无忧传媒

听I'm Joanna Guy, lead technologist at 无忧传媒 Hamilton, and I'm here at the Helix Center for Innovation to talk about 无忧传媒's A2E2, AI for Edge Environments, and I'm here with Zane Price, lead technologist on A2E2 development. Can you tell me more about A2E2? Yeah, absolutely. A2E2 is AI for edge environments, and what we've done is we've proven the capability to take AI large language models to space. And have shown that we're able to use them on the space station and we've developed this capability that has no reliance on outside connectivity so these language models can actually be used without the Internet at the edge to provide information where we might be relying on analog information or paper information and so we've really proven that we can take this to space, we can take this anywhere and that we can actually have access to these technologies that are transforming life here on Earth. Wow, so how is this different from what astronauts were using before to access information in space? Yeah, well, astronauts definitely do not have access to AI or language models. What they've been using is analog information, right? That might be binders of technical documents or information, medical information, whatever they might need on the space station. As you can imagine, the problem is that we don't have a lot of space there, no pun intended. And so this type of technology could be really transformative for information storage and information retrieval and then intelligent summarization of of queries and information that an astronaut might need in space. This technology is obviously clearly helpful to the human. How might it have implications in terms of large language models or even agentech AI for autonomous capabilities like satellites and maybe even robots on the moon, which I think is the only Thing that inhabits the moon and Mars right now. Sure, yeah, that's a great question. Um, what this does, I think it's very important in in the consequences that we've shown that AI can operate where there is no connectivity where we don't have access to a remote server. And so as you mentioned, if we're thinking about satellites in orbit, we might be able to implement language models as agents. Without that necessary connectivity that we currently see these models being constrained by, we might be able to deploy these language models as agents to satellites, but they can actually make agentic and autonomous decisions without access to the internet. Well, I'd love to try A2E2, but before we dive into that, what, uh, large language model was this trained on? Yeah, so we're actually empowering A22 with what we're calling space llama. That was a llama model that we have trained in collaboration with Meta to be more knowledgeable and more truthful about information in space. Awesome. Well, can I give it a try? Absolutely. And, uh, may I just ask it a voice to text question perhaps as if I am an astronaut? All right. How should I torque the bolts on my space born computer? 0 test access controls. Wow, to torque the bolts on your spaceboard computer too, use a high torque locker tool with a 1/4 inch socket and a torque wrench 1/4 inch drive. Tighten the fasteners to 30 pounds as specified, and it looks like I also get a number of reference documents. That's right. So that's a feature that we've intentionally built in here for tractability, for traceability, so that the operator can trust the information that's being provided. For some reason they say, hey, that just doesn't sound right. Perhaps they do have an idea of the answer. And they want to verify, they can go ahead and look at the sources that Space llama has based its answers on. That's fantastic. And uh what happens if I ask it, uh, an off the wall question? Uh, for example, Space llama, what are the origins of the universe? Uh, and we could see that it comes back and it says I don't actually have information because we are basing the answers on the information that's been uploaded to the model, right? And so in this case it's based on technical manuals in order to repair the space born computer. So this is a very important feature so that if we're asking information it does not have access to that it doesn't respond and hallucinate. So that is a big challenge with language models is that they can provide information that sounds convincing. But actually is not factual. And in order to make updates to this information on HPE's Space born computer on the ISS National Lab, is that all software driven, uh, those updates? It is. This is completely software driven. Um, we've developed this solution in a way that it can be uploaded over low bandwidth, uh, connectivity like we do have with the International Space Station. Um, for example, we've already made a patch update of 22 kilobytes, which is very small. We uploaded that almost instantly. We restarted the solution and we were able to see those improvements in real time. And so making updates to the system, we've designed it in a way that that's very simple and it's very quick. So A2E2 powered by space llama, obviously a fantastic solution for delivering information in this AI system format in space, but I imagine if you can put it in space, you can also put it at the bottom of the ocean or in airplane mode. So in your mind, what are some of the other use cases for A2E2? Yeah, you know, one that comes to mind that I think that could be very, very impactful is in medical evacuation flights. So when we have an operator or some a soldier that's in the battlefield right that gets critically injured, they might need to be medically evacuated back back to the states, um, and so currently what's going on in these flights is that surgeons or medics have access to information perhaps there's an injury that they they don't treat every day, right? And so they might need to look some things up. Um, they're currently doing that with paper manuals. They might have binders of standards of care for particular injuries. Um, that could take a long time to find the answer, right as you understand, someone who's in a traumatic has a traumatic injury, we might not have a lot of time for that, not to mention they might just not have the information, right? We can only have so many books of information in a medical evacuation flight. So I see a scenario where we could take A2E2. And put it on a medical evacuation flight, upload all of the information we need. We can upload a lot more information than a surgeon can parse in an emergency, uh, and provide the the standard of care or the information that the surgeon might need almost instantly summarized and uh with a response that's relevant to how the surgeon might need that information. Thank you so much, Zane and thanks for the demo. It was great to talk to A2E2, and if you would like to learn more about this capability, visit booz Allen.com/space.听