O1’s Chain of Thought: I Built a System to Mimic It-Here’s How It Went!

[@engineerprompt]

Code for the project is up on github: github.com/PromtEngineer/Agent-0.git

[@kai_s1985]

This is a great idea! Imagine having multiple LLMs specialized in different types of skills like reasoning, math, coding, general knowledge, writing, creativity, etc... working in parallel or sequentially, and criticizing each other, and outputting a final agreed answer like ensemble models in random forest or xgboost. Imagine each of those models are better than PhD's in their own area of expertise.

[@shuffle1659]

Thank you for the idea of my next project.

[@engineerprompt]

I think for specialized LLMs, they probably needs to be finetuned for specific task but I think its worth exploring.

[@Juan-n6k3c]

​@@engineerprompt even with RAG? Is finetuning necesary?

[@Anas-dd5ud]

@@Juan-n6k3c yes in my opinion. Rag only adds information to the prompt. but the generation "format" result is much more dependant on the fine tuning. You can notice that if you compare the generation results of instruct finetuned model and the base one.

[@rousabout7578]

Great video and first try! I was waiting for someone to replicate 01 with an agentic system. Controversy aside Matt Schumers reflection 70b has a surprisingly good system prompt. Worth trying on a few models. You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside tags, and then provide your final response inside tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside tags.

[@engineerprompt]

Prompting definitely needs improvement here. The reflection is a tricky one without fine-tuning the model though. Was hoping Matt Schumers will actually just release the data that they created.

[@1littlecoder]

I was working on something similar and here you managed to get it out already! Great work!

[@engineerprompt]

thank you, mean a lot coming from you. big fan of your content. keep up the good work sir.

[@1littlecoder]

@@engineerprompt 🙏🏾

[@HassanAllaham]

and here is a big fat "LIKE" to one of the best channels and one of the best youtubers and prompt engineers. Thanks for the good content 🌹🌹🌹

[@BenoitChabord]

I think your approach is brilliant, and I suppose the same way with other models you could find a COT good at writing, with different agents prompts. You can start to categorise all of the kind of problems and the first step is an agent prompt knows which Multi stage COT worflow it should use.

[@cheaptube4918]

Thanks for putting this together!

[@tollington9414]

Good work, love this approach

[@supercurioTube]

Good job ! It'll be interesting to try it in parallel, from 2 to n and identity if more attempts help or not. Did you leave the temperature to the 1.0 default? To get more diversity of outputs, temperature at the max 2.0 might help.. or not 😅

[@engineerprompt]

The temperature is set to 1.0. I plan to experiment with multiple different LLMs at some point. I think that will give diverse outputs.

[@xXWillyxWonkaXx]

In simple terms, this is a "routing prompt system" aka orchestration or multi-agent prompting.

[@engineerprompt]

Kinda, specially the parallel system will work in that way.

[@epokaixyz]

Consider these actionable insights from the video: 1. Understand Chain-of-Thought Reasoning: Learn how this technique enables AI to solve complex problems by breaking them into smaller, manageable steps. 2. Explore Multi-Stage Systems: Study the concept of using multiple AI agents, each with a specialized role, to enhance the problem-solving process. 3. Learn how to Use Code for Analytical Reasoning: Discover how AI leverages code to efficiently process information and arrive at accurate solutions. 4. Research Solutions to Existing Challenges: Investigate how to address the limitations of Chain-of-Thought reasoning, such as sequential bottlenecks and lack of diversity in AI models. 5. Dive Deeper into Resources: Consult the provided books and research papers to gain a more comprehensive understanding of prompt engineering and Chain-of-Thought prompting.

[@keshavkasat9465]

ai comments

[@nusquama]

@@keshavkasat9465and so ? Question is : is it useful?

[@micbab-vg2mu]

I use my own reflextion prompt with claude 3.5 sonnet that gives me better results in tests than o1:)

[@raphaelchristi6424]

@@micbab-vg2mu you have repo for that?

[@IRazorLeaf]

Could you share a bit more details on what you do or is that top-secret? :)

[@aaron6235]

Yes, share, and don't monetize😮

[@unclecode]

You’ve built a moat for yourself :)) Nice job! The end result of "O1" isn’t exactly new, the "moat" (if there is one) lies in how they figured out bringing the process into inferences within the model itself. I hope they eventually share the details and reasoning tokens. Right now, it could just be like what you did, or maybe they’ve found a way to embed "search" natively during inference.

[@engineerprompt]

Its a mystery but I think it is possible to replicate that behavior with multi-agent system especially if they are specialized. Seems like the Qwen team has something in works. will be interesting to see their approach.

