awesome!

are you sure 16x16 is the number of patches? 16x16 should be the size of the patch (number of pixels) and 14x14 should be the number of patches.

I am struggling to understand how the search would work if we needed to handle some kind of structured data. Anyone having thoughts? I would appreciate your assistance

In building agents, which is preferred; langgraph or crewai ?

RAGAS must go. I dont really understand why such weak method promoted so much.

Thank you for the video! The notebook seems to have been removed from the repository. Is it still available?

If you want to go deep into the rabbit hole of Neuro Symbolic AI there is some interesting recordings here www.youtube.com/@asimmunawar/videos

Very Helpful, subscribed !!

The font at the beginning is hard to read

Too much talk and very few illustrations

Love your videos and wlakthroughs - really easy to understand!! thank you james

By the way, do check out Max Tegmark's recent paper "The Geometry of Concepts: Sparse Autoencoder Feature Structure" - points in the same direction

I've loved embeddings for a long time, so very powerful for semantic AND analogical search. But their power relies on the quality of the embedding model. We've been looking at SAE's (mechanistic interpretability of LLMs) as embedders. All concepts in the embeddings are unique (sparse_, and decomposed, ie Interpretable embeddings from an LLM. Goodfire intend to make this into a service, do take a look at their Goodfire SDK, specifically the Latent Explorer. Their SDK is new this week for the ApartLabs Hackathon tomorrow.

Just curious if there is a way to pass required param option to the function schema definition and fail the function call if not provided.

Not a great description of what's going on

Great RAG walkthrough! With Web3NS, you can combine this tech with YourName.Web3 to create decentralized AI agents that integrate seamlessly with Web3 ecosystems. 🚀

Are you suggesting the image encoder never have seen any cassette player during it's training and still it has formed an understanding of it in the latency space? I would disagree to this.

The way I am starting to think about neural logic circuits (logic = a dog has 4 legs, 2 ears, long nose, etc) is a matrix of real numbers which house the weights that makes up the "rule/logic". You would start off like, is this an object, that would be at a certain location in the matrix from 0.0 to 1.0. Is this object organic/in-organic next cell, is mammal, etc etc. This would look like a grey scale mxn image for example. This image would represent a dog not in the way a CNN would identify a dog, but in a graphical sense. A nxn Perceptron. It's another type of embedding, but it's not a word embedding. You could experiment with mnist autoencoder but instead of encoding graphics representations of numbers, it would be rules for things.

That’s a very keen observation about boxing yourself and if you’re only using an LLM. One of the things I’m curious is how do you envision the semantic router working with non-LLM based inquiries combined with LLM based function calling together to create a cohesive strategy that also incorporates API Potentially as micro services and how would you envision this working together in sort of an AI middleware similar to llama stack? Also, do you see if integration with your semantic router and the logic around the routes that you create using NLP and potentially knowledge graph frameworks? Also, I’m interested if you see any framework where there are spontaneous creations of agents based upon API request coming into an agent that doesn’t necessarily have all of the internal services to accommodate so it builds the agent on the fly creating a more dynamic workflow.

Congrats on the birth of your son 🎉

thanks - great video:)

Hi James, a fantastic video. Thank you for the validation. We are building an AI system that uses the semantic router, and our system resembles your diagram. I look forward to the next video on this topic. Cheers Chris

Yeah, that it's the way... But i do not agree with de symbolic treatment... It would be deeper using another "symbolic carrier"...

The GOAT is back !

