{"id":12411,"date":"2024-10-15T20:24:03","date_gmt":"2024-10-16T03:24:03","guid":{"rendered":"https:\/\/mattfife.com\/?p=12411"},"modified":"2024-12-31T22:26:22","modified_gmt":"2025-01-01T05:26:22","slug":"12411","status":"publish","type":"post","link":"https:\/\/mattfife.com\/?p=12411","title":{"rendered":"More details of the AI piloted fighter jet"},"content":{"rendered":"\n
\n

“It\u2019s a security risk not to have it. At this point, we have to have it [AI flight capabilities]”<\/p>\n<\/blockquote>\n\n\n\n

Back in 2016, I wrote about how a lone grad student created a fighter pilot AI that was able to defeat a retired U.S. Air Force Colonel every engagement<\/a>. Back in April 2024, I wrote about how a new project put AI into an real F16 called X-62A, or codenamed Vista<\/a>. At the time, they didn’t reveal the results of those tests other than saying they ‘went well’.<\/p>\n\n\n\n

Well, now we know what happened. First off, when the VISTA AI test aircraft flew against humans in a dogfight, it was able to do so with no safety rule violations logged. During simulation tests named AlphaDogfight, kicked off in 2020, AI agents developed by Heron Systems beat an experienced Air Force F-16 pilot in simulated dogfights 5-0<\/a>. Essentially duplicating what the grad student was able to achieve in 2016 simulations.<\/p>\n\n\n\n

In further live tests, Air Force Secretary Frank Kendall said, “We were up against a pilot who had two or three thousand hours of experience. He was very good. It was roughly an even fight. But against a less experienced pilot the AI \u2026 [and] the automation would have performed better,\u201d<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

We also learned that in this most recent experiment with a real plane, the re-enforcement learning based AI pilot wasn’t trained to fly from point A to point B or do any kind of normal ‘civilian’ flying. Instead, it was meant to figure out a specific dogfighting problem<\/a>:<\/p>\n\n\n\n

\u201cWe did about 10 or a dozen different situations where \u2026 I was in the front seat and I had a button on my stick where basically I initiated the automation,\u201d Air Force Secretary Frank Kendall, the real pilot in the AI plane, said<\/a> at an AI expo hosted by the Special Competitive Studies Project.<\/p>\n\n\n\n

\u201cIt was meant to do something much more sophisticated [than just fly the plane], but limited in that we had to reason over a highly dynamic situation. So, where is the adversary, what is the adversary doing, how do I put my airplane in position to solve for both safety and weapons effectiveness?\u201d<\/p>\n\n\n\n

This distinction of using AI to help solve getting the plane in position against an opponent leaves the critical decision to fire up to the human; while still solving one of the most critical problems of air combat – which is out-foxing your opponent to get a weapons lock on them.<\/p>\n\n\n\n

At the end of the hourlong flight, Kendall climbed out of the cockpit grinning. He said he\u2019d seen enough during his flight that he\u2019d trust this still-learning AI with the ability to decide whether or not to launch weapons in war.<\/p>\n\n\n\n

\n