Why lasers? Most deep-space communication still relies heavily on radio waves. Laser communication can pack much more data into a narrow beam while using smaller, lighter equipment. The difficult part is hitting something hundreds of thousands of kilometres away with that beam. At lunar distances, tiny pointing errors matter; the signal arriving back at Earth is incredibly faint, and Earth’s atmosphere adds another layer of distortion. For this test, Chinese researchers used high-precision tracking and extremely sensitive single-photon detectors to recover the signal after its journey of more than 400,000 km. The eventual goal is much bigger than faster Moon photos. High-bandwidth links like this could connect lunar spacecraft, astronauts, surface bases and Earth - essentially forming part of a future internet around the Moon.

This isn’t just hypothetical. Researchers are already building some of the pieces a future “dog translator” would need. A University of Michigan team found that an AI model originally trained on human speech could also learn patterns in dog barks. Using recordings from 74 dogs, it could distinguish things including the context of a bark, as well as characteristics such as the dog’s age, sex and breed. That same research group is now running CrowdBark, collecting videos of real dogs so AI can analyse vocalisations, facial expressions and body posture together, rather than treating a bark in isolation. They’re specifically looking for patterns connected to emotions, intentions, needs and even whether dogs use signals to refer to particular objects or situations. Other research is moving in the same direction. A 2026 study used deep learning to recognise canine emotional states from images, with its best model reaching about 84% accuracy on its dataset. Another 2026 study trained an audio model to classify dog emotions from vocalisations and reported very high accuracy on its own test data. And there’s a good reason this may be possible: experiments show that dogs themselves extract meaningful information from other dogs’ vocalisations. Their behaviour changes depending on whether another dog sounds hostile, distressed or playful. But we are nowhere near converting a bark into a sentence like: “Please open the back door.” Emotion recognition is much easier than proving that a particular sound has a specific word-like meaning. Dogs also communicate through smell, posture, gaze, movement, and context, so a real translator would probably need to watch and listen at the same time. There’s another warning too: researchers have found that general-purpose AI models can misread dog emotions and introduce very human biases. A useful system will probably need to be trained specifically on canine behaviour rather than simply asking a normal chatbot what a dog looks like it’s feeling. So the first useful “dog translator” probably won’t give us full conversations. It may simply understand dogs well enough to tell us what they’re feeling, what they need, and what they’re trying to communicate. And even that would be pretty extraordinary.

This isn’t an AI being sent to the Moon in a robot. It’s been trained to make sense of the enormous amount of data we’ve already collected there. NASA and IBM pulled together more than 30 different layers of lunar data, using observations from nine instruments across four missions, including NASA’s Lunar Reconnaissance Orbiter and GRAIL, plus Japan’s Kaguya mission. One of the most useful jobs is searching the Moon’s permanently shadowed craters for signs of water ice. That matters because lunar ice could eventually provide drinking water, oxygen, and even hydrogen/oxygen rocket propellant for a long-term Moon base. The model can also identify and classify craters, study volcanic features and combine observations taken at very different resolutions. On one ice-detection benchmark, it reduced error by up to 22% compared with the model researchers used as a baseline. On a larger-scale crater-mapping task, it performed nearly 19% better while using only half as much training data. The interesting part is that NASA doesn’t want to build a completely new AI for every scientific question. The idea is to create a general model of the Moon that researchers can adapt for different jobs. Basically, decades of lunar observations are becoming something an AI can explore for patterns humans might otherwise miss.

This is much bigger than just getting an AI to answer a difficult maths question. The Navier–Stokes equations describe how fluids move, things like air, water and blood. They’re used in aircraft design, weather forecasting and fluid dynamics. The unsolved question was whether a perfectly smooth three-dimensional fluid described by those equations could eventually develop a mathematical “singularity”: essentially a point where the calculated speed becomes unbounded. OpenAI says its new internal model found a proof that this can happen under the official Millennium Prize formulation. To get there, it used roughly 10,000 coordinating AI agents working in parallel. During the Navier–Stokes effort alone they exchanged around 2.7 million messages and generated roughly 130 billion output tokens. The agents reached the result after about 88 hours. GPT-6 Astra then spent another 17 hours helping formalise and verify the proof in Lean, a system used to check mathematical proofs mechanically. OpenAI says the model that did the main work is still being trained and is significantly more capable than GPT-6 Astra. The important caveat is that this has only just been released. A Lean formalisation is powerful evidence, but mathematicians will still examine the assumptions, construction and whether it fully satisfies the Millennium Prize requirements. OpenAI also says it doesn’t intend to claim the $1 million prize. If the proof survives scrutiny, the bigger story may be less about one equation and more about what happens when thousands of AI researchers can attack the same scientific problem at once. Will AI solve all the problems humans have been unable to solve?