Of course you have to use something like artificial intelligence, to
learn, then make decisions. The time lag for communications makes
it impossible for it to wait for human command to make any movement.

from
https://www.space.com/artificial-intelligence-chatgpt-mars

ChatGPT on Mars: How AI can help scientists study the Red Planet

By Leonard David published 2 days ago
"This is not trivial. It is not just a paper. It is about who we really
want to become as a species."

(opens in new tab)
(opens in new tab)
(opens in new tab)
(opens in new tab)
(opens in new tab)
Comments (6)
robots on Mars
Could robots on Mars be supercharged with AI to perform on-the-spot
research, relaying their discoveries in real-time? (Image credit:
NASA/JPL-Caltech)
The world is abuzz, perhaps even befuddled, about the growing use of
artificial intelligence. One of the most popular artificial intelligence
(AI) tools available to the public today is ChatGPT, an AI-powered
language model that has been "trained" and fed vast amounts of online
information. After taking all that in, ChatGPT can regurgitate
human-like text responses to a given prompt. It can respond to queries,
discuss a lot of topics and crank out pieces of writing.

It isn't difficult to imagine a robot wheeling and dealing on the
surface of Mars, factory-wired with ChatGPT or a similar artificial
intelligence language model. This smartbot could be loaded with a suite
of science devices. It could analyze what its scientific instruments are
finding "on-the-spot," perhaps even collating any evidence of past life
it uncovers nearly instantly.

That data could be digested, assessed, appraised and assembled in some
scientific form. The product, in well-paginated condition, with
footnotes to boot, could then be transmitted directly from the robot to
a scientific journal, like Science or Nature, for publication. Of
course, that paper would then be peer reviewed — maybe by AI/ChatGPT
reviewers. Sound far-fetched?

I reached out to several leading researchers, presenting this off-Earth,
on-Mars scenario, with a variety of reactions in return.

Related: Artificial intelligence could help hunt for life on Mars and
other alien worlds

Click here for more Space.com videos...
CLOSE

Prone to hallucination
"It could be done but there could be misleading information," said
Sercan Ozcan, Reader in Innovation and Technology Management at the
University of Portsmouth in the United Kingdom. "ChatGPT is not 100%
accurate and it is prone to 'hallucination.'"

Ozcan said he's not sure if ChatGPT would be valuable if there is no
prior volume of work for it to analyze and emulate. "I believe humans
can still do better work than ChatGPT, even if it is slower," he said.

His advice is to not use ChatGPT "in areas where we cannot accept any
error."

Humans in the loop
Steve Ruff, associate research professor at Arizona State University's
School of Earth and Space Exploration in Tempe, Arizona, is keenly tied
to studying Mars.

"My immediate reaction is that it's highly unlikely that 'on-the-spot'
manuscripts would be a realistic scenario given how the process involves
debates among the team over the observations and their interpretation,"
Ruff said. "I'm skeptical that any AI, trained on existing observations,
could be used to confidently interpret new observations without humans
in the loop, especially with new instrument datasets that have not been
available previously. Every such dataset requires painstaking efforts to
sort out."

For the near term, Ruff thinks AI could be used for rover operations,
like picking targets to observe without humans in the loop, and for
navigation.

NASA's Ingenuity Mars helicopter, photographed by the agency's
Perseverance rover on April 16, 2023. The rover captured this
enhanced-color image using its Mastcam-Z instrument.

NASA's Ingenuity Mars helicopter, photographed by the agency's
Perseverance rover on April 16, 2023. Ingenuity has been used as a
'scout' to help identify locations for Perseverance to study. (Image
credit: NASA/JPL-Caltech/ASU/MSSS)
First things first
In what world do we want to live?

Perhaps that is the strongest question, said Nathalie Cabrol, Director
of the Carl Sagan Center for Research at the SETI Institute in Mountain
View, California.

"First things first," Cabrol said. "AI is a formidable tool and should
be used as such to support humans in their activity. We actually do that
already every day, in one form or another," she added, "and improved
versions might make things better."

On the other hand, like any human tools, they are double-edged swords
and sometimes lead people to start thinking "nonsense," Cabrol added,
and she believes that to be the case here.

"I do personally like writing papers. It is a great time where I see my
work coming to fruition and can put my ideas together on paper," Cabrol
said, and sees that as an important part of her creative process.

"But let's assume for a moment that I let this algorithm write it for
me. Then, I am being told that it's okay because the paper will be
reviewed," Cabrol said. "But by whom? I would assume that if you let
algorithms do the job for you it's because you assume they will be less
biased and do a better job? Following that logic, I would assume that a
human is not qualified to review that paper."

Specters of "transhumanism"
Cabrol senses that a next question is: Where do we stop? What if all
researchers ask AI to write their research grant proposals? What if they
do and don't tell?

