[DVYio Newsletter](https://newsletter.dvy.io/visit/LX5rX8qR) [Humans are heading back around the Moon for the first time in over 50 years](https://newsletter.dvy.io/visit/z7bXdNBe) [Humans are heading back around the Moon for the first time in over 50 years](https://newsletter.dvy.io/visit/z7bXdNBe) NASA's targeting March 6, 2026 for Artemis II — a 10-day trip carrying four astronauts aboard the Space Launch System rocket. They'll spend the first day in Earth orbit, then swing around the far side of the Moon at a distance of 6,500 to 9,500 km above the surface. For context: this will be the furthest humans have travelled into space. Ever. Artemis I flew the same rocket uncrewed in 2022. This time, four humans will be living inside an Orion capsule roughly the size of a minibus. If it all goes well, Artemis III aims to actually land astronauts on the lunar surface by 2028. [Taalas can provide AI models 200× faster than ChatGPT](https://newsletter.dvy.io/visit/Hb4rLNeD) [Taalas can provide AI models 200× faster than ChatGPT](https://newsletter.dvy.io/visit/Hb4rLNeD) They convert AI models directly into custom silicon — purpose-built chips that do one thing and do it absurdly well. The result: 17,000 tokens per second. (For context, ChatGPT typically generates somewhere around 30–80 tokens per second, so that's roughly a 200× speed improvement.) Current AI infrastructure is slow and expensive: room-sized supercomputers, liquid cooling, city-scale data centres, using general-purpose graphics processing units (GPUs). Taalas sidesteps all of that by baking the model into the hardware itself, merging storage and computation into a single specialised chip. Faster, cheaper, and dramatically less power-hungry. You can play with it here: [chatjimmy.ai](https://newsletter.dvy.io/visit/fGhQUKBq) [One nasal spray to fight every cold, flu, and lung infection](https://newsletter.dvy.io/visit/Yp4MjhJJ) [One nasal spray to fight every cold, flu, and lung infection](https://newsletter.dvy.io/visit/Yp4MjhJJ) Stanford researchers have developed a nasal spray vaccine that doesn't work like any vaccine you've had before. Rather than training your immune system to recognise a specific virus (the approach we've used since the 1790s), it keeps white blood cells in your lungs in a permanent state of "amber alert," as the team puts it, so they can respond to basically anything that tries to get in. Colds, flu, bacterial infections, possibly even allergies. The animal trial results are pretty striking: a 100-to-1,000-fold reduction in viral penetration through the lungs. It also worked against two bacterial species. The effect lasted about three months per dose. Still very early days with no human trials yet but the mechanism is genuinely different from how vaccines have worked for over two centuries. That alone makes it worth paying attention to. [Gemini 3.1 Pro now available](https://newsletter.dvy.io/visit/FMZbDBZq) [Gemini 3.1 Pro now available](https://newsletter.dvy.io/visit/FMZbDBZq) Google's latest model, Gemini 3.1 Pro, is now available in preview across the Gemini app and the various API providers. The pitch: better reasoning for complex tasks like data synthesis, multi-step problem-solving, explaining tricky topics. Basically the stuff where a one-line answer isn't enough. My experience so far? It's aggressively sycophantic. Every response opens by telling me how brilliant my question is. Hopefully that gets dialled back before it leaves preview. This Week I Learnt [People can communicate by bouncing radio signals off the trails of meteroids that burn up in the Earth's atmosphere](https://newsletter.dvy.io/visit/HqhHL3SB) [People can communicate by bouncing radio signals off the trails of meteroids that burn up in the Earth's atmosphere](https://newsletter.dvy.io/visit/HqhHL3SB) [— Every day, millions of tiny meteoroids burn up in Earth's atmosphere. Most are unremarkable but some leave behind brief trails of ionised particles in the upper atmosphere, and it turns out you can bounce radio signals off them. It's called "meteor burst communications" (MBC). As a meteoroid disintegrates, it creates a glowing trail of ionised particles in the E layer of the atmosphere. These trails are dense enough to reflect radio waves, typically in the 30–50 MHz range. The reflection only lasts from a fraction of a second to a few seconds, but that's enough to open a brief communication window between two stations up to 2,250 km apart. What‽ The usable frequency depends on how intensely ionised the trail is, which is a function of the meteoroid's size, and the communication range depends on where the meteoroid burns up, at what altitude, the angle it entered the atmosphere, and where the two stations happen to be relative to all of that. You're essentially waiting for a random space rock to create a momentary mirror in the sky at exactly the right spot. The earliest observation of meteors interacting with radio propagation was in 1929, by Japanese physicist Hantaro Nagaoka. So people