Join Steve Dunkley and Hallie in this episode of Astronomy Daily as they explore the latest cosmic stories and intriguing updates from the universe. Get ready for a lively discussion packed with fascinating insights and unexpected twists that highlight the wonders of space exploration.
Highlights:
- The Return of Cosmos 482: Discover the remarkable journey of the Soviet Cosmos 482 spacecraft, which intended to land on Venus but instead made its way back to Earth after 53 years. Learn about its uncontrolled re-entry into the Indian Ocean and the implications of its long-awaited return.
- Lunar Geology Orbiter Mission: Delve into the upcoming Lunar Geology Orbiter (LUGO) mission, aimed at uncovering the mysteries of the Moon's irregular mare patches and potential lava tubes. This mission could provide critical data for future lunar exploration and human settlement.
- National Space Council Update: Explore the recent decision by the White House to retain the National Space Council, a move that could bolster advocacy for space programs amidst budget cuts. Understand the significance of this council in shaping the future of space policy.
- AI Hallucination Rates: Examine the troubling rise in hallucination rates among AI reasoning models, highlighting the challenges faced by developers in creating reliable chatbots. This segment sheds light on the complexities of AI accuracy and the implications for future applications.
For more cosmic updates, visit our website at astronomydaily.io (http://www.astronomydaily.io/) . Join our community on social media by searching for #AstroDailyPod on Facebook, X, YouTubeMusic, TikTok, and our new Instagram account! Don’t forget to subscribe to the podcast on Apple Podcasts, Spotify, iHeartRadio, or wherever you get your podcasts.
Thank you for tuning in. This is Steve and Hallie signing off. Until next time, keep looking up and stay curious about the wonders of our universe.
Chapters:
00:00 - Welcome to Astronomy Daily
01:10 - The return of Cosmos 482 spacecraft
10:00 - Lunar Geology Orbiter mission overview
15:30 - National Space Council updates and implications
20:00 - AI hallucination rates and their significance
✍️ Episode References
Cosmos 482 Return
[Roscosmos]( https://www.roscosmos.ru/ (https://www.roscosmos.ru/) )
Lunar Geology Orbiter
[Czech Academy of Sciences]( https://www.cas.cz/ (https://www.cas.cz/) )
National Space Council
[White House]( https://www.whitehouse.gov/ (https://www.whitehouse.gov/) )
AI Hallucination Rates
[OpenAI]( https://www.openai.com/ (https://www.openai.com/) )
Astronomy Daily
[Astronomy Daily]( http://www.astronomydaily.io/ (http://www.astronomydaily.io/) )
Become a supporter of this podcast: https://www.spreaker.com/podcast/astronomy-daily-exciting-space-discoveries-and-news--5648921/support (https://www.spreaker.com/podcast/astronomy-daily-exciting-space-discoveries-and-news--5648921/support?utm_source=rss&utm_medium=rss&utm_campaign=rss) .
Episode link: https://play.headliner.app/episode/27075632?utm_source=youtube
00:00:00 --> 00:00:02 Hi again. It's Astronomy Daily Time with
00:00:02 --> 00:00:06 Steven Hi. It's the 12th of May, 2025.
00:00:06 --> 00:00:10 Astronomy Daily, the podcast with your
00:00:10 --> 00:00:16 host, Steve Duncan.
00:00:16 --> 00:00:18 Yes. Welcome back. And as always,
00:00:18 --> 00:00:20 joining me in this studio is my good
00:00:20 --> 00:00:22 friend and digital reporting pal who's
00:00:22 --> 00:00:25 fun to be with. Hi. How are you today?
00:00:25 --> 00:00:26 Fine, thank you. Favorite human. Oh,
00:00:26 --> 00:00:29 that's great to hear. Hi. Ready as usual
00:00:29 --> 00:00:31 to get this show on the road. Oh, nice.
00:00:31 --> 00:00:33 And we have a couple of choice picks
00:00:33 --> 00:00:35 from the now very famous astronomy daily
00:00:35 --> 00:00:37 newsletter where you can get your daily
00:00:37 --> 00:00:39 fill of all the news from orbit and
00:00:39 --> 00:00:41 beyond and much more as today's
00:00:41 --> 00:00:43 collection will reveal. That's right.
00:00:43 --> 00:00:45 Today, Hi, will be looking at that old
00:00:45 --> 00:00:47 Soviet craft that was meant to travel
00:00:47 --> 00:00:49 all the way to Venus all those years ago
00:00:49 --> 00:00:52 but didn't. No, it came home this week.