"This depends in which world you want to live and what part you want
left to humanity," Cabrol said. "We are creative beings and we are not
perfect," she continued, "but we learn from our mistakes and that's part
of our evolution. Mistake and learning are other words for
'adaptation'," she said.

By letting AI getting into what makes us human, we are messing with our
own evolution, Cabrol added, and she sees specters of "transhumanism" in
all of this. Transhumanism can be defined as a loose ideological
movement united by the belief that the human race can evolve beyond its
current physical and mental limitations, especially by means of science
and technology.

"Of course, that's not a chip in our brain and that's only a paper, you
will say. Unfortunately, it is part of a much broader, and very
disturbing, discourse on the (mis) use of AI," Cabrol concluded. "This
is not trivial. It is not just a paper. It is about who we really want
to become as a species. Personally, I see AI useful as a tool, and I
will confine it as that."

Knowledge cutoff
"How funny that we still argue about the definition of life as we know
it, and we're starting to use a tool in that search that also stretches
the definition of life," said Amy Williams, assistant professor in
Geological Sciences at the University of Florida in Gainesville. She is
a participating scientist on the NASA Curiosity and Perseverance rover
missions that have robots scouting about on Mars.

Williams reacted to the AI-ChatGPT off-world setting in full disclosure
mode. "The first time I used ChatGPT was in preparing for this response,
asking it: 'What organic molecules have the Mars rovers found?' The
question was based on my particular field of expertise," she told
Space.com.

The Perseverance rover on the surface of Mars

Perseverance rover on the surface of Mars. (Image credit:
NASA/JPL-Caltech/MSSS)
"It was illuminating in that it did a great job providing me with
statements that I would describe as robust and appropriate for a summary
that I could give in an outreach talk to the general public about
organic molecules on Mars," said Williams.

But it also demonstrated to Williams its limitation in that it could
only access data from, in her case, September 2021 — flagging it as a
"knowledge cutoff."

"So its responses did not encompass the full breadth of published
results about organics on Mars that I know about since 2021," she said.

Emphasizing that she is not a specialist in AI or machine learning,
Williams said that future iterations of ChatGPT + AI will likely be able
to incorporate more recent data and generate a complete synthesis of the
recent results from any given scientific exploration.

"However, I still see these as tools to use in step with humans, instead
of in place of humans," Williams remarked. "Given the limitations in
data uplink and downlink with our current Deep Space Network, it is
difficult for me to see a way to upload the knowledge base for something
as complex as, for example, the current and historic data and context
for the sources, sinks, and fates of organic molecules on Mars so that
the onboard AI could generate a manuscript for publication," she said.

On Wednesday, May 17, 2023 at 1:56:13 AM UTC+10, a425couple wrote:
> Of course you have to use something like artificial intelligence, to
> learn, then make decisions. The time lag for communications makes
> it impossible for it to wait for human command to make any movement.
>
> from
> https://www.space.com/artificial-intelligence-chatgpt-mars
>
> ChatGPT on Mars: How AI can help scientists study the Red Planet
return.
....

I'm an advocate of using AI in various science-based projects. "Unmanned science" is what we called it back in the 2010s. :)

But one thing the growing body of media coverage hasnt pointed out so far is -- AI techniques can be incredibly brittle.

If you write code you know (after a while ;) that sometimes code handles what it expects very well but you wander off the path a bit and use a tool for something it was never intended for and it
falls in a heap. Sometimes noisly (good) and sometimes quietly (bad!!!).

A project we did on Kaggle.com underscored the problem. We had been writing s/w to recognize things in photos. E.g. recognizing individual whales
bu the pattern of bumps on their noses. Recognizing numberplates, or the faces of people in your pix library.

But one project turned that on its head. Try to find ways to alter images -- usually by the insertion of odd-colored pixels here and there or swapping rows of pixels and other stuff like that --
so that the images was mis-identified. I.e. adjust the image of something like a mouse so the AI s/w would identify it as a motor vehicle. Or vice-versa.

Turned out to be hysterically easy. If you knew vaguely how the image recog algorithm worked you could screw with it and make it dance like a little monkey.

The next part of the project was almost impossible. Create an algorithm that could not be screwed with. :)

So it turns out AI is ripe for hacking attacks. Any system "protected" by an AI software of most current designs can be messed up. You might be able to get access to whatever it is protecting but you might also be able to fix it so the people that are SUPPOSED to have access will be denied access.

And until the algorithms have been found that can protect themselves against hacking -- and they probably will never be found -- you had better watch out. The onlyway we found on that Kaggle project was to use several different recog algorithms that did not "seem" to have a common problem and majority vote the decisions.
It wasnt a hard and fast fix but it at least made it harder to give the thing hallucinations.