We put our model on a satellite. Then we ran it tens of thousands of times. Johannes Waldstein on Tech Command with Alain Gavin, on what it takes to run AI on hardware you can’t reach.

Earlier this year, @NASA selected PiLogic to test Exact AI for spacecraft fault detection. The SBIR Phase I covers three spacecraft subsystems: electrical power, thermal management, and attitude control. The goal is to catch failures that rule-based monitoring misses, including the ones that spread from one system into another. “This selection lets us apply Exact AI to the missions where the margin for error is thinnest,” said @waldsteinj, CEO and co-founder. “Deep space and Earth-independent operations don’t give you the luxury of a ground team troubleshooting in real time.” None of this happens without the people who backed us early. Thank you to @ScoutVentures @seraphim_space @SovereignsCap @flexcapital @foundersnetwork

Up early on the East Coast this morning? That glowing “space jellyfish” was a @SpaceX Falcon 9 headed to the ISS with supplies at 6:33am ET. As the rocket climbed, the air got thinner and its exhaust spread into a huge cloud. Up that high, it caught sunlight while the ground was still dark. The glowing cloud made the bell, and the trail below made the tentacles. People saw it from Florida all the way up to D.C., about 750 miles away. Did you catch it? Drop your photo below 👇 Photos: @KevinHarlanEfct @Peter_Res_USA @moreno_iv

🚀🌍 Happy World Space Week! 🌍🚀 This year’s theme is Rocket Revolution. Here are five of our favorite rockets from 1957 to now, left to right: •⁠ ⁠1957 — R-7 •⁠ ⁠1967 — Saturn V •⁠ ⁠1981 — Space Shuttle •⁠ ⁠2010 — Falcon 9 •⁠ ⁠2023 — Starship 🌌 World Space Week runs October 4–10

There are 3 types of AI (only one is ChatGPT). We build AI for spacecraft and radar. Here's the cheat sheet we wish everyone had: Every AI does the same basic job: take in information and reach a conclusion a person would otherwise have to work out. The three types differ in where their knowledge comes from. 1. Machine learning: 📚 learns from a huge pile of examples. Engineers train machine learning models on large sets of examples. The model finds patterns, then predicts the most likely output for a new input. ChatGPT and Gemini take in your prompt and predict the next word, over and over, until they've written a reply. Netflix and Spotify take in what you've watched, played, and skipped, and predict what you'll want next. It works well when data is plentiful and the stakes are low. Show it something its training data never covered, and it answers with the same confidence whether it's right or wrong. 2. Symbolic AI: 📋 uses a checklist of rules, followed exactly as written. People write the rules and logic by hand, and the system applies them the same way each time. TurboTax walks you through tax law. Google Maps and UPS search possible routes for the shortest one. Because people write every rule ahead of time, the system never goes off-script. You can trace every decision back to the rule behind it, which makes symbolic AI easy to audit. It also means the system is limited by what its authors thought to include. 3. Probabilistic reasoning: ⚖️ weighs the evidence and updates the odds. You start with a model of how a system works: parts, how they connect, and how they fail. Each new reading shifts the odds toward whichever explanation fits it best, and the system keeps updating as readings come in. Your phone's GPS dot blends noisy satellite and sensor signals into a best guess at your location. Xbox matchmaking and Chess.com ratings estimate your skill from a running record of wins and losses. Because it reasons from how the system works, it can handle cases it hasn't seen. But someone has to build that model first, and exact odds take so much computing power that most systems settle for approximations. 🛰️ Bonus: PiLogic's Exact AI Sapce and radar are hard for the first two types of AI. Machine learning depends on large amounts of training data, but some spacecraft failures have only happened a few times. Symbolic AI depends on someone writing a rule for every problem in advance, and no one can predict them all. PiLogic builds on the third type, probabilistic reasoning, and computes exact odds instead of approximating them. It can't hallucinate, and it shows the evidence behind every conclusion. PiLogic has made exact calculation practical for large, complex systems, and uses it in two products: Manifest for spacecraft diagnostics and Resolve for radar tracking. What did we miss?

Having a human in the loop doesn't mean much if the person can't see what the model based its answer on. Otherwise they're just signing off on it. 🔗 spacenews.com/if-ai-cannot-b… via @SpaceNews_Inc

Congratulations to the whole @SpaceX team!

In machine learning, 92 or 95 percent accuracy is often a win. For a government satellite, being wrong 5 or 8 percent of the time isn’t acceptable.

By 2032, the Air Force aims to field 500 autonomous drone wingmen. Defense One reports that reaching that goal will take new doctrine, training, permissions and confidence that the technology behaves predictably. That predictability is foundational. You can't build doctrine or training around a system when you don't know what it will do. A unpredictable system that always sounds certain will do more harm than good. PiLogic is building a different kind of system: one that tells an operator exactly how sure it is and what that confidence is based on so they have something they can act on. 🔗 defenseone.com/technology/20… via @DefenseOne

Not every AI tool is right for every job. The tool has to match the stakes. Chatbots are useful for a lot of work. But an AI report that sounds right is not good enough when the stakes are this high. Exact AI is different. It reasons over the evidence you actually have, scores how likely each reading is, and leaves every step open to inspection. Same inputs, same answer. Every time. For the decisions that really matter, you have to be able to see why the system said what it said. And a human has to own the decision.

Nine years ago today, NASA flew Cassini into Saturn on purpose, ending 13 years of exploring the planet and its moons. 🪐 Here are some of its most iconic images. Images: NASA/JPL/Space Science Institute

Agentic AI is going to play a major role in future space operations. As satellites sense, decide, and act at machine speed, the reasoning underneath those decisions has to be something you can verify. More autonomy makes exact, auditable reasoning all the more important. Read the full article: spacenews.com/agentic-ai-in-… via @SpaceNews_Inc

Getting a new system onboard used to be measured in quarters. We're now measuring it in weeks. In the case of @AFResearchLab, 45 minutes. Exact AI isn't just more accurate than what it replaces. It's faster to get onboard. @waldsteinj with @AlainGavin on Tech Command Investing

September 3, 1976. Viking 2 touches down at Utopia Planitia. Mars was roughly 20 light-minutes away. Nobody on the ground could fly it down. The lander had to detect its own state and act on it, alone, in real time. Designed for 90 days. It ran for more than three years.