วีดีโอ

Impressed with your bit on self reinforcing systems find it difficult to break out of their viewpoint, nice talk John

I've only just started this video but...All from my memory so accuracy may be variable: 1990s ALN (Artificial Logic Networks), University of Alberta, Canada I think. Nodes, irc, are trained to be AND, LEFT, RIGHT of 2 inputs. There was some success in work with tar sands and some sort of imminent mechanical failure. I started looking into this in the adaptive optics field with the idea of FPGA implementation. I didn't get very far as it wasn't what I should have actually been working on.

Thanks so much for delving deeper into this, wish you a Happy New Year from Italy.

Great talk, thank you!

All the Ai Chat apps are full of glitches o1 included. The problem is you don't get a vivid presentation of the mistakes and glitches like you do with generative images. You need to be well versed in a subject and use chatApps a fair bit to tell how far off the market they still are. Very synthetic algorithmic patterns in its writing, and you can't really get away from that.

53:08 those two phrases don't appear in the article?

Maybe some tree search is done during training in the same style as stream of search (arXiv:2404.03683) by training on serialized suboptimal search strategies where errors are potentially recovered through backtracking which might overcome error accumulation, but during inference it seems implausible to me for o1 at least as for o1-pro it seems different (see 17th dec AMA session from OpenAI's API team) probably using multiple samples. And I don't think a reward model is used during inference neither, as there is QwQ 32B-Preview that is open weight reasoning model and has similar performances as o1-preview/o1-mini on AIME, MATH and LiveCodeBench and it's the same architecture and inference code (with different default temperature, top_p,..) as the standard llm Qwen2.5-32B you can try it in huggingface. All this points to the fact that reasoning models are likely just llms with a custom post-training.

Part 1 here: th-cam.com/video/-6J0S1q03Ds/w-d-xo.html

Part 2 here: th-cam.com/video/f5obaHiOog4/w-d-xo.html

Where did OpenAI say they didn't use tree search? I think they do use tree search, specifically MCTS for generating the synthetic data for o1, then at inference time they don't use tree search. The magic is in creating the synthetic data - they take a variety of paths including some wrong paths of the tree search and chain those with keywords like "but wait, the above is getting me stuck. Let's try this instead" then jump to another branch (the branch frequently does lead to the correct answer) of the tree. The key is MCTS + "let's verify step by step" in my opinion, so they linearize the MCTS thoughts chains and train on that. Somewhere in there they're using RL also as another key ingredient. Looking forward to hear your thoughts

I tried o1 with moderately complex questions regarding solar astronomy and o1 completely fell apart. It was useless. I would point the contradictions in the answers and it kept apologizing.

Prompt that makes 4o behave like o1: ``` [Problem] Do it out by the following format, taking care to reflect, verify, clarify all assumptions: ###Thoughts### <Subtask 1 Name> <Subtask 1 Result> <Subtask 2 Name> <Subtask 2 Result> ... ###Final Answer### <Final Result> ```

Great content, thanks so much for sharing all of your videos!

What is the purpose of generating synthetic data from the model which would be used to improve itself? Wouldn't the synthetic data it produced contain the exact same biases as the model? How do you remove the inherent bias? More importantly, if it can produce expert data, why would it be used to fine-tune itself over it again considering the model was already able to produce the very same data? Does this feel like CoT or ReAct with extra steps?

AgentJo GitHub repo here: github.com/tanchongmin/agentjo

Great work and very practical too, kudos 👍 How much more time for going from agentjo to agentjohn? LOL

Check out AgentJo GitHub here! github.com/tanchongmin/agentjo

I like this more. But how does the tech differ from hypothes.is ?

Thank you John for this explanation! Will try it out!

i am quite interested to know, the distribution of XOR/XNOR distribution here. They are very good with non linearity classification. after the training is complete, does the distribution of xor/xnor seem higher than other gates?

Can't wait to see how this will revolutionize math

may you provide code pls?

Since the connections are randomly initialized and then fixed, so it's impossible to guarantee a 100% acc even the data is guaranteed no noise?

i hear the size being of concern, but i feel the comparison is a bit off: while the number of neurons in the logic gate net scale n^2 with the input size, where as in classic neural net scale linear. the number of weights in a classic FC-NN also scale quadratically with the number of neurons. so the learnable parameter count scales quadratically in both. the logic gate net when drawn on paper may look larger, but it wont require more memory i think. in the paper they mention that they just save the random connections in form of a seed number, so that does not require any space

There’s no choice, but it feels like too many interfaces have changed. 😭

It will not work well...

No, if you have a pulse signal, it will have an infinite representation in the latent (frequency) space. No advantage! Also, this JEPA is nothing new. In the past it was called PLS (Partial Least Squares and the likes -- Kernel PLS).

Related video (TaskGen Paper): th-cam.com/video/F3usuxs2p1Y/w-d-xo.html

Here's my discord group for the AgentJo discussion + Logo Design competition: discord.gg/bzp87AHJy5

Looking forward to following the development of AgentJo ( very cool name), keep up the excellent work Jon💪

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

TaskGen Repo: github.com/simbianai/taskgen AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

Book Link: www.amazon.com/Will-ChatGPT-Take-Job-Strategies-ebook/dp/B0D6Y8ZX5Y DALL-E course: www.udemy.com/course/the-dall-e-master-course/ Upcoming Video Generation Course in Singapore (Thursday 24 Oct 2024 4pm): lu.ma/p0g5ksm9

AgentJo Repo (building on TaskGen for multi-agent systems): github.com/tanchongmin/agentjo TaskGen Repo: github.com/simbianai/taskgen

Thank you for this tutorial. Would love to see integrating LanceDB (multi-modal vector db) to TaskGen.

cool wrappers

The Design Philosophy and whole architecture for taskgen was a pleasure to watch. Got to learn so much from it. I would love to contribute to the Project. Do we have any next ideas/features that we want to develop for the Project ? I have joined your Discord Channel as well

Pretty powerful. Thanks!!

Hi John. Was chugging along well and fine with Ollama-mistral-nemo model. All works well right till Tutorial 2 - Shared variables. improvised with another function with sqlite dbase - works fine as well. Then came this 3rd Tutorial on Memory, having OpenAIError: The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable. Is your memory method strictly requires OpenAI API key to continue? Look forward. Cheers and appreciate your work very much.

Update: Proudly rejected at ENNLP Industry Track 2024 for not having enough experiments!