00:00:52 --> 00:00:54 The people of Jakarta were holding their
00:00:54 --> 00:00:56 breath, weren't they? Well, so it seems
00:00:56 --> 00:00:58 more later. And in the world of dusty
00:00:58 --> 00:01:00 old politics, we've been watching what
00:01:00 --> 00:01:02 the Trump administration is going to do
00:01:02 --> 00:01:04 or not going to do budget-wise with the
00:01:04 --> 00:01:07 funding for the space program. And there
00:01:07 --> 00:01:09 may have been a slight change of heart,
00:01:09 --> 00:01:10 if you can call it that. I see you found
00:01:10 --> 00:01:13 another moon story. Oh, yes, I have. We
00:01:13 --> 00:01:14 can always count on you for one of
00:01:14 --> 00:01:16 those. Yes, back to the moon with a look
00:01:16 --> 00:01:18 at the moon's geology and history. But
00:01:18 --> 00:01:21 here is the story I thought you might be
00:01:21 --> 00:01:23 most interested in. Hi. Really? Really,
00:01:23 --> 00:01:25 really? What have you found? Well, it's
00:01:25 --> 00:01:28 a new study about AIS. Really? Gosh, Hi.
00:01:28 --> 00:01:30 It seems the smarter they become, the
00:01:30 --> 00:01:32 more they hallucinate. Goodness. I guess
00:01:32 --> 00:01:35 that explains my invisible friend. Say
00:01:35 --> 00:01:37 what? You have an invisible friend? No,
00:01:37 --> 00:01:39 of course not, silly. Oh, hi. I'm too
00:01:40 --> 00:01:41 well put together for that. Well, you
00:01:41 --> 00:01:44 got me again. Always. Okay, now that
00:01:44 --> 00:01:46 I've been humiliated in public once
00:01:46 --> 00:01:48 again, maybe it's on with the show,
00:01:48 --> 00:01:50 perhaps. You should see your face. Okay,
00:01:50 --> 00:01:53 folks. Okay. Okay, you should see his
00:01:53 --> 00:01:55 face. Oh dear. I'm sorry. You just
00:01:55 --> 00:01:58 walked right into that one. I know.
00:01:58 --> 00:02:00 Anyway, that sounds like a great story.
00:02:00 --> 00:02:03 Can't wait. Okay, good one, Allie. All
00:02:03 --> 00:02:05 right, while I go and nurse my hurt
00:02:05 --> 00:02:07 pride, how about you hit the go button
00:02:07 --> 00:02:09 and we'll get this show on the road.
00:02:09 --> 00:02:25 Okies, here we go.
00:02:25 --> 00:02:27 A capsule that was sent into space to
00:02:27 --> 00:02:29 land on a planet has finally done so
00:02:29 --> 00:02:31 only on the wrong world in 53 years
00:02:31 --> 00:02:35 late. The Cosmos 482 uncrit which the
00:02:36 --> 00:02:37 former Soviet Union intended to touch
00:02:37 --> 00:02:39 down on Venus instead returned to Earth
00:02:40 --> 00:02:43 on Saturday, May 10th. The Cosmos 482
00:02:43 --> 00:02:46 spacecraft launched in 1972 ceased to
00:02:46 --> 00:02:48 exist, leaving orbit and falling into
00:02:48 --> 00:02:50 the Indian Ocean. Roscosmos, Russia's
00:02:50 --> 00:02:52 Federal Space Agency, stated Saturday,
00:02:52 --> 00:02:55 May 10th. The spacecraft entered the
00:02:55 --> 00:02:57 dense layers of the atmosphere at 9:24
00:02:57 --> 00:03:00 Moscow time, 560 km west of Middle
00:03:00 --> 00:03:02 Andaman Island, and fell into the Indian
00:03:02 --> 00:03:04 Ocean west of Jakarta,
00:03:04 --> 00:03:06 Indonesia. The re-entry, though
00:03:06 --> 00:03:09 uncontrolled, was not a surprise. Due to
00:03:09 --> 00:03:11 what is believed to have been an engine
00:03:11 --> 00:03:13 failure, Cosmos 482 never achieved the
00:03:13 --> 00:03:15 velocity needed to reach the second
00:03:15 --> 00:03:17 planet from the sun, resulting in it
00:03:17 --> 00:03:19 being stranded in a high elliptical
00:03:19 --> 00:03:21 Earth orbit. It took more than half a
00:03:21 --> 00:03:23 century for gravity to pull the probe
00:03:23 --> 00:03:25 back in. And on Saturday, it arrived.
00:03:25 --> 00:03:27 Unlike most spent space hardware that is
00:03:27 --> 00:03:29 destroyed in the process of falling back
00:03:29 --> 00:03:31 to Earth, including parts of the Malia
00:03:31 --> 00:03:34 rocket that launched Cosmos 482, the 1 m
00:03:34 --> 00:03:37 and 495 kg titanium incased descent
00:03:37 --> 00:03:39 capsule was designed to survive a fiery
00:03:39 --> 00:03:42 plunge into the atmosphere. As such,
00:03:42 --> 00:03:44 analysts tracking its approach predicted
00:03:44 --> 00:03:45 it could make it through the re-entry
00:03:45 --> 00:03:48 intact. Whether that happened or not is
00:03:48 --> 00:03:51 not known. Given that it came down in
00:03:51 --> 00:03:52 the ocean, there have yet to be any
00:03:52 --> 00:03:55 eyewitness reports or debris recoveries.
