Видео

awesome work! thanks for sharing

Every paper you talk about is a gem. I am a fellow AI enthusiast trying to create Jarvis from Marvel! and your videos always help me to get closer to my goal. Please never stop making youtube content!!

Impressive work. Many good insights. Thank you.

Paper at 1:23:52 in which we talked about human priors in Atari Games, and how human priors are important for humans to solve games. Investigating Human Priors for Playing Video Games: arxiv.org/pdf/1802.10217

Nice presentation! I found this in the wake of the new DeepMind AlphaProof announcement. It seems they are doing something very similar with AlphaProof like here or with GPT-f, but their way of generating data seems fancier: Reinforcement learning on problem statements that are translated from natural language into a formal language, and if I'm interpreting their writing correctly, they even do some online learning. Anyway. For the sampling from the policy network, I could even imaging the usage of MCTS inside MCTS :D Instead of greedily selecting next tokes, or using beam search, the most promising token sequences could be found using a monte carlo tree search, a bit like in "AlphaZero-Like Tree-Search can Guide Large Language Model Decoding and Training". I kinda like ideas about combining LLMs and MCTS.

Nice thanks :) I like the discussion at 1:40:00

you are talking no sense

Slides: github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/Intelligence_Is_Sampling_And_Filtering.pdf

Great video John

Your videos are such a rich source of inspirations. Thank you for presenting these, it really helps me learn a great deal about the interplay between mathematics, language, and other areas of knowledge.

Curious how far you are 😅 i relistened to this and now second probably we are trying to enhance for the same solution

great to have this interview!

0:00 Speaker Introduction 0:45 Introduction to ARC-DSL 8:26 Data Generation 12:13 How close is DSL to human priors 14:42 How to decide which DSL to keep and which to add in 19:53 Introduction to RE-ARC 23:45 Overview of RE-ARC 25:32 Task Generalisation in RE-ARC 26:00 Example Verification in RE-ARC 26:47 Example Difficulty in RE-ARC 31:48 Limitations of RE-ARC 33:37 Examples of RE-ARC 35:22 Using RE-ARC to gauge model learning 37:48 Vision for meta-learning beyond RE-ARC 39:08 Can arbitrary DSL be generated with RE-ARC? 43:17 Discussion

Repo: github.com/michaelhodel/re-arc Paper: arxiv.org/abs/2404.07353

Regarding to the official paper: 1. In inference part, we're stacking R,s,a but a is empty, therefore how can we stack them? 2. If we're using a placeholder, i.e. R1, S1, NOP (that will stand for "no action" before a1 because it's not taken yet), we're getting a1 as a prediction in the inference 3. But in training we're not shifting actions, i.e. there's no mention that we're passing into the loss something like [NOP, a1, a2, ..] with [a_pred1, a_pred2,, a_pred3 ...] Therefore, the model is trained to predict current action 4. Since (R, s, a) after stacking stands for one token with dimension size which has tripled, inferring implies that we're getting (R1, s1, NOP) => a1 but the model clearly is trained (R1, s1, a1) => a1 because no shift is mentioned in paper. Moreover, if model is a function (R,s,a) => a, there's no point of R, s, it will just make it 1 and pass a => a. In teacher forcing, shifting is necessary 5. Nothing in the model points that the target is to maximize the return; it just generalizes relations between return, state, action tuples and some actions, unclear, the very same actions or the next actions. So I'm deeply interested in this topic but this concrete paper seems questionable despite being fundamental

What is the difference between this and the symbolic equation that KANs give?

"maybe the text itself is way more expressive than the knowledge graph" ouch 1:16:10 in the age of LLM, it seems that any representation that deviates from the source text is a serious gamble

Check out the repo here (Tutorial 7): github.com/simbianai/taskgen

This Conversation Class shown here is for TaskGen v2.6.0. TaskGen/strictjson has been greatly revamped in v3.0.0 to work better for smaller LLMs

Forgive me if my question is not correct, Are we using LLMs to build a Knowledge graph here?

Hi John, any reason not going with yaml vis-a-vis JSON because yaml format consume less token consumption compared to json.

Cool!

Update. This is live in TaskGen 4.2.1. Refer to Tutorial 6 for example of how to do Code Actions. My take: The code generated can be quite flimsy and harder to error-correct as compared to modular function calling. I think this method of using code is not very robust XD.

John, do you have perspective on other multiagent frameworks like TaskWeaver, CrewAI (based on LC) and LangGraph?

Great content!!

great content and thorough walk through. cheers!

Agents can eat breakfast three times in a row 😂🎉🎉 great video and accent 😅

that is freaking awesome

thanks for sharing!

As always, Fantastic video!!. Your content is just awesome. As you say "Food for thought", your video gives me a lot of content to explore :)

Hi John can ChatDev help develop a Webapp/Website?

TaskGen repo: github.com/simbianai/taskgen

My slides: github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/Code%20as%20Actions.pdf

Github: github.com/xingyaoww/code-act Paper: arxiv.org/abs/2402.01030

Thank you! Excellent teaching style 😊

This seems very useful! Can't wait to get deeper into this, I will be needing it soon for an upcoming project

Fun fact, at 1:11:50 onwards I repeated "I am such a failure with life" three times because of a buggy code XD (had to try 3 times to get it right) Haha live coding is fun.

Key Highlights: 29:31 Escape Room Solver 1:09:30 Psychology Counsellor The Math Quizbot it is better done with rule-based methods, you can see 41:26 for how it can be done

Notebook can be found here: github.com/simbianai/taskgen/blob/main/Tutorial%207%20-%20Conversation%20Class.ipynb

Here is a similar research for the modulo operator. project-archive.inf.ed.ac.uk/msc/20172390/msc_proj.pdf

Gemini 1.5 Pro: This video is about whether a neural network can model if-else statements. The speaker, John Tan Chong Min, investigates this by coding a neural network to model a simple if-else statement. The code defines a function that takes in an input x. If x is greater than 30, then the function returns 0, otherwise it returns 1. The speaker then trains the neural network on this function. The speaker finds that the neural network has difficulty modeling the if-else statement. The network can only learn a simple linear function, rather than the piecewise linear function required for the if-else statement. The speaker concludes that traditional MLPs may not be well-suited for modeling if-else statements. Here are the key points: * The speaker defines a function that takes in an input x. If x is greater than 30, then the function returns 0, otherwise it returns 1. * The speaker trains a neural network on this function. * The neural network has difficulty modeling the if-else statement. * Traditional MLPs may not be well-suited for modeling if-else statements.

Can you compare Autogen, task and langchain agents ? Which one is more mature ? Do you know any other one ?

Thanks

Is it function calling for Llama3?

Companion Notebook to this session: github.com/simbianai/taskgen/blob/main/TaskGen%20AMA_18May2024.ipynb

After this session, I went to massively improve TaskGen so that it works better for non-structured tasks and also not return so many errors due to list constraining or enum constrainer in Agent subtask generation logic. The Agent is also much more receptive to environment outputs and can modify it's plan more dynamically as it is not constrained to a list of steps anymore. The notebook for this session has been updated accordingly. Try out the new TaskGen v2.0.1 today!

Note: The information in this Notebook is largely correct, but for Functions, do take note variable names need to be enclosed by <> now, to let LLM understand better that it is an input. I will be releasing more updated tutorial videos in my StrictJSON/TaskGen channel once things become more concrete.

Great stuff! I've been looking for ways to generate output in JSON, and this helps me a lot. Thank you

Looking forward to more on task gen. Great work

Amazing work John, I hope I can be part of future meetings, I love the points that you are most focused on in these developments.