Bro it is an absolute pitty that you made this awesome video with such a extremely shity audio quality?!!!! my friendly advice is spens just a 20$ on a good quality michrophone, you yourself are trashing your product with this awfull microphone what the heck was that really? have you even once listen to it afterward? it goooos beyond my nerve when I see impressive works are just fucked up only for 20$ microphone, come on man what the heck?!!!!!!
7:10 oh, so the "revolutionary" part of alphageometry is not using an LM for the rabbit out of the hat or the neuro-symbolic engine, but the synthetic data generation process that allows for such a task to be completed. neat.
I heard somebody say something recently along the lines of "soon we'll be able to juggle around concepts like math and physics and use them like we use objects". This seems to provide a glimpse into that future. I think most problems can be rephrased as a geometric problem and understood much more nicely with human spatial intuition. This is a huge step forward in human interpretation of machine logic.
Since AlphaGeometry is open-source, Can I run alphageometry on my personal computer? Will it work? If yes, To what do I need to set my BATCH_SIZE, BEAM_SIZE and DEPTH?
Just got it to run! First observation is how tiny it is! I've had a suspicion that a well- crafted mathematical model could be small and highly quantized. I think you are demonstrating that!
Since you got it to run, Can I run alphageometry on my personal computer(Ram: 8GB) ? Will it work? If yes, To what do I need to set my BATCH_SIZE, BEAM_SIZE and DEPTH?
Very cool work and great explanation. Sounds to me like keep reaching into the hat until what you pull is a rabbit. Which is like a human doing trial and error and tinkering around with a problem. Excited to see where else this leads.
Great work! Thank you for taking the time to make it accessible for the rest of us. Looks like you've found a way of generating pages straight from 'The Book' as Erdos would say.
Hey Trieu Awesome Work! Doing this with 150M parameter Model seems to be Amazing! Will be going through the paper. But few things: 1. If I understand correctly, there is still manual intervention to convert the natural language to mathematical language right. If yes, is there a way to avoid this manual step as well 2. Also neural networks are meant to avoid symbolic reasoning right. So why not approach it completely from neural network perspective. Can Bigger neural networks solve this without any symbolic AI. 3. nit pick: At 3:05, should it be 9^(x+2)?
So, what exactly is Euclid's rabbit in Ex: 1? Is it an understanding of the nature of prime numbers? If so, where does it come from? This particular proof in Elements is in Book IX, prop 20. To get here you have to do a lot of work. And if you do the work, you might have conditioned your mind enough where the rabbit isn't so strange at all.
I guess the first rabbit is that you decide to do the proof by induction, and the second rabbit is the construction of the new prime. If you have access to these rabbits, the rest is just easy mechanical applications of simple proof steps. (EDIT: well, it's not really induction but that's not the point EDIT2: thinking about it, maybe this is not a rabbit, as it's basically the problem statement) "And if you do the work, you might have conditioned your mind", and this conditioning is exactly what makes rabbits possible, the equivalent to "conditioning your mind" in AlphaGeometry is the training of the LLM on synthetic proofs.
@@andrea-mj9ce Tbh, I don't have fully grasped the process, and I may have misunderstood things. How I understand so far what they are doing to get a single synthetic proof: that they generate a bunch of initial facts. Using these facts, they use the symbolic engine DDAR to deduce a random new fact based on these initial facts. So the new synthetic theorem would read "from follows that ", and the proof would be the tree of steps that DDAR found. Now, sometimes it happens that the initial facts contain a statement, that is implied by the other facts. For example, if the initial facts are ["a
Would be interesting to see how this works when you connect this to the Wolfram Language, so you have access to a very powerful symbolic solver. Have you considered doing this?
Bro it is an absolute pitty that you made this awesome video with such a extremely shity audio quality?!!!! my friendly advice is spens just a 20$ on a good quality michrophone, you yourself are trashing your product with this awfull microphone what the heck was that really? have you even once listen to it afterward? it goooos beyond my nerve when I see impressive works are just fucked up only for 20$ microphone, come on man what the heck?!!!!!!
Such passion.
This is not entertainment, if you wanted entertainment look elsewhere
rAbbIT Out oF My ASSSssss!!! MicROPhoNE iS cRAApppp!!!!!!one!!!! 😂
Audio was fine for me
7:10 oh, so the "revolutionary" part of alphageometry is not using an LM for the rabbit out of the hat or the neuro-symbolic engine, but the synthetic data generation process that allows for such a task to be completed. neat.
Most impressive work in 2024. He must be a leading researcher brining AGI to our civilization.
Impressive and important work! And your walkthrough is so clear. Thanks!
Listening to this while downloading the repo. The next few days are going to be interesting.
Thank you for your work on this project!