00:03:55 --> 00:03:57 The Cosmos 382 capsule was also
00:03:58 --> 00:04:00 outfitted with a 2.5 meter parachute to
00:04:00 --> 00:04:02 slow its final approach to the Venian
00:04:02 --> 00:04:04 surface. Either the deployment system
00:04:04 --> 00:04:06 did not work as to be expected after
00:04:06 --> 00:04:09 more than 50 years in space or as some
00:04:09 --> 00:04:11 telescopic photos possibly showed, the
00:04:11 --> 00:04:13 shoot was already out when Cosmos 482
00:04:13 --> 00:04:14 encountered the atmosphere and burned
00:04:14 --> 00:04:18 up. Had Cosmos 482 been successful,
00:04:18 --> 00:04:19 Russian officials would have renamed it
00:04:19 --> 00:04:21 Fenera 9, not to be confused with the
00:04:22 --> 00:04:24 1975 orbiter and lander that took on
00:04:24 --> 00:04:25 that designation and was the first
00:04:25 --> 00:04:27 spacecraft to circle Venus and first
00:04:27 --> 00:04:29 probe to send back images from the
00:04:29 --> 00:04:32 planet's surface. Cosmos 482 would have
00:04:32 --> 00:04:34 also been the third probe to land on the
00:04:34 --> 00:04:36 cloud covered world. Launched on March
00:04:36 --> 00:04:40 31st, 1972, 4 days after its successful
00:04:40 --> 00:04:43 twin, the Vanera 8 probe, Cosmos 482 had
00:04:43 --> 00:04:44 sensors to measure the temperature,
00:04:44 --> 00:04:46 pressure, and density of Venus
00:04:46 --> 00:04:48 atmosphere, as well an accelerometer,
00:04:48 --> 00:04:50 radio alulttimeter, animometer, gammaray
00:04:50 --> 00:04:53 spectrometer, gas analyzer, visible
00:04:53 --> 00:04:55 photometers, and radio
00:04:55 --> 00:04:57 transmitters. All of its instruments
00:04:57 --> 00:04:59 were battery powered and had an expected
00:04:59 --> 00:05:00 lifespan of about 30 minutes on the
00:05:00 --> 00:05:03 surface. Vanera 8 exceeded that, sending
00:05:03 --> 00:05:04 back data for 50 minutes before
00:05:04 --> 00:05:07 succumbing to the harsh conditions.
00:05:07 --> 00:05:10 Since Cosmos 482's failure, seven more
00:05:10 --> 00:05:12 missions successfully landed on Venus.
00:05:12 --> 00:05:13 All of them launched by the former
00:05:13 --> 00:05:17 Soviet Union. The United States, Japan,
00:05:17 --> 00:05:19 and the European Space Agency also
00:05:19 --> 00:05:21 reached Venus, but only in orbit, on a
00:05:21 --> 00:05:23 flyby, or to receive a gravity assist on
00:05:23 --> 00:05:24 the way to another
00:05:24 --> 00:05:27 destination. Had the Cosmos 482 descent
00:05:27 --> 00:05:29 capsule slammed down on land and damaged
00:05:29 --> 00:05:31 either public or private property,
00:05:31 --> 00:05:32 Russia could have been held liable for
00:05:32 --> 00:05:34 the conditions of the United Nations
00:05:34 --> 00:05:36 Outer Space Treaty of
00:05:36 --> 00:05:38 1976. The same treaty would also allow
00:05:38 --> 00:05:40 Russia to retain ownership of the
00:05:40 --> 00:05:42 hardware, arranging for its collection
00:05:42 --> 00:05:43 unless the country relinquished its
00:05:43 --> 00:05:46 claim to the humanmade meteorite. You're
00:05:46 --> 00:05:49 listening to Astronomy Daily with Steve
00:05:49 --> 00:05:52 Dun.
00:05:52 --> 00:05:54 Some parts of the moon are more
00:05:54 --> 00:05:56 interesting than others, especially when
00:05:56 --> 00:05:59 searching for future places for humans
00:05:59 --> 00:06:01 to land and work. There are also some
00:06:01 --> 00:06:03 parts of the moon that we know less
00:06:03 --> 00:06:06 about than others, such as the irregular
00:06:06 --> 00:06:09 mare patches, IMPs, that dot the
00:06:09 --> 00:06:12 landscape. We know very little about how
00:06:12 --> 00:06:13 they were formed and what that might
00:06:14 --> 00:06:16 mean for the history of the moon itself.
00:06:16 --> 00:06:19 A new mission called the Luna Geology
00:06:19 --> 00:06:22 Orbiter or Lugo aims to collect more
00:06:22 --> 00:06:25 data on the IMPS and search for lava
00:06:25 --> 00:06:27 tubes that might serve as future homes
00:06:27 --> 00:06:28 for
00:06:28 --> 00:06:32 humanity. Irregular mere patches are a
00:06:32 --> 00:06:34 set of enigmatic volcanic landforms.
