How Can AI Be Used for Space Applications?

[sanman]

Another round in the AI wars  --- Musk has formed a consortium which is now trying to buy OpenAI in a sort of hostile takeover -- and he's apparently willing to put up his stakes in both Tesla and SpaceX to secure the deal:

[Star One]

UK and US refuse to sign international AI declaration

https://www.bbc.co.uk/news/articles/c8edn0n58gwo

[InterestedEngineer]

... that long explanation is way over my head!

As far as I can see, nothing in this interaction shows that Grok "knows" that 5th power scaling is the correct answer. It could just be regurgitating the input.

If you give it the exact same prompt with other scaling exponents, does it correct the prompt? 

My bold. 

As it turns out, even the short explanation stumps me, especially now, since "regurgitation" is on the table.

I mean, try "axing" Grok or ChatGPT...

"Given a constant mass ratio of a rocket, show me how the energy rate scales with the 4th power of travel time between two points when using a conventional rocket"

(etc, trying a bunch of different exponents other than "5th")

By "regurgitating," I'm suggesting that the LLMs haven't shown that they actually know the fifth power scaling is correct, since they might just be parroting the prompt back at you.

did you not see the screen snapshot above?  Chatgtp showed its work, it would have made any good  math or physics teacher happy.

[Twark_Main]

... that long explanation is way over my head!

As far as I can see, nothing in this interaction shows that Grok "knows" that 5th power scaling is the correct answer. It could just be regurgitating the input.

If you give it the exact same prompt with other scaling exponents, does it correct the prompt? 

My bold. 

As it turns out, even the short explanation stumps me, especially now, since "regurgitation" is on the table.

I mean, try "axing" Grok or ChatGPT...

"Given a constant mass ratio of a rocket, show me how the energy rate scales with the 4th power of travel time between two points when using a conventional rocket"

(etc, trying a bunch of different exponents other than "5th")

By "regurgitating," I'm suggesting that the LLMs haven't shown that they actually know the fifth power scaling is correct, since they might just be parroting the prompt back at you.

did you not see the screen snapshot above?  Chatgtp showed its work, it would have made any good  math or physics teacher happy.

I understood that. My question was if you don't prompt it with the correct answer first, can it still solve the problem?

And more importantly, if you just give it the post I was responding to, is it smart enough to call attention to the fifth power scaling all on its own with no hints? Or was a human still required in this loop? 

Which part actually takes the most "smarts" in this problem-solving task: some elementary algebra on physics equations (which a human solved and one LLM half-solved), or knowing which elementary algebra was probably being overlooked and needs attention called to it, based on the context of the conversation (a task for which LLMs have zero demonstrated competency AFAIK)?

Obviously the latter is what you really want for your onboard mission AI. Just curious how close we actually are...


I agree though, ChatGPT definitely performed better than Grok! We have no evidence Grok was able to work the math, whether or not it's given the correct answer. At least ChatGPT could, with help, work through a high-school level physics problem. 

[Twark_Main]

From the Grok 3 presentation:

The next [transit window] would be November of next year roughly, and if all goes well SpaceX will send Starship rockets to Mars, with Optimus robots and Grok.

[JohnFornaro]

My question was if you don't prompt it with the correct answer first, can it still solve the problem?

You have to prompt it with the correct answer first?  Am I the only one in this thread who would like to get an answer from [your AI here]?  Because you don't already know what the answer is?

[InterestedEngineer]

My question was if you don't prompt it with the correct answer first, can it still solve the problem?

You have to prompt it with the correct answer first?  Am I the only one in this thread who would like to get an answer from [your AI here]?  Because you don't already know what the answer is?

I'm a little confused about the disagreement here, but I'll note three characteristics of the current LLMs related to this

1. You have to able to ask a question reasonably close to what you want the answer to be.  You can frame it in terms of "here's one way it might work, here's another, please analyze which is more correct for me", which helps.

2.  The LLMs are FAR to eager to please and won't call you out strongly enough on your bad ideas.  You have to literally ask it "why might this idea be wrong" to get any pushback at all.

3.  At least with chatGTP you can share the conversation you had with the LLM via hyperlink so all others can see it. I failed to do so in the example I provided above.

[Twark_Main]

My question was if you don't prompt it with the correct answer first, can it still solve the problem?

