Not a graduate level problem, any first/second-year physics undergrad should be able to chug through this easily

Boring

Thank you for the video! The cynic in me says that the original version of the problem might be on these models' training data and that's enough for them to solve it. So i'd like to try these out with a novel problem from one of my upcoming exams. I dont have access to o1 pro though...

American products always cost too much.

Your content makes me love mathematics and partial differential equations again. ❤

prononciation is "diriclay"

These "tests" are absolutely useless and do not paint a picture That helps u compare these things. Same with all the benchmarks he is referring to.

1337 and 69421... I see what you did there!

Capitalize on DeepSeek for real that’s free views right now

What people seem to be forgetting is that everyone is making technology improvements and especially when they're open like this, others can also build from them. If you consider this makes training and inference require only 1/45th of the hardware, then in practice what this really means is OpenAI will copy some of these techniques and end up with the equivalent of 45 times more hardware than they currently have. It effectively takes their $500B StarGate project and supersizes it to the equivalent of a $22 trillion HyperGate. Their bigger budget still ultimately allows them to do more. This in no way reduces the need for NVidia or similar hardware, it just means we will do more with the same amount and accelerate more quickly -- for better or worse.

Image generation with temp 1? That's maximum creativity…

People bailing on NVidia is like people during the industrial revolution being like "See those English guys made a smaller, more powerful and more efficient steam engine? They're not going to need nearly the amount of iron to make these things - I'm bailing on my iron mining stocks!" - what is it about a better, cheaper AI that will cause NVidia to sell less tech? Better, cheaper AI will result in more people using it - more servers being made and used - thus more need for the chips that run them. NVidia stock is going to be just fine. (not financial advice... don't sue me)

Hi Kyle, enjoy your videos and like the variety of video run times - would like to see more testing as carried out in this video and, if possible, with challenges that are not on the web / not possible to have been used in any training data.

Nice Haircut, looking fresh😎

now generate 100x more synthetic data and increase parameter count by another 100x

crazy times this year started nicely 💯👍

This architecture is interesting. The current results are not. Maybe it will be really good after some time... but not now.

Hi, Kyle, please excuse if this question is silly, but I'm confused about the boundary conditions in the problem: If the cube has a fixed potential at top and bottom, and zero on the sides, isn't there some contradiction at the edges and corners, where the top/bottom and sides meet? For example, what is the potential supposed to be at the origin? z=0 says a non-zero potential, but x=0 or y=0 says zero potential.

I actually witnessed it's reasoning capabilities today I basically had this high school level problem (yeah me in highschool) And, to test it's inference and logic, I just described the situation to it, with very little hints. Like, the problem couldn't be understood by it's description. It just understood what the question was about, tried a bunch of different ways to solve it, then checked the options I had provided, tried all the goofy approaches, and just said, " I am speculating but this might be the answer, as all other answers don't quite fit" It felt so human, It guessed the right answer, tho it wasn't confident about it. That's intelligence, unlike other models who literally lie with confidence. Crazy!

I love this shorter format! Very impressive results!

I am waiting for day when it will solve any problems and it will make avg person to genius

The gemini model works around 1 temperature the best for my evals that involve a bit of calculation and coding

Can't wait to see you test o3 mini!

More videos like this

You've been busy, Kyle.. lol

bro for google o1 try to increase the temperature and you will be surprised

numbners hard. cant brain.

Even the running 8b model locally is good enough to solve average math and physics problems 🥲

Google was so close there, and faster. I wonder why it was an order of magnitude faster.

Can you make a video teaching me how to verify different expressions of partial differential equations in Mathematica to check if they are actually equivalent? For example, using numerical methods or graphing. I am very interested in this technique and am just curious. Ones that also involved summation notation,Thank you!

Plot twist theres a large team of chinese scientists at the other end of the line writing out the deepseekR1 answers.

Also Gemini 2.0 thikingking beat Deepseek r1 on SimpleBench he got 7/10 which is very big improvement. And o1 only get 5/10 if i remember

Waiting for Gemini 2 Thinking regular version, non-flash

I gave deepseek a geenral relativity test, It correct solved the Christoffel symbol of the metric g_tt = (c^2 - (2GM/r)) g_tr = 0 g_rt = 0 g_rr = ((-1)/(1-(2GM)/(rc^2))) Metric signature (+,-) It calculated all christoffel symbols corrected. It also inject them into the geodesic equation correctly in the second prompt. However when I telling it to calculate the Christoffel symbols, then Riemann tensor, the Ricci tensor, it calculated it wrong, probably because too many heavy problems all at once, but that still very impressive! Here is its answer The non-zero Christoffel symbols for the given metric are: \[ \boxed{ \begin{aligned} \Gamma^t_{tr} = \Gamma^t_{rt} &= \frac{GM}{r^2 \left(c^2 - \frac{2GM}{r} ight)}, \\ \Gamma^r_{tt} &= \frac{GM}{r^2} \left(1 - \frac{2GM}{r c^2} ight), \\ \Gamma^r_{rr} &= -\frac{GM}{c^2 r^2 \left(1 - \frac{2GM}{r c^2} ight)}. \end{aligned} } \]

no, it has been verified by looking the the released models and the paper and doing the math

Could you do further physics tests on this new site?

Thank you kyle! Love to see more like these!

Very cool 😮

First

What a coincidence bro I was just using that AI when this video came out😅

OpenAi try to slow down and squeeze more profit from there Models And deepseek doesn't care about money and they just keep accelerating 😅

It's wasn't a bad idea to pull out, and then get back in at the bottom, which is probably soon if not yesterday.

Soon people wont be using their brains, they wont be able to even think for themselves it's not looking good for the future, ever seen that movie called idiocracy that's where their heading.

I miss the original yugioh

Now you can go to bed.

can you test same problems on deepseek?

hahaha man that book… that book is evil. im gonna try and tackle it again with gpt help

PLEASE translate all your videos

Hey i have a really important question i want to ask about your PHD... Before you got your PHD in Physics or whatever, how did you managed distractions and mind wandering? Especially in your teenage years how did you avoided distractions? Whats your moto for motivation and if you could not stick to your motivation how did you get back on track??

Hi Kyle! Could you show us how to create your own LLM using big LLM specially costumed to solve physics problems?