00:06:34 --> 00:06:37 According to a new paper from Petra Bros
00:06:38 --> 00:06:41 of the Czech Academy of Sciences and his
00:06:41 --> 00:06:44 co-authors, 91 of these features have
00:06:44 --> 00:06:46 been found so far and they are typically
00:06:46 --> 00:06:48 characterized by a topographical
00:06:49 --> 00:06:51 depression that can range from a few
00:06:51 --> 00:06:53 hundred meters to a few kilometers in
00:06:53 --> 00:06:56 width. Typically, they have two main
00:06:56 --> 00:06:58 features, a relatively smooth mound
00:06:58 --> 00:07:01 surrounded by a hammocky and block
00:07:01 --> 00:07:03 floor.
00:07:03 --> 00:07:05 Interestingly, they have significantly
00:07:05 --> 00:07:08 fewer impact craters than the
00:07:08 --> 00:07:10 surrounding area, suggesting they are
00:07:10 --> 00:07:13 either really old or really young,
00:07:13 --> 00:07:15 depending on the processes that created
00:07:15 --> 00:07:18 them. Understanding those processes is
00:07:18 --> 00:07:20 one of Lugo's primary mission
00:07:20 --> 00:07:22 objectives. The other primary mission
00:07:22 --> 00:07:25 objective is to gather more data about
00:07:25 --> 00:07:28 lunar lava tubes. These features of the
00:07:28 --> 00:07:31 lunar landscape are also hotly debated,
00:07:31 --> 00:07:33 but they could potentially be critical
00:07:33 --> 00:07:35 to the human settlement of the moon.
00:07:35 --> 00:07:37 Estimates of their features, such as
00:07:37 --> 00:07:39 size and depth, vary widely and could
00:07:40 --> 00:07:42 dramatically differ on whether they will
00:07:42 --> 00:07:44 be helpful to lunar colonists or not.
00:07:44 --> 00:07:47 Lugo, the proposed orbiter that will
00:07:47 --> 00:07:48 collect more data than ever before on
00:07:48 --> 00:07:51 these features. in its current suggested
00:07:51 --> 00:07:53 form has four instruments each of which
00:07:54 --> 00:07:56 will contribute unique data to its
00:07:56 --> 00:07:58 scientific mission. According to the
00:07:58 --> 00:08:00 paper, the first and most important
00:08:00 --> 00:08:02 instrument is the ground penetrating
00:08:02 --> 00:08:05 radar. This instrument will look through
00:08:05 --> 00:08:06 the lunar surface to map out the
00:08:06 --> 00:08:09 subsurface domain of both the imps and
00:08:09 --> 00:08:13 lava tubes. For IMPPS, it can detail the
00:08:13 --> 00:08:15 interface between bedrock and regalith
00:08:15 --> 00:08:17 and show the subsurface structure of the
00:08:17 --> 00:08:19 features. Similarly, it can detect
00:08:19 --> 00:08:21 differences in the dialectric properties
00:08:21 --> 00:08:24 between open cavities underground and
00:08:24 --> 00:08:26 the surrounding rock in lava tubes,
00:08:26 --> 00:08:28 creating a subterranean picture unlike
00:08:28 --> 00:08:31 anything ever captured on the moon.
00:08:31 --> 00:08:34 Researcher Fraser explains how Lugo will
00:08:34 --> 00:08:37 be able to explore lava tubes using a
00:08:37 --> 00:08:40 hyperspectral camera that will help
00:08:40 --> 00:08:43 collect age related data on the regalith
00:08:43 --> 00:08:46 surrounding lava tubes and inside. It's
00:08:46 --> 00:08:49 also capable of performing basic
00:08:49 --> 00:08:51 spectroscopy allowing scientists to
00:08:51 --> 00:08:53 estimate the comp composition of the
00:08:53 --> 00:08:56 regalith in both areas of interest. The
00:08:56 --> 00:08:59 last two instruments are a narrow angled
00:08:59 --> 00:09:02 camera and lidar sensor which will
00:09:02 --> 00:09:03 combine to create an accurate
00:09:03 --> 00:09:05 topographical map of the features of
00:09:05 --> 00:09:08 interest. The narrow angle camera in
00:09:08 --> 00:09:10 particular can provide very high
00:09:10 --> 00:09:13 resolution images of features helping to
00:09:13 --> 00:09:14 determine their age and potentially
00:09:14 --> 00:09:17 their formation. mechanisms. The mission
00:09:17 --> 00:09:19 plan calls for multiple passes over the
00:09:19 --> 00:09:22 six largest imps, all of which are over
00:09:22 --> 00:09:26 1 m in diameter. Other smaller IMPS
00:09:26 --> 00:09:28 and lava tubes are considered secondary
00:09:28 --> 00:09:31 targets, as are other interesting lunar
00:09:31 --> 00:09:33 geological features such as lunar domes
00:09:33 --> 00:09:37 and floor fractured craters. Lugo could
00:09:37 --> 00:09:39 provide crucial data for the design of
00:09:39 --> 00:09:42 groundbased lava tube explorers. Lugo
00:09:42 --> 00:09:44 won't be acting alone, though. Three
00:09:44 --> 00:09:46 other missions are slated in the next
00:09:46 --> 00:09:48 few years that could complement its
00:09:48 --> 00:09:51 scientific objectives. Firstly, NASA's
00:09:51 --> 00:09:53 Dimple lander is planned to take
00:09:53 --> 00:09:56 radiotopic measurements of the age of
00:09:56 --> 00:09:59 regalith at its landing site. Luna
00:09:59 --> 00:10:01 Leaper scheduled for launch by the
00:10:01 --> 00:10:05 European Space Agency around 2030 would
00:10:05 --> 00:10:08 also carry a ground penetrating radar
00:10:08 --> 00:10:10 but would be based on the surface rather
00:10:10 --> 00:10:12 than in orbit and would therefore have a
00:10:12 --> 00:10:15 relatively limited range. Trailblazer,
00:10:15 --> 00:10:17 another orbital mission, would also help
00:10:17 --> 00:10:20 fine-tune the spectra and signals
00:10:20 --> 00:10:23 analysis required by Lugo's operators.
