AI learnt "something" from the Physical & Life Sciences Forum.

[sjastro]

Have you thought about using RAG (retrieval augmented generation)?

RAG is largely based on supervised learning from information provided by humans.
Intelligent agents on the other hand train themselves using reinforcement learning without any human interaction.

​

The point of the exercise was to give the intelligent agent minimum information about the image.
Here is the code produced by the intelligent agent in extracting information from the image.

​

GPT-4o did attempt to import a file to read what is known as the FITS - Wikipedia file which contains information along with the image but since it was saved as a JPEG file the information was lost, and GPT had to rely on other resources to extract the information.
While it is nowhere near in depth compared to humans accessing databases, given it was able to identify the brighter galaxies of the cluster and their approximate positions is an impressive feat.

 

[AlexB23]

RAG is largely based on supervised learning from information provided by humans.
Intelligent agents on the other hand train themselves using reinforcement learning without any human interaction.

View attachment 352882
​

The point of the exercise was to give the intelligent agent minimum information about the image.
Here is the code produced by the intelligent agent in extracting information from the image.

View attachment 352883
​

GPT-4o did attempt to import a file to read what is known as the FITS - Wikipedia file which contains information along with the image but since it was saved as a JPEG file the information was lost, and GPT had to rely on other resources to extract the information.
While it is nowhere near in depth compared to humans accessing databases, given it was able to identify the brighter galaxies of the cluster and their approximate positions is an impressive feat.

Whew, this stuff is a little too heavy for me. But, it is impressive how GPT-4o managed to get some info. Do you feel that Meta's Llama 3.1 405B could catch up with GPT-4o, once it has vision capabilities? I use Mistral 7B, as my 8 GB RAM laptop can not run anything > 7 billion parameters.


Meta Llama:

Llama (language model) - Wikipedia

en.wikipedia.org

 

[sjastro]

Whew, this stuff is a little too heavy for me. But, it is impressive how GPT-4o managed to get some info. Do you feel that Meta's Llama 3.1 405B could catch up with GPT-4o, once it has vision capabilities? I use Mistral 7B, as my 8 GB RAM laptop can not run anything > 7 billion parameters.


Meta Llama:

Llama (language model) - Wikipedia

en.wikipedia.org

I have no idea.

While GPT-4o can give impressive answers, LLMs (large language models) can still make serious errors such as this simple maths problem.
What is the smallest integer when squared gives a value between 10 and 50?

The correct answer is -4 as integers can take on positive and negative values.

There are rumours Google will release Gemini combined with AlphaProof and AlphaGeometry 2 which is streets ahead in mathematical ability having achieved a silver medal performance solving 4 out of 6 questions from the International Mathematical Olympiad.

 

[AlexB23]

I have no idea.

While GPT-4o can give impressive answers, LLMs (large language models) can still make serious errors such as this simple maths problem.
What is the smallest integer when squared gives a value between 10 and 50?

View attachment 352918
The correct answer is -4 as integers can take on positive and negative values.

There are rumours Google will release Gemini combined with AlphaProof and AlphaGeometry 2 which is streets ahead in mathematical ability having achieved a silver medal performance solving 4 out of 6 questions from the International Mathematical Olympiad.

Yeah, AI sucks at math. Hopefully, Google can release Gemini with math capabilities. Isn't the correct answer -7, not -4, as -7 squared is 49, which is between 10 and 50?

However, AI could use tools. Folks are working on getting AI to use calculators. Llama 3.1 has the capability to call Wolfram|Alpha, apparently, if one pays for the Wolfram API. Who knows, open source tools could be developed in the future.

 

[sjastro]

Yeah, AI sucks at math. Hopefully, Google can release Gemini with math capabilities. Isn't the correct answer -7, not -4, as -7 squared is 49, which is between 10 and 50?

However, AI could use tools. Folks are working on getting AI to use calculators. Llama 3.1 has the capability to call Wolfram|Alpha, apparently, if one pays for the Wolfram API. Who knows, open source tools could be developed in the future.

Yes the answer is -7, l should have paid more attention to the question.
GPT-4o is variable when it comes to 'mathematical abilities'.
In another thread I found it could do calculus if I typed the question instead of scanning the question.
It struggles if the input is presented as an image file.

 

[AlexB23]

Yes the answer is -7, l should have paid more attention to the question.
GPT-4o is variable when it comes to 'mathematical abilities'.
In another thread I found it could do calculus if I typed the question instead of scanning the question.
It struggles if the input is presented as an image file.

Yeah, so true. AI is not there yet. And yes, I got the math right

 

[AV1611VET]

AI learnt "something" from the Physical & Life Sciences Forum.

Who's Al?

 

[AlexB23]

Who's Al?

AI stands for Artificial Intelligence.

 

[AV1611VET]

AI stands for Artificial Intelligence.

 

[AlexB23]

Yep, though the capital letter "I/i" looks like a lowercase "L\l" on most sans serif fonts.

 

[sjastro]

For my final AI test (for a while) this one might prove even a challenge to a human expert.

The image was taken through a grating filter used on my telescope and records the spectra of objects in the field which are on the right hand side of the object.
The object of interest is PGC41121 also known as 3C 273 - Wikipedia which is a quasar at a distance of 2.4 billion light years and due to its extreme faintness took a six hour total exposure to get a sufficiently high S/N ratio to produce the resulting spectrum.