*Rethinking AI Agents as Neuro-Symbolic Systems* * *0:00** Introduction:* James Briggs introduces his perspective on redefining AI agents beyond the current narrow view of LLMs interacting with tools. * *2:04** ReAct Agents as a Foundation:* He revisits the concept of ReAct agents, which involve an LLM going through reasoning and action steps, including calling external tools, to answer a question. * *7:28** Expanding the Definition of Agents:* Briggs argues that the common perception of agents as solely LLM-driven systems with tool interactions is too limited, both practically and in the broader AI literature. * *8:37** Neuro-Symbolic AI as a Framework:* He proposes the neuro-symbolic framework from the MRKL paper as a more comprehensive way to understand agents. * *9:27** Symbolic AI:* Symbolic AI represents the "good old-fashioned AI" approach using handwritten rules, logic, and ontologies, similar to the syllogistic logic example provided. * *12:48** Connectionist AI (Neural AI):* Connectionist AI, the foundation of modern neural networks, was inspired by the structure of the human brain. Rosenblatt's perceptron is highlighted as a key development. * *17:23** Neuro-Symbolic Systems Blend the Best of Both Worlds:* Neuro-symbolic systems combine the learning capabilities of neural networks with the reasoning and structure of symbolic AI. * *21:09** Agents Beyond LLMs:* Briggs emphasizes that agents can leverage various neural network models beyond just LLMs, citing semantic router as an example where embedding models guide routing to either RAG pipelines or LLMs. * *25:21** Practical Implications of a Broader Definition:* He demonstrates how a neuro-symbolic perspective enables more flexible agent design, where different intents can trigger specialized toolchains and workflows. * *27:34** Conclusion:* Briggs concludes by advocating for a more encompassing understanding of AI agents as neuro-symbolic systems, allowing for greater adaptability and pushing beyond the confines of purely LLM-centric approaches. I used gemini-1.5-pro-exp-0827 on rocketrecap dot com to summarize the transcript. Cost (if I didn't use the free tier): $0.03 Input tokens: 20966 Output tokens: 433

21:19 Why in the lowest level there aren't 1.000.000 values since the dataset is 1 million? What happened to the rest?

Was this run GPU accelerated or just done on the CPU?

So it's basically skip list but in vector space?

Super 👌 Needed that clearly delivered to me. Thank you.

Very good video even 3 years after it was uploaded. Would’ve been amazing if you also included how to plot loss and accuracy graphs and confusion matrix. And although finding out the precision, recall and F1 scores isn’t that difficult it would’ve a nice add to the video.

brilliant video. it's exactly what i'm looking for. Thank you! I've been using LangGraph and trying to find any good practical reasons to use Llamaindex for next agentic app. Btw, i think LangGraph also support async implementation. But i can see the design concept is quite different. look forward to more videos from you. :)

goated video

link for the code?

What is the difference between a graph and a state machine?

Great video! Any plans of doing something similar with CrewAI?

hey james, i wanna ask out of scope this video. do you drink coffe?

Thanks for the video really interesting. This works with Azure Open AI AD authentication? ( No API KEY )

Hi james. I know this video is already a year old and there has been a lot of new development, but didn't contriever already outperform BM25 at the time on most benchmarks? I believe Contriever fine tuned on MS MARCO basically outperformed BM25 on everything.

Thanks for the video, helped me!

Super good explanations

Great introduction! Will you make another video with more in-depth or additional examples?

What about the google embedding model? Have you tried? It's a lot cheaper than open ai's embedding model.

Great tutorial!! Thanks!

@James Briggs Newbie here, was wondering if it was necessary to store the chunk with the vector, it seems like a lot of data duplication and a good way to fill your disk. I like the idea of storing the title, I was thinking about storing the document path and filename also. I haven't been able to find good info about what data besides vectors is also kept in the vector db. I understand that the vectors need to correlate to data, I just don't understand what data is actually represented in the vectors. If you just have an ID and the vectors, can't that ID point back to the document with the content?

I have a question, you say I do this if I wanna finetune my model on my custom say domain dataset and at the same time using the same approach the model used to pretrain originally. What if I wanna do pretraining from scratch? Say I want bert to pretrain it not on english data, say japanese, and I wanna pretrain it from scratch, will I follow your approach? But instead of .from_pretrained I would use .from_pretrained('a saved BertForPretrainingModel') or what exactly?

this is shitty as fuck, you people are sick

47091 Schneider Stream

Jackson Nancy Hernandez Timothy Jones Ronald

This is really good! Would love to see the whole journey of building a Dhivehi language model. How you trained models? What is your plan for instruct-tuning? Also on how you guys build the dataset, how you pre-processed and cleaned the dataset, removed the redundancy, etc.

Harris Nancy Lopez Christopher Williams Thomas