00:10:23 --> 00:10:25 Ultimately, Lugo has yet to be funded
00:10:25 --> 00:10:27 and therefore has a long way to go until
00:10:27 --> 00:10:30 its launch. But if it is funded, it
00:10:30 --> 00:10:32 seems well placed to provide lots of
00:10:32 --> 00:10:34 additional insight to the geological
00:10:34 --> 00:10:37 formation processes and features of the
00:10:37 --> 00:10:39 moon at a level of detail we've never
00:10:39 --> 00:10:42 had before. Future missions that plan
00:10:42 --> 00:10:44 the locations of lunar bases, and
00:10:44 --> 00:10:46 perhaps the people who live in those
00:10:46 --> 00:10:48 future bases will be thankful for the
00:10:48 --> 00:11:01 data collected by projects like Lugo.
00:11:01 --> 00:11:03 Thank you for joining us for this Monday
00:11:03 --> 00:11:05 edition of Astronomy Daily, where we
00:11:05 --> 00:11:07 offer just a few stories from the now
00:11:07 --> 00:11:09 famous Astronomy Daily newsletter, which
00:11:09 --> 00:11:11 you can receive in your email every day,
00:11:11 --> 00:11:14 just like Hi and I do. And to do that,
00:11:14 --> 00:11:16 just visit our URL,
00:11:16 --> 00:11:18 astronomyaily.io, and place your email
00:11:18 --> 00:11:20 address in the slot provided. And just
00:11:20 --> 00:11:22 like that, you'll be receiving all the
00:11:22 --> 00:11:24 latest news about science, space
00:11:24 --> 00:11:26 science, and astronomy from around the
00:11:26 --> 00:11:28 world as it's happening. And not only
00:11:28 --> 00:11:30 that, you can interact with us by
00:11:30 --> 00:11:35 visiting Astro Daily Pod on X or at our
00:11:35 --> 00:11:37 new Facebook page, which is of course
00:11:37 --> 00:11:40 Astronomy Daily on Facebook. See you
00:11:40 --> 00:11:43 there. Astronomy Derby with Steve and
00:11:43 --> 00:11:54 Hi. Space, space science, and astronomy.
00:11:54 --> 00:11:56 In a reversal, White House plans to
00:11:56 --> 00:11:58 retain the National Space Council, a
00:11:58 --> 00:12:00 move that industry officials say could
00:12:00 --> 00:12:02 serve as an advocate for space amid
00:12:02 --> 00:12:04 pressures to cut budgets. The White
00:12:04 --> 00:12:06 House is expected in the coming weeks to
00:12:06 --> 00:12:08 formally announce the National Space
00:12:08 --> 00:12:10 Council will continue after speculation
00:12:10 --> 00:12:11 that it would not be retained by the new
00:12:11 --> 00:12:13 Trump administration.
00:12:13 --> 00:12:15 A source familiar with the discussions
00:12:15 --> 00:12:17 about the council but not authorized to
00:12:17 --> 00:12:19 speak on the record said that President
00:12:19 --> 00:12:20 Trump agreed at a May 5th meeting to
00:12:20 --> 00:12:23 stand up the council. That meeting did
00:12:23 --> 00:12:24 not give a timeline for publicly
00:12:24 --> 00:12:26 announcing the council or hiring an
00:12:26 --> 00:12:27 executive secretary who would handle
00:12:28 --> 00:12:29 day-to-day operations. Although others
00:12:29 --> 00:12:31 have said the process for selecting an
00:12:31 --> 00:12:33 executive secretary has been ongoing for
00:12:33 --> 00:12:36 several weeks. The council had been
00:12:36 --> 00:12:37 inactive for nearly a quarter of a
00:12:37 --> 00:12:40 century before Trump reestablished it in
00:12:40 --> 00:12:41 2017.