I heard somebody say something recently along the lines of "soon we'll be able to juggle around concepts like math and physics and use them like we use objects". This seems to provide a glimpse into that future. I think most problems can be rephrased as a geometric problem and understood much more nicely with human spatial intuition. This is a huge step forward in human interpretation of machine logic.
Humans have always been able to do that, you're late to the party
Very concise talk! Impressive work!
Một nghiên cứu tuyệt vời! Cảm ơn anh!
Since AlphaGeometry is open-source, Can I run alphageometry on my personal computer? Will it work? If yes, To what do I need to set my BATCH_SIZE, BEAM_SIZE and DEPTH?
Thank you for your contribution to the field of AI. Making my dream of automated scientific and technological advancement a little closer to reality!!
What? Just a little closer?
Nice work! The idea to combine "make a guess engine" with a deterministic solver is clever, could be used in other domains,
Impressive! Great work, looking forward to the future!
Thank you for inspiration! Will dive deeper in details and looking forward to further research!
Great work, and amazing presentation!
The rabbit-out-of-the-hat is like the human intuition that selects potential chess moves.
Just got it to run! First observation is how tiny it is! I've had a suspicion that a well- crafted mathematical model could be small and highly quantized. I think you are demonstrating that!
Since you got it to run, Can I run alphageometry on my personal computer(Ram: 8GB) ? Will it work? If yes, To what do I need to set my BATCH_SIZE, BEAM_SIZE and DEPTH?
@@monaluthra4769 I ran it on a fractional cloud A40 GPU, 8GB vRAM, 10GB RAM
Very cool work and great explanation. Sounds to me like keep reaching into the hat until what you pull is a rabbit. Which is like a human doing trial and error and tinkering around with a problem. Excited to see where else this leads.
Great work! Thank you for taking the time to make it accessible for the rest of us. Looks like you've found a way of generating pages straight from 'The Book' as Erdos would say.
This is amazing! Probably going to be my favourite paper of the year :)
Parsing and understanding the problem statement is probably harder than solving the problem once understood.
Hey Trieu Awesome Work! Doing this with 150M parameter Model seems to be Amazing! Will be going through the paper. But few things:
1. If I understand correctly, there is still manual intervention to convert the natural language to mathematical language right. If yes, is there a way to avoid this manual step as well
2. Also neural networks are meant to avoid symbolic reasoning right. So why not approach it completely from neural network perspective. Can Bigger neural networks solve this without any symbolic AI.
3. nit pick: At 3:05, should it be 9^(x+2)?
So, what exactly is Euclid's rabbit in Ex: 1?
Is it an understanding of the nature of prime numbers? If so, where does it come from?
This particular proof in Elements is in Book IX, prop 20. To get here you have to do a lot of work. And if you do the work, you might have conditioned your mind enough where the rabbit isn't so strange at all.
I guess the first rabbit is that you decide to do the proof by induction, and the second rabbit is the construction of the new prime. If you have access to these rabbits, the rest is just easy mechanical applications of simple proof steps. (EDIT: well, it's not really induction but that's not the point EDIT2: thinking about it, maybe this is not a rabbit, as it's basically the problem statement)
"And if you do the work, you might have conditioned your mind", and this conditioning is exactly what makes rabbits possible, the equivalent to "conditioning your mind" in AlphaGeometry is the training of the LLM on synthetic proofs.
@@dariusduesentrieb Agree, each time you begin the proof with the word "suppose" you are entering rabbit territory
@@dariusduesentrieb Can you explain how the synthetic proofs were generated? I don't get it
@@andrea-mj9ce Tbh, I don't have fully grasped the process, and I may have misunderstood things. How I understand so far what they are doing to get a single synthetic proof: that they generate a bunch of initial facts. Using these facts, they use the symbolic engine DDAR to deduce a random new fact based on these initial facts. So the new synthetic theorem would read "from follows that ", and the proof would be the tree of steps that DDAR found. Now, sometimes it happens that the initial facts contain a statement, that is implied by the other facts. For example, if the initial facts are ["a
@@andrea-mj9ce Did you by any chance get my reply? It doesn't seem to show up for me on RUclips.
Such a good presentation, thank you!
Try proving pythagorean theorem, how many leafs would it take?
3
I still don't get how the theorem-proof pairs were generated
Would be interesting to see how this works when you connect this to the Wolfram Language, so you have access to a very powerful symbolic solver. Have you considered doing this?
Congratulations on your fantastic work. Will you publicly release the synthetic dataset?
Great stuff man
You're Great!
From Viet Nam with love and proud
Great video! Thank you for making it, I think it is a great complement to the paper.
Classical logic constructive logic, is not a two way arrow. There are classical theorems without constructive proofs, so the two way arrow is wrong.
increíble
😁
very inspiring!
Pulling out the rabbits from the hat, i.e., Generating insights, thats the exciting part
Can you share the PPT of this video
Türkçe çeviri istiyorum 😢
WHAT A LAD!
hy vọng một ngày nào đó pro như m