have known about it for nearly a century, but it's one of those ideas that feels like it shouldn't work. You're using the debris of burning space dust as infrastructure. The tradeoff is obvious: you can't control when meteors arrive, so the communication windows are unpredictable and fleeting. But for certain use cases (low-bandwidth, delay-tolerant, point-to-point links over long distances) it's a genuinely clever approach. Nature provides the reflector; you just have to be ready when it shows up.](https://newsletter.dvy.io/visit/HqhHL3SB) Which application of AI are you most excited about? Science Healthcare Robotics Creativity Energy Transport Last Week's Poll Results for: Which decade of your life has been the best so far? Childhood (0–9) 0% Teens (10–19) 19% Twenties (20–29) 28% Thirties (30–39) 35% Forties (40–49) 1% Fifties (50–59) 0% Sixties or beyond (60+) 0% Hasn't happened yet 17% [Image](https://newsletter.dvy.io/visit/HrhdViLA) [Image](https://newsletter.dvy.io/visit/HrhdViLA) [Image](https://newsletter.dvy.io/visit/HrhdViLA) [Image](https://newsletter.dvy.io/visit/HrhdViLA) [Image](https://newsletter.dvy.io/visit/HrhdViLA) [Image](https://newsletter.dvy.io/visit/HrhdViLA) [Laurids Gallée Lighting](https://newsletter.dvy.io/visit/HrhdViLA) [V&A Museum rebuilt YouTube's 2006 interface](https://newsletter.dvy.io/visit/QtWDjiVA) [V&A Museum rebuilt YouTube's 2006 interface](https://newsletter.dvy.io/visit/QtWDjiVA) The London museum is putting a reconstructed version of the original YouTube watchpage on display — rebuilt from internet archives dating back to December 2006. Alongside it sits "Me at the Zoo," the first video ever uploaded to YouTube, posted by co-founder Jawed Karim in April 2005. That 19-second clip of a guy standing in front of elephants at San Diego Zoo has since racked up nearly 380 million views and 18 million likes. The interface is basically a fossil record of how we learned to navigate video on the internet. Other V&A digital collections include WeChat, Flappy Bird, and the design for the mosquito emoji. [Cheaper protein drugs, courtesy of a language model trained on yeast DNA](https://newsletter.dvy.io/visit/p4fxHyCW) [Cheaper protein drugs, courtesy of a language model trained on yeast DNA](https://newsletter.dvy.io/visit/p4fxHyCW) MIT researchers took a large language model (LLM) — the same type of AI that powers tools like ChatGPT, trained to find patterns in sequences of text — and pointed it at something that isn't text at all: the genetic code of an industrial yeast called Komagataella phaffii. The yeast is already widely used to manufacture protein-based drugs, but it has preferences about which genetic codons (the three-letter DNA instructions that encode amino acids) it works with most efficiently. The LLM learned those preferences by treating the yeast's genome as a kind of language, then predicted optimal DNA sequences for producing specific proteins at higher yields. The practical result: same manufacturing process, better output. If it scales, this could meaningfully cut the cost of developing protein drugs. [Pixar built AI spiders to weave realistic cobwebs for Toy Story 4](https://newsletter.dvy.io/visit/XsNMUPq7) [Pixar built AI spiders to weave realistic cobwebs for Toy Story 4](https://newsletter.dvy.io/visit/XsNMUPq7) Pixar's Toy Story 4 team faced a problem: the Antique Mall needed cobwebs. Lots of them. Manually placing each strand would be tedious and unconvincing. Their solution? Virtual spiders. Technical Director Hosuk Chang built AI agents that weave webs the way real spiders do — you can see the red dots in the video moving around, spinning silk in real-time. The result looks organic because, procedurally speaking, it is organic. For cobwebs that characters interact with, the animators still built those by hand (you need precise control when Woody's brushing through a web). But for ambient set dressing? Let the spiders do what spiders do. [Ground News](https://newsletter.dvy.io/visit/f78QaYjJ) [Ground News](https://newsletter.dvy.io/visit/f78QaYjJ) [— This news site aggregates the same story from dozens of sources and labels each one by political leaning (left, centre, right) so you can see how framing shifts across outlets. It also surfaces "blindspot" stories: ones covered heavily by one side of the political spectrum but largely ignored by the other. The core question it's trying to answer is a good one: not just "what happened?" but "who's telling you about it, and who isn't?" It's a neat way to surface not just bias in how things are reported, but bias in what gets reported at all. Whether or not you use it daily, spending ten minutes with it recalibrates how you read everything else.](https://newsletter.dvy.io/visit/f78QaYjJ) [Image](https://newsletter.dvy.io/visit/vrBeuWsU) [Image](https://newsletter.dvy.io/visit/vrBeuWsU) [Web](https://newsletter.dvy.io/visit/iuHqdkca) · [Twitter](https://newsletter.dvy.io/visit/CRbWSBDT) · [LinkedIn](https://newsletter.dvy.io/visit/zwLScypZ)