You have to prompt it with the correct answer first?  Am I the only one in this thread who would like to get an answer from [your AI here]?  Because you don't already know what the answer is?

I'm a little confused about the disagreement here, but I'll note three characteristics of the current LLMs related to this

1. You have to able to ask a question reasonably close to what you want the answer to be.  You can frame it in terms of "here's one way it might work, here's another, please analyze which is more correct for me", which helps.

2.  The LLMs are FAR to eager to please and won't call you out strongly enough on your bad ideas.  You have to literally ask it "why might this idea be wrong" to get any pushback at all.

3.  At least with chatGTP you can share the conversation you had with the LLM via hyperlink so all others can see it. I failed to do so in the example I provided above.

Well that's rather disappointing.

Circling back to what started this all:

Cross posting

Does that mean that energy is to the fourth power of time decrease?

Yes, if you hold mass ratio constant then the energy scales as the inverse fourth power of travel time.

So if you want to use the same propellant to get you there in half the time, you need 16x as much energy (and of course, 2⁵ = 32x as much power).   

The 2020s version of TL;DR, here's the prompt you feed chatGTP to derive the 5th power

Given a constant mass ratio of a rocket, show me how the energy rate scales with the 5th power of travel time between two points when using a conventional rocket

Alas it's not possible to paste the result here.  Try it yourself. It appears to be a valid derivation.

Note it's the inverse of the 5th power of the travel time.  Even the wrong prompt will get you the right answer.


I need to make a "let me chatgtp that for you" website.

Given the admission above, I would say such a website is premature.

What would ChatGPT have done for me in this instance? Do the math that I had already done myself? 

[InterestedEngineer]

Given the admission above, I would say such a website is premature.

What would ChatGPT have done for me in this instance? Do the math that I had already done myself? 

It would have shown its work ;-)

they either implemented the feature very recently or I just found out about it very recently.

[sanman]

You have to prompt it with the correct answer first?  Am I the only one in this thread who would like to get an answer from [your AI here]?  Because you don't already know what the answer is?

What I've noticed is that Grok goes out of its way to incorporate contents of your past inquiries to it in its answer.
Even several questions later, it may still refer back to something you'd discussed before, even if just in an off-handed way.

I have to admit, I'm liking Grok more than the other LLMs now, because I find its responses so personable.
Anybody else been using Grok much? If so, what do you think of it?

[Ke8ort]

You have to prompt it with the correct answer first?  Am I the only one in this thread who would like to get an answer from [your AI here]?  Because you don't already know what the answer is?

What I've noticed is that Grok goes out of its way to incorporate contents of your past inquiries to it in its answer.
Even several questions later, it may still refer back to something you'd discussed before, even if just in an off-handed way.

I have to admit, I'm liking Grok more than the other LLMs now, because I find its responses so personable.
Anybody else been using Grok much? If so, what do you think of it?

ChatGPT does a similar thing as well, sometimes if I ask it a general question, it’ll offer to apply it to something mentioned earlier.

I’ve messed around with Grok a little bit and I can agree, it is quite personable. It may not be the greatest in raw performance benchmarks but it just has something about its responses that makes it stand out. First time I tried it was on LMArena, and after a good little bit of doing the blind head to head comparisons, I was consistently picking Grok over other models, even ChatGPT 4o and Llama 3.1.

The responses Grok would give when I was trying it out hit that balance of informative without feeling like I was drinking from a firehouse. Sometimes ChatGPT spits out way too much information even for simple responses, and I resort to telling it to explain it to me briefly. Grok felt more like asking someone a question and it giving an appropriately detailed response, which would be good for a more conversational chat with a robotic crew member… HAL

[sanman]

Google has announced its new "AI Co-Scientist" which is supposed to act as a virtual collaborator to assist scientists in accelerating scientific discovery.

https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

Apparently, it only took a couple of days to solve a problem that had taken a research team close to a decade.

https://www.bbc.com/news/articles/clyz6e9edy3o





https://blog.google/feed/google-research-ai-co-scientist/

Scientists who are part of our Trusted Tester Program will have early access to AI co-scientist

When can we have an AI Co-Engineer?

[Star One]

Actually that’s not the whole story. To quote from another forum.

That is the BBC quoting marketing from Google that isn't telling the whole truth.