[@unclecode]

@@engineerprompt Qwen became good in math and me too, waiting for that 😎

[@obetishere9215]

Nice. I have my personal project similar to this built using react and using mistral nemo api. My use case is developer, code reviewer, unit tester. I feel like i am missing something but i do not know yet.

[@engineerprompt]

nice, I love when people are solving practical problems with LLMs.

[@aaron6235]

Why didn't you put the link to your repo in the description? I would very much like to clone it

[@engineerprompt]

Wanted to test it a bit more and put up a Readme. There are some failure cases that needs to be highlighted. Will do it soon

[@mirkakonest]

Open Source?

[@engineerprompt]

github.com/PromtEngineer/Agent-0.git

[@raphaelchristi6424]

There's some repo with that?

[@ingenierofelipeurreg]

Pls share repo

[@engineerprompt]

github.com/PromtEngineer/Agent-0.git

[@ahmadrezaatashbartehrani3617]

where is code ? thanks for sharing that

[@engineerprompt]

Look at the pinned comment

[@rajapulau]

10:45 can this solution use autogen?

[@engineerprompt]

Yes, that could be a good option.

[@zeusconquers]

hi, nice work. could you provide the source code please

[@engineerprompt]

github.com/PromtEngineer/Agent-0.git

[@Dr.UldenWascht]

Reflection reborn 😄

[@engineerprompt]

😄😄😄

[@大支爺]

Just review it's own answer and answer it once more time.

[@monstercameron]

this isnt 01 this is gpt4 with code interpreter

[@clray123]

These attempts show your fundamental lack of understanding of what OpenAI did to improve their models. Which is excusable because they no longer publish anything worthwhile and are so vague in their marketing material that it leaves you guessing. But I suspect it went as usual: they created a big mountain of synthetic data demonstrating those "chain of thoughts" on a well defined set of problems, for which reward functions can be defined using RL, and then they trained a very normal LLM on that very mountain of "solved problem" data. What you are missing is not some shitty prompts. You are missing the mountain of training data (and another mountain of compute resources to pull off the training, but it's mostly the data which gives them a competitive advantage).

[@georgearistides7704]

genuine question, how do you know this. I am asking because I built a program to implement chain of thought using the lightweight llama models today, and it does of course give improved responses with respect to amount detail and accuracy. Do you have a source for this claim, and if so can you please drop the link here. I want to try implementing this and possibly combining with what I have already done

[@clray123]

@@georgearistides7704 The source is my knowledge of how LLMs work and what it takes to get better scores on benchmarks. Also that OpenAI did not mention any improvements to model architecture, which they presumably would have done if they had done anything better than provide more training data.

[@georgearistides7704]

@@clray123 source: trust me bro

[@clray123]

@@georgearistides7704 Well, you can go watch AI Explained's review which pretty much does say the same thing after reading 40 pages of their release docs. Believe it or not, these LLMs don't grow on trees and some people out there are fairly well aware how the entire process works.

[@paulocacella]

This simply does not matter. What matters is the output quality goal and cost/benefit. You must show a problem that openai approach is in a way that is superior than agents, RAG and tools in terms of privacy and cost/benefit. That is it. The other aspects simply are not relevant outside a research world. It is a simply as that, you do not need a Ferrari to go to the market. Up to now I can make locally outputs that are better than the approach of o1 on meaningful and useful general problems. When they have something better we can go deeply. Up to now this is only a toy using chatgpt because people to not have the agentic systems setup and have no idea on how to build them.

[@alexpetrov8871]

there is no "thought", there is no "answer": this is a big math function calculating a "next word"

[@StevenSSmith]

and....?

[@alexpetrov8871]

​@@StevenSSmith and what is your question? can you formulate what you are actually asking?

[@ronilevarez901]

Exactly like you.

[@alexpetrov8871]

@@ronilevarez901 >Exactly like you. No, I am quite different. Because to me, “words” are just a language that I use to convey information about the reality that I perceive by analyzing the signals of my senses over time. My senses are imperfect, however, they reflect reality to one degree or another. On the contrary, this LLM has not the slightest connection with reality, all they have is to study this human “language”, a set of phonemes that people use to convey the information that they perceived through their imperfect senses and were able to comprehend and translate into "words". Think about the difference.

[@ronilevarez901]

@@alexpetrov8871 yeah, yeah. The problem is that you and many others would be saying the same about AI if it was embodied and had learned language through real world experience: "it doesn't think". Mhm.