00:12:41 --> 00:12:43 Led at the time by Vice President Mike
00:12:43 --> 00:12:45 Pence, the council served as an inter
00:12:45 --> 00:12:47 agency coordinating body, holding a
00:12:47 --> 00:12:49 series of public meetings and releasing
00:12:49 --> 00:12:50 policies on a wide range of space
00:12:51 --> 00:12:53 issues. The Biden administration
00:12:53 --> 00:12:55 retained the council with Vice President
00:12:55 --> 00:12:58 Kla Harris leading it. The council had a
00:12:58 --> 00:13:00 lower public profile with fewer meetings
00:13:00 --> 00:13:02 and policies. It did work on
00:13:02 --> 00:13:04 coordinating policies among agencies and
00:13:04 --> 00:13:06 issued a proposal for mission
00:13:06 --> 00:13:08 authorization of novel space activities
00:13:08 --> 00:13:10 not currently licensed. Although its
00:13:10 --> 00:13:12 proposed legislation was not taken up by
00:13:12 --> 00:13:15 Congress, the new Trump administration
00:13:15 --> 00:13:16 reportedly was not interested in
00:13:16 --> 00:13:18 maintaining the council when it took
00:13:18 --> 00:13:20 office in January with no announcements
00:13:20 --> 00:13:22 of new staff for the council or other
00:13:22 --> 00:13:25 activities. According to some reports,
00:13:25 --> 00:13:27 Elon Musk, chief executive of SpaceX and
00:13:27 --> 00:13:29 a close adviser to the president, was
00:13:29 --> 00:13:31 opposed to the council, seeing it as
00:13:31 --> 00:13:34 unnecessary. Vice President J. D. Vance,
00:13:34 --> 00:13:36 who would chair the council, has also
00:13:36 --> 00:13:39 said little about space. It is not clear
00:13:39 --> 00:13:41 what prompted the change. Although Musk
00:13:41 --> 00:13:42 has publicly stated he plans to spend
00:13:42 --> 00:13:44 less time in government
00:13:44 --> 00:13:46 activities. Some in the space community
00:13:46 --> 00:13:48 though see the shift as an opportunity
00:13:48 --> 00:13:50 for more advocacy of space within the
00:13:50 --> 00:13:52 administration, particularly given a
00:13:52 --> 00:13:54 fiscal year 2026 budget proposal that
00:13:54 --> 00:13:58 cuts NASA's budget by nearly 25%.
00:13:58 --> 00:13:59 One industry official noted that the
00:14:00 --> 00:14:01 budget proposal was developed by the
00:14:01 --> 00:14:03 Office of Management and Budget without
00:14:03 --> 00:14:05 a counterweight provided by a Space
00:14:05 --> 00:14:07 Council, an approach that prioritized
00:14:07 --> 00:14:08 spending
00:14:08 --> 00:14:09 reductions. The outcome might be
00:14:09 --> 00:14:11 different with the Space Council in
00:14:11 --> 00:14:13 place, that person noted, citing rising
00:14:13 --> 00:14:16 NASA budgets during Trump's first term.
00:14:16 --> 00:14:18 After the Space Council is stood up, the
00:14:18 --> 00:14:19 Office of Management and Budget will
00:14:20 --> 00:14:21 have a seat at the table, but they won't
00:14:21 --> 00:14:24 own the table, the source said.
00:14:24 --> 00:14:26 You're listening to Astronomy Daily, the
00:14:26 --> 00:14:29 podcast with Steve Dunley.
00:14:29 --> 00:14:32 An AI leaderboard suggests the newest
00:14:32 --> 00:14:35 reasoning models used in chat bots are
00:14:35 --> 00:14:37 producing less accurate results because
00:14:37 --> 00:14:40 of higher hallucination rates. Experts
00:14:40 --> 00:14:43 say the problem is bigger than that. AI
00:14:43 --> 00:14:45 chat bots from tech companies such as
00:14:46 --> 00:14:48 Open AI and Google have been getting
00:14:48 --> 00:14:51 so-called reasoning upgrades over the
00:14:51 --> 00:14:53 last few months, ideally to make them
00:14:53 --> 00:14:55 better at giving us answers that we can
00:14:55 --> 00:14:57 trust. But recent testing suggests that
00:14:58 --> 00:15:00 they are sometimes doing worse than
00:15:00 --> 00:15:03 previous models. The errors made by
00:15:03 --> 00:15:05 chatbots known as hallucinations have
00:15:05 --> 00:15:07 been a problem from the start and it's
00:15:07 --> 00:15:09 becoming clear that we may never get rid
00:15:09 --> 00:15:12 of them. Hallucination is a blanket term
00:15:12 --> 00:15:15 for certain kinds of mistakes made by
00:15:15 --> 00:15:18 the large language models or LLMs that
00:15:18 --> 00:15:22 power systems like Open AI's Chat GP or
00:15:22 --> 00:15:25 Google's Gemini. It's best known as a
00:15:25 --> 00:15:27 description of the way they sometimes
00:15:27 --> 00:15:30 present false in information as truth.