Professor José R Penadés is the only person seeming to endorse Co Scientist, which is just a version of an LLM aimed at scientific papers.

Whilst Professor José R Penadés had not published his paper, he forgot he had uploaded it a year earlier before official publication and the LLM had access to it's findings. It really wasn't that astounding. Co Scientist can only sift through existing papers, which is actually quite useful, but it certainly can't find up with new discoveries. Very crudely It looks at a lot of data and spits it back out in a different way very quickly.

Notice in the article that there is no quote from him saying he was impressed with the software. What he actually says is that he was shocked to find it came out with something he hadn't published yet (but had uploaded to the internet).

[Star One]

Generative AI’s Greatest Flaw - Indirect Prompt Injection

Apparently this is not truly fixable with current architecture. It also sounds like people with a bit of knowledge will be able to ‘game’ LLMs.

[InterestedEngineer]

Here's a simple example of coming up with an estimate about how much solar power surface area you need to have to grow a square meter of spinach:

https://chatgpt.com/share/67c89410-5da0-8013-9e4f-b65ae00bc6c5

it's roughly 4:1 on Mars, 2:1 on Earth, and probably 1:1 or less in a high enough orbit out of the shadow of Earth most of the time.   The inputs can (and should!) be cross checked with a google search but the math is understandable and check-able by any competent engineer.

about 25% of the 1000t cargo mass budget per 10 colonists is needed to grow potatoes at calorie rates sufficient to sustain life  on mars.   It's mostly energy storage to last 2 week solar outages.

If one needed to jury-rig something on Mars to a first order approximation this is the type of thing that AI can give you answers for in minutes instead of hours, days, and an extremely good expert on various topics available.

[Twark_Main]

Given the admission above, I would say such a website is premature.

What would ChatGPT have done for me in this instance? Do the math that I had already done myself? 

It would have shown its work ;-)

Well what fun exercise would that leave for the rest of the class? 

If you had bothered to ask me (just like you asked her  ) then of course I would have given it, so I don't consider that an (apples-to-apples, or perhaps I should say prompt-for-prompt) advantage to ChatGPT.

Honestly I just thought it was such a trivial thing that it would be practically insulting to make people read a tedious explainer. It would be like writing a long explanation that the rate of change of energy over time in physics is given by power (admit it, your eyes glazed over for a second), or that force times velocity is power.

I just figured everyone would look at the mathematical hazard I had called attention to—a capability ChatGPT still hasn't matched btw, GPT would've blithely let you stumble upon it—picture the exponents for a second, and say to themselves "of course that's right."

And if not, that person could surely benefit from a quick refresher exercise in their elementary physics. Win-win! Everyone got the right meal. You're welcome. 




Humanity giveth, ChatGPT taketh away... 

[t43562]

What you would use AI for in Space?

The Royal Airforce has a plan to use an LLM with the "temperature" settings on low that has been trained on all their maintenance manuals.  This is to help engineers to get answers without having to search the textbook but it tries to ensure they get the same answer every time.

If that works then spaceships will probably have the same sort of thing - an LLM that makes the manual FAR more useful than a big stack of paper could be.  It wouldn't be very creative, just like a more useful search.

AI in general will probably help with spotting things - from debris to asteroids to landing spots.  It might let us put processing work on the spacecraft so we can know about interesting things before we have to download all the images.

I'm not an aerospace engineer but I can imagine they need to be able to know what's going on in the world of research - e.g. materials research and the models will again act like supercharged search engines where one can find required data more intelligently.

The most interesting thing to me, however, is not about using pre-trained models but in sending out models that can learn. Current training methods are pretty expensive - especially with power.  There's an approach called "neuromorphic computing" which uses more sophisticated neurons and learning strategies that are closer to the way real brains work.  These could turn out to be much much more efficient at learning.  This would be interesting because rather than being exhaustively trained and then sent out, our spacecraft and robots might be able to learn as they went along.   e.g a rover might learn how to drive itself better on Mars.

https://spectrum.ieee.org/neuromorphic-computing-2671121824

[Twark_Main]

What you would use AI for in Space?

https://en.wikipedia.org/wiki/List_of_NASA_flight_control_positions

Most of the work can be done by regular software, but there could be a role for LLMs in "operating" that software by handling high-level decisions.