This is a particularly difficult problem for LLMs because each stage in the pattern is independent, so figuring out a potential solution for one stage doesn't provide any clues for the next. Furthermore, equations like 7x + yz = 9, 9x + yz = 12, and 12x + yz = 48 have many solutions. Humans tend to search for patterns with small, manageable numbers, assuming the solution will be elegant and concise, but LLMs are not inclined to choose numbers based on these conditions. It would be a smarter approach for an LLM with access to external tools to formulate this problem with an additional constraint (e.g. x^2+y^2+z^2
I freaking love it, damn good coder yeah, but, I can talk to GPT-4o for hours defining a project, then have o1 read over everything and consolidate the most practical ideas 🤯so handy, the new voice features make it so easy to talk to as well.. I think Altman might be right about the whole Chat bots < reasoners < agents < innovators < super intelligence progression ..!
Here's a concise overview of what you need to know: 1. Think of O1 as a research assistant you can collaborate with. 2. Provide O1 with context, clues, and guidance rather than just questions. 3. Embrace O1's unconventional thinking processes, as they can lead to breakthroughs. 4. Experiment with different clues and prompts, observing O1's responses. 5. View O1 as an active collaborator in problem-solving, not just a tool for answers.
I have e theory. I am lazy, thats why I am not testing it. 1. Give a base prompt to any llm e.g. 3.5 opus ( why not because its one of the best model) and ask if you think this is complex task is hard give me step by step guide to solve it. Don't give me solution just a guide to solve it. 2. take that guide and input it again {base} + {guide} + do you think these steps are going in a right direction. If not tell me how to solve. and give me new step by step fixed guide. [This is self-critique recently open AI showed self-critique can improve response. 3. Base prompt + { new fixed guide} + are there any other options to solve this. 4. Run those 3 steps until there are not other way to solve that problem. 5. Choose the best and solve it. Because gpt 4 internally creating a lot of response thats why it is so much expensive. And they can't show they real thinking process cz if they do anyone can know how this is working. I think this strongly suggests my theory is right.
I did that :) with Gemini, Llama -> It's an improvement. But not as close as o1. In fact I was planning to make a video on it. But didn't do as the results weren't great. I'll see if I still have the code somewhere.
@@1littlecoder I found that it doesn't follow the instruction quite well. I think this where reinforcement learning comes into play. They probably created a new model that can give guide and follow that guide very well.
claude cannot find B as the answer. but if i tell it that B is the answer, this is what i get: 7 + 2 = 9 9 + 3 = 12 (previous operation + 1) 12 * 4 = 48 (switch to multiplication, previous number + 1) 48 + 124 = 172 (return to addition, where 124 = ((48 + 4) * 2) + 20) 172 + 718 = 890 (continuing the pattern, where 718 = ((172 + 124) * 2) + 20)
Not from India but these types of questions are kind of pseudomath more like IQ puzzles that test creativity to try various path than skill in math. From a math perspective: 7, 9, 12, 48, ?, 890 Given: f(1) = 7 f(2) = 9 f(3) = 12 f(4) = 48 f(5) = ? f(6) = 890 5 constraints so solution should involve
I prompted Given f(1) = 7, f(2) = 9, f(3) = 12, f(4) = 48, f(6) = 890 find f(5) Hint: set up a recurrence relation that finds f(n) as a function of f(n-1) and n, do not forget about operations like (-1)^n for the Wolfram GPT and it tried somethings but finally gave up and said (48+890)/2 = 469
It got it correct with this prompt that gave it the format of the relation: Given f(1) = 7, f(2) = 9, f(3) = 12, f(4) = 48, f(6) = 890 find f(5). A) 594 B) 172 C) 560 D) 343 Hint: set up a recurrence relation that finds f(n) as a function of f(n-1) and n, do not forget about operations like (-1)^n Consider f(n)=f(n−1)⋅(n−1)+(−1) ^ n ⋅(an ^ 2 +bn + c) set up equations to solve for a, b, c then find f(5)
Man now I have to learning coding and Building agent also ahh man one more reason to earn money to buy a good laptop or pc or work ass of to get in such company atleast
At this point it's just engineering and not real innovation, what's more concerning is OpenAI has this habit of anthromorphizing everything that they build, is the model really thinking?
Nandri for Shoutout Man🙏🏻!!! Didn’t expect the video to drop this soon. Great Analysis of O1 preview’s multifaceted deterministic approach✨
Thank you for sharing the problem, That sparked my interest to dig deeper and which in turn helped me turn into a video!