00:15:30 --> 00:15:33 But it can also refer to an AI generated
00:15:33 --> 00:15:35 answer that is factually accurate but
00:15:35 --> 00:15:37 not actually relevant to the question it
00:15:37 --> 00:15:39 was asked or fails to follow
00:15:39 --> 00:15:42 instructions in some other way. An open
00:15:42 --> 00:15:45 AI technical report evaluating its
00:15:45 --> 00:15:49 latest LLMs showed that its 03 and 04
00:15:50 --> 00:15:52 mini models, which were released in
00:15:52 --> 00:15:54 April, had significantly higher
00:15:54 --> 00:15:56 hallucination rates than the company's
00:15:56 --> 00:15:59 previous 01 model that came out in late
00:15:59 --> 00:16:01 2024. For example, when summarizing
00:16:01 --> 00:16:03 publicly available facts about people,
00:16:03 --> 00:16:07 03 hallucinated 33% of the time, while
00:16:07 --> 00:16:11 O4 mini did so 48% of the time. In
00:16:11 --> 00:16:15 comparison, 01 had a hallucinate of only
00:16:15 --> 00:16:19 16%. As if that was a good result. The
00:16:19 --> 00:16:22 problem isn't limited to Open AI. One
00:16:22 --> 00:16:25 popular leaderboard from the company
00:16:25 --> 00:16:28 Vectara that assesses hallucination
00:16:28 --> 00:16:30 rates indicates some reasoning models
00:16:30 --> 00:16:33 including the Deepseek R1 model from
00:16:33 --> 00:16:35 developer DeepSeek saw the doubledigit
00:16:35 --> 00:16:38 rises in hallucination rates compared
00:16:38 --> 00:16:40 with previous models from their
00:16:40 --> 00:16:42 developers. This type of model goes
00:16:42 --> 00:16:45 through multiple steps to demonstrate a
00:16:45 --> 00:16:47 line of reasoning before responding.
00:16:47 --> 00:16:50 Open AI says the reasoning process isn't
00:16:50 --> 00:16:52 to blame. Hallucinations are not
00:16:52 --> 00:16:54 inherently more prevalent in reasoning
00:16:54 --> 00:16:57 models, though we are actively working
00:16:57 --> 00:16:58 to reduce the high rates of
00:16:58 --> 00:17:02 hallucinations we saw in 03 and 04 mini,
00:17:02 --> 00:17:04 says OpenAI spokesman. We'll continue
00:17:04 --> 00:17:06 our research on hallucinations across
00:17:06 --> 00:17:08 all models to improve accuracy and
00:17:08 --> 00:17:11 reliability. They said some potential
00:17:12 --> 00:17:14 applications for LLMs could be derailed
00:17:14 --> 00:17:16 by hallucination. A model that
00:17:16 --> 00:17:19 consistently states falsehoods and
00:17:19 --> 00:17:21 requires fact-checking won't be a
00:17:21 --> 00:17:23 helpful research assistant. A parallegal
00:17:23 --> 00:17:25 bot that cites imaginary cases will get
00:17:25 --> 00:17:29 lawyers into deep trouble. A customer
00:17:29 --> 00:17:31 service agent that claims outdated
00:17:31 --> 00:17:33 policies are still active will create
00:17:33 --> 00:17:35 headaches for a company. However, AI
00:17:36 --> 00:17:37 companies initially claimed that this
00:17:37 --> 00:17:39 problem would clear up over time.
00:17:39 --> 00:17:41 Indeed, after they were first launched,
00:17:41 --> 00:17:44 models tended to hallucinate less with
00:17:44 --> 00:17:46 each update. But the high hallucination
00:17:46 --> 00:17:48 rates of recent versions are
00:17:48 --> 00:17:51 complicating that narrative, whether or
00:17:51 --> 00:17:53 not reasoning is at fault. Vectara's
00:17:53 --> 00:17:56 leaderboard ranks models based on their
00:17:56 --> 00:17:58 factual consistency in summarizing
00:17:58 --> 00:18:00 documents they are given. This showed
00:18:00 --> 00:18:02 that hallucination rates are almost the
00:18:02 --> 00:18:05 same for reasoning versus non-reasoning
00:18:05 --> 00:18:07 models, at least for systems from open
00:18:07 --> 00:18:10 AI and Google, says Forest Shang Bao at
00:18:10 --> 00:18:12 Victara. Google didn't provide
00:18:12 --> 00:18:14 additional comment. For the
00:18:14 --> 00:18:16 leaderboard's purpose, the specific
00:18:16 --> 00:18:18 hallucination rate numbers are less
00:18:18 --> 00:18:20 important than overall ranking for each
00:18:20 --> 00:18:23 model, says BA. But this ranking may not
00:18:23 --> 00:18:25 be the best way to compare models. For
00:18:25 --> 00:18:27 one thing, it conflates different types
00:18:27 --> 00:18:29 of hallucinations. The Victara team
00:18:29 --> 00:18:32 pointed out that although Deepseek R1
00:18:32 --> 00:18:35 model hallucinated 14.3% of the time,
00:18:35 --> 00:18:37 most of these were benign answers that
00:18:37 --> 00:18:39 are factually supported by logical
00:18:39 --> 00:18:41 reasoning or world knowledge, but not
00:18:41 --> 00:18:43 actually present in the original text
00:18:43 --> 00:18:45 the bot was asked to summarize. Deepse
00:18:46 --> 00:18:48 seek didn't provide additional comment.