My question to GPT-4o was to find the redshift of the object and identify the type of object using the spectrum.
I imagine this would also be extremely different for a human expert to answer since the peaks have been redshifted and there is no reference point such as using a laboratory spectrum which is in the rest frame.

This was GPT-4o's response which took a couple of days to complete as I had used up my runtime.

As expected GPT-4o got the wrong answer so I provided it more information which is in the next post......

 

[sjastro]

Even though GPT-4o got the wrong answer it did 'recognize' there could be red shifted hydrogen peaks in the spectrum and also correctly identified their wavelengths in a rest or laboratory frame.
So I provided more information in the spectrum.

​

There are two questions for GPT-4o here whether it could understand the significance of the extra information and addressing the more difficult question of what type of object produces this spectrum.
Here was its response.

On the basis of the extra information GPT-4o correctly calculated the redshift of z = 0.158.
Quasars are not star forming galaxies but they are class of AGN which GPT-4o identified as a possible alternative.
My overall impression of GPT-4o is favourable.

 

[AlexB23]

For my final AI test (for a while) this one might prove even a challenge to a human expert.

View attachment 353048​

The image was taken through a grating filter used on my telescope and records the spectra of objects in the field which are on the right hand side of the object.
The object of interest is PGC41121 also known as 3C 273 - Wikipedia which is a quasar at a distance of 2.4 billion light years and due to its extreme faintness took a six hour total exposure to get a sufficiently high S/N ratio to produce the resulting spectrum.

View attachment 353049​

My question to GPT-4o was to find the redshift of the object and identify the type of object using the spectrum.
I imagine this would also be extremely different for a human expert to answer since the peaks have been redshifted and there is no reference point such as using a laboratory spectrum which is in the rest frame.

This was GPT-4o's response which took a couple of days to complete as I had used up my runtime.

View attachment 353050

As expected GPT-4o got the wrong answer so I provided it more information which is in the next post......

Wow, your telescope can spot a quasar that is over 2 billion light-years away? That is far away. Thought only space telescopes and powerful ground based scopes could make out quasars.

 

[sjastro]

Wow, your telescope can spot a quasar that is over 2 billion light-years away? That is far away. Thought only space telescopes and powerful ground based scopes could make out quasars.

An amateur sized telescope will reveal the quasar as a spot of light.
The challenge is when using a diffraction grating filter on the CCD requires very long exposures to get a sufficiently high signal to noise ratio to separate out the spectrum information from the noise.

 

[AlexB23]

An amateur sized telescope will reveal the quasar as a spot of light.
The challenge is when using a diffraction grating filter on the CCD requires very long exposures to get a sufficiently high signal to noise ratio to separate out the spectrum information from the noise.

That is cool. How does diffraction grating work.

 

[sjastro]

That is cool. How does diffraction grating work.

Here is a simple explanation.

 

[Hans Blaster]

Even though GPT-4o got the wrong answer it did 'recognize' there could be red shifted hydrogen peaks in the spectrum and also correctly identified their wavelengths in a rest or laboratory frame.
So I provided more information in the spectrum.

View attachment 353051
​

There are two questions for GPT-4o here whether it could understand the significance of the extra information and addressing the more difficult question of what type of object produces this spectrum.
Here was its response.

View attachment 353052

On the basis of the extra information GPT-4o correctly calculated the redshift of z = 0.158.
Quasars are not star forming galaxies but they are class of AGN which GPT-4o identified as a possible alternative.
My overall impression of GPT-4o is favourable.

When you said it was a quasar, I cheated and guessed that the big line red of H-alpha was H-alpha, made a crude estimate of the peak wavelength and then tried to remember the wavelength for H-alpha. I got z = 0.17. (I cannot quote any more digits as I can't claim the requisite precision. )

The actual value on the plot falls very close to 7600 and if I had gone with 6550 for H-alpha, it would have been a lot closer. ( I knew it was 655-something, but it did feel like it was close to 6550. I just didn't feel I could justify more precision so I "rounded" to 6500. [6563 A, I need to remember that. I should, as I once measure the difference between the H-alpha and D-alpha lines to 0.01 A precision in a lab class.] )

Your python looks pretty familiar, except for the image handling stuff. I mostly make line plots and it always takes a lot of trial and error to find the right way to generate an image.

 

[AlexB23]

Here is a simple explanation.

Wow, so this is the same tech used to determine the structure of DNA, but with visible light instead of X-rays?

 

[sjastro]

Wow, so this is the same tech used to determine the structure of DNA, but with visible light instead of X-rays?

The professionals use the same techniques for spectroscopy but the technology is vastly more sophisticated than the grating filters we amateurs use.

ESO - Overview

ESO is the European Organisation for Astronomical Research in the Southern Hemisphere. It operates the La Silla Paranal Observatory in Chile and has its headquarters in Garching, near Munich, Germany.

www.eso.org

 

[sjastro]

Another test for AI on a spectrum of a very rare type of star known as a Wolf–Rayet star - Wikipedia.
The subject is a binary star system identified as a Gamma Velorum where the brighter component Suhail has a distinctive spectrum as shown in my image.

​

GPT-4o was asked to identify what type of object produced this spectrum and if possible identity the star.

Here is the full detail.

It correctly identified the object as a Wolf-Rayet star, the class WC8 and the carbon emission lines but missed the Hα peak and could not specifically identify the object as Suhail.
Still quite an impressive feat.