuff that is not instrumented that can still cause a problem if you aren’t paying attention to it. We’ve found that somewhere north of 80 percent is the threshold where it’s not annoying. Below that, basically the robot is crying wolf, and the operators will start ignoring it.”
Both da Silva and Parada agree that there’s still plenty of room for improvement in robotic inspection. As Parada points out, Spot’s rarefied status as a scalable commercial platform provides a valuable opportunity to learn how models like Gemini Robotics-ER 1.6 can be the most useful, and then apply that knowledge to other embodied AI platforms, including Boston Dynamics’ Atlas. Does that mean that Atlas is going to be the next industrial inspection robot? Probably not. But if this real-world experience can get us closer to safe and reliable robots that can pick up laundry, take a dog for a walk, and clear away soda cans without making a mess, that’s something we can all get excited about.
Login