This is a particularly difficult problem for LLMs because each stage in the pattern is independent, so figuring out a potential solution for one stage doesn't provide any clues for the next. Furthermore, equations like 7x + yz = 9, 9x + yz = 12, and 12x + yz = 48 have many solutions. Humans tend to search for patterns with small, manageable numbers, assuming the solution will be elegant and concise, but LLMs are not inclined to choose numbers based on these conditions. It would be a smarter approach for an LLM with access to external tools to formulate this problem with an additional constraint (e.g. x^2+y^2+z^2
I freaking love it, damn good coder yeah, but, I can talk to GPT-4o for hours defining a project, then have o1 read over everything and consolidate the most practical ideas 🤯so handy, the new voice features make it so easy to talk to as well.. I think Altman might be right about the whole Chat bots < reasoners < agents < innovators < super intelligence progression ..!
Here's a concise overview of what you need to know:
1. Think of O1 as a research assistant you can collaborate with.
2. Provide O1 with context, clues, and guidance rather than just questions.
3. Embrace O1's unconventional thinking processes, as they can lead to breakthroughs.
4. Experiment with different clues and prompts, observing O1's responses.
5. View O1 as an active collaborator in problem-solving, not just a tool for answers.
OpenAI o1 has transformed my journey as a CSE student!
that's what I like to hear, I just signed up for Engineering 😆
Enjoyed thoroughly.
So glad!
Can you share what tools to use fully end to end full stack applications especially for participating in hackathons?
2025 is going to be the year of CoT
I have e theory. I am lazy, thats why I am not testing it.
1. Give a base prompt to any llm e.g. 3.5 opus ( why not because its one of the best model) and ask if you think this is complex task is hard give me step by step guide to solve it. Don't give me solution just a guide to solve it.
2. take that guide and input it again {base} + {guide} + do you think these steps are going in a right direction. If not tell me how to solve. and give me new step by step fixed guide. [This is self-critique recently open AI showed self-critique can improve response.
3. Base prompt + { new fixed guide} + are there any other options to solve this.
4. Run those 3 steps until there are not other way to solve that problem.
5. Choose the best and solve it.
Because gpt 4 internally creating a lot of response thats why it is so much expensive. And they can't show they real thinking process cz if they do anyone can know how this is working. I think this strongly suggests my theory is right.
I did that :) with Gemini, Llama -> It's an improvement. But not as close as o1. In fact I was planning to make a video on it. But didn't do as the results weren't great. I'll see if I still have the code somewhere.
@@1littlecoder I found that it doesn't follow the instruction quite well. I think this where reinforcement learning comes into play. They probably created a new model that can give guide and follow that guide very well.
claude cannot find B as the answer. but if i tell it that B is the answer, this is what i get:
7 + 2 = 9
9 + 3 = 12 (previous operation + 1)
12 * 4 = 48 (switch to multiplication, previous number + 1)
48 + 124 = 172 (return to addition, where 124 = ((48 + 4) * 2) + 20)
172 + 718 = 890 (continuing the pattern, where 718 = ((172 + 124) * 2) + 20)
Not from India but these types of questions are kind of pseudomath more like IQ puzzles that test creativity to try various path than skill in math.
From a math perspective:
7, 9, 12, 48, ?, 890
Given:
f(1) = 7
f(2) = 9
f(3) = 12
f(4) = 48
f(5) = ?
f(6) = 890
5 constraints so solution should involve
I prompted Given f(1) = 7, f(2) = 9, f(3) = 12, f(4) = 48, f(6) = 890 find f(5)
Hint: set up a recurrence relation that finds f(n) as a function of f(n-1) and n, do not forget about operations like (-1)^n
for the Wolfram GPT and it tried somethings but finally gave up and said (48+890)/2 = 469
It got it correct with this prompt that gave it the format of the relation:
Given f(1) = 7, f(2) = 9, f(3) = 12, f(4) = 48, f(6) = 890 find f(5).
A) 594
B) 172
C) 560
D) 343
Hint: set up a recurrence relation that finds f(n) as a function of f(n-1) and n, do not forget about operations like (-1)^n
Consider f(n)=f(n−1)⋅(n−1)+(−1) ^ n ⋅(an ^ 2 +bn + c)
set up equations to solve for a, b, c then find f(5)
Man now I have to learning coding and Building agent also ahh man one more reason to earn money to buy a good laptop or pc or work ass of to get in such company atleast
At this point it's just engineering and not real innovation, what's more concerning is OpenAI has this habit of anthromorphizing everything that they build, is the model really thinking?
Ai is dying.no chatgpt can do what a normal human being can do
o1 sucks !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
If you liked o1 Preview, you'll absolutely love o1.
❤preety cool insightful video🫡
Glad you liked it