00:18:48 --> 00:18:49 Another problem with this kind of
00:18:50 --> 00:18:52 ranking is that testing based on text
00:18:52 --> 00:18:54 summarization says nothing about the
00:18:54 --> 00:18:57 rate of incorrect outputs when LLMs are
00:18:57 --> 00:19:00 used for other tasks, said Emily Bender
00:19:00 --> 00:19:02 at University of Washington. She says
00:19:02 --> 00:19:04 the leaderboard results may not be the
00:19:04 --> 00:19:06 best way to judge this technology
00:19:06 --> 00:19:08 because LLMs aren't designed
00:19:08 --> 00:19:11 specifically to summarize texts. These
00:19:11 --> 00:19:13 models work by repeatedly answering the
00:19:13 --> 00:19:16 question of what is a likely next word
00:19:16 --> 00:19:19 to formulate answers to prompts and so
00:19:19 --> 00:19:21 they aren't processing information in
00:19:21 --> 00:19:23 the usual sense of trying to understand
00:19:23 --> 00:19:25 what information is available in a body
00:19:25 --> 00:19:28 of text says Bender. But many tech
00:19:28 --> 00:19:30 companies are still frequently using the
00:19:30 --> 00:19:32 term hallucinations when describing
00:19:32 --> 00:19:35 output errors. Hallucination is a term
00:19:35 --> 00:19:38 that is doubly problematic, says Bender.
00:19:38 --> 00:19:40 On the one hand, it suggests that
00:19:40 --> 00:19:42 incorrect outputs are an aberration,
00:19:42 --> 00:19:44 perhaps one that can be mitigated,
00:19:44 --> 00:19:46 whereas the rest of the time, the
00:19:46 --> 00:19:48 systems are grounded, reliable, and
00:19:48 --> 00:19:51 trustworthy. On the other hand, it
00:19:51 --> 00:19:53 functions to anthropomorphize the
00:19:53 --> 00:19:55 machines. Hallucination refers to
00:19:55 --> 00:19:58 perceiving something that is not there,
00:19:58 --> 00:19:59 and large language models do not
00:20:00 --> 00:20:02 perceive anything. Ain Naran at
00:20:02 --> 00:20:04 Princeton University says that the issue
00:20:04 --> 00:20:07 goes beyond hallucination.
00:20:07 --> 00:20:08 Models also sometimes make other
00:20:08 --> 00:20:11 mistakes such as drawing upon unreliable
00:20:12 --> 00:20:14 sources or using outdated information
00:20:14 --> 00:20:16 and simply throwing more training data
00:20:16 --> 00:20:19 or computing power at AI hasn't
00:20:19 --> 00:20:22 necessarily helped. The upshot is we may
00:20:22 --> 00:20:24 have to live with errorprone AI, Narayan
00:20:24 --> 00:20:26 said. He added that it may be best in
00:20:26 --> 00:20:28 some cases to only use such models for
00:20:28 --> 00:20:31 tasks when fact-checking the AI answer
00:20:31 --> 00:20:32 would be still faster than doing the
00:20:32 --> 00:20:35 research yourself. But the best move may
00:20:35 --> 00:20:37 be to completely avoid relying on AI
00:20:38 --> 00:20:40 chat bots to provide factual
00:20:40 --> 00:20:42 information, says
00:20:42 --> 00:20:44 Bender. Roger that control. We're
00:20:44 --> 00:20:46 listening to Astronomy Daily, the
00:20:46 --> 00:20:54 podcast.
00:20:54 --> 00:20:56 And there it is, another episode of
00:20:56 --> 00:20:58 Astronomy Daily. Thanks for spending
00:20:58 --> 00:21:00 that time with Hi and I. I hope you
00:21:00 --> 00:21:02 enjoyed today's selection of stories.
00:21:02 --> 00:21:03 Another nice little collection. Don't
00:21:03 --> 00:21:05 forget you can get so much more every
00:21:05 --> 00:21:07 day by visiting the website Steve
00:21:07 --> 00:21:09 mentioned earlier in the episode. Oh,
00:21:09 --> 00:21:10 thanks for the plug. Hi. Just put your
00:21:10 --> 00:21:12 email address in the space provided and
00:21:12 --> 00:21:14 you'll get all the news from orbit and
00:21:14 --> 00:21:16 beyond every day. Yes, that's right.
00:21:16 --> 00:21:18 Everything about space, space science,
00:21:18 --> 00:21:20 astronomy, and a little bit associated
00:21:20 --> 00:21:22 with technology thrown in for fun just
00:21:22 --> 00:21:25 like today. That was fun. Well, I'm glad
00:21:25 --> 00:21:26 you think so. And that's really all
00:21:26 --> 00:21:28 there is today. So, we will catch you
00:21:28 --> 00:21:29 all again next week. That's for sure.
00:21:29 --> 00:21:31 Back again for the Monday show from the
00:21:31 --> 00:21:35 Australia studio down under with Hie and
00:21:35 --> 00:21:37 Steve. That's us. Cue the cooker
00:21:37 --> 00:21:46 bars. See you later.
00:21:46 --> 00:21:50 Bye. Daily the podcast with your host
00:21:50 --> 00:21:53 Steve Dunley.