- 16
- 1 771
sanjay jayaram
Registrace 10. 10. 2011
Demo of an LLM assisted command line
This video is a demo of an LLM assisted command line built ontop of gemma2
zhlédnutí: 156
Video
Creating a Neural Network from scratch using only Numpy!
zhlédnutí 111Před měsícem
In this video we will create a neural network from scratch using only Numpy!
Demo of a LLM connected to a SQL database
zhlédnutí 126Před měsícem
This video is a demo of a LLM connected to a SQL database. You can chat with the database from the terminal.
Basic Transformer Tutorial in Pytorch!
zhlédnutí 50Před měsícem
My Attention is all you need paper reading: czcams.com/video/pV8mfFZKmfg/video.html
Mixture of Experts Tutorial using Pytorch
zhlédnutí 108Před měsícem
This video is a sequel to my Vision Transformer video: czcams.com/video/U0VJDNSQ9Ew/video.html
Readthrough of the "Contextual Position Encoding" paper by FAIR at Meta from stream
zhlédnutí 35Před měsícem
The paper read in this stream: arxiv.org/pdf/2405.18719 This video is from a twitch stream at www.twitch.tv/12344man Draw.io Board for the streams: drive.google.com/file/d/1N7bc7ZdhW0l3XZznqasg2zLG2MG2g5uP/view?usp=sharing
Introduction to Vision Transformers using Pytorch
zhlédnutí 167Před měsícem
This video is an introduction to Vision Transformers using PyTorch. Starts with the theory and then ends with building a model to classify the CIFAR-10 dataset.
Readthrough of the "Attention is all you need" paper from stream
zhlédnutí 57Před měsícem
The paper read in this stream: arxiv.org/pdf/1706.03762 This video is from a twitch stream at www.twitch.tv/12344man Draw.io Board for the streams: drive.google.com/file/d/1N7bc7ZdhW0l3XZznqasg2zLG2MG2g5uP/view?usp=sharing
Readthrough of the "Chameleon: Mixed-Modal Early-Fusion FoundationModels" paper from stream
zhlédnutí 76Před měsícem
The paper read in this stream: arxiv.org/pdf/2405.09818 This video is from a twitch stream at www.twitch.tv/12344man Draw.io Board for the streams: drive.google.com/file/d/1N7bc7ZdhW0l3XZznqasg2zLG2MG2g5uP/view?usp=sharing
Readthrough of the "An Image is worth 32 tokens" paper from stream
zhlédnutí 53Před měsícem
The paper read in this stream: arxiv.org/pdf/2406.07550v1 This video is from a twitch stream at www.twitch.tv/12344man Draw.io Board for the streams: drive.google.com/file/d/1N7bc7ZdhW0l3XZznqasg2zLG2MG2g5uP/view?usp=sharing
Web View of LLM based Mind Map Creator
zhlédnutí 61Před 2 měsíci
A LLM based Mind Map Creator. Powered by ExaAI's(exa.ai/) search and Cohere's(cohere.com/) Command R and graphed with cytoscape(js.cytoscape.org/)
LLM based Mind Map creator
zhlédnutí 90Před 2 měsíci
A LLM based Mind Map Creator. Powered by ExaAI's(exa.ai/) search and Cohere's(cohere.com/) Command R .
Demo of an LLM assisted Calendar
zhlédnutí 396Před 4 měsíci
This video is a demo of an LLM assisted calendar that is powered by Google Cloud's Vertex AI Platform
Introduction to LLMs with Ollama and Python
zhlédnutí 120Před 6 měsíci
This video is a quick tutorial and introduction to using LLMs in Python with Ollama. Link to Ollama Download: ollama.ai/download
Sanjay Jayaram - Showcasing the 30 CA$ Proximity Sensor to save from Covid spread.
zhlédnutí 114Před 4 lety
Sanjay Jayaram - Showcasing the 30 CA$ Proximity Sensor to save from Covid spread.
Best simplistic explanation!
Hey thanks for this walkthrough! I'm a data scientist but I don't work on these types of models. I don't understand what the latent tokens are that are used as part of the *input.* i.e. at 21:34 Fig 3c. I understand that the model does compression and that's what the latent space is inside the model. But what are the latent tokens that are used as input? Where do they come from, and what do they represent?
In essence, those latent tokens in the input are the output of another model that adds extra information to the image, for example, If you have a caption that goes along with the image(e.g This is a picture of an orange cat), you can feed those into this model to get a more accurate encode and later decode.
@@12344man Ah, ok. I had seen that in a few other papers but I didn't realize that's what they were showing here. Thanks so much!
🔥🔥🔥
🔥🔥🔥
🔥🔥
Also, is the favorable scaling of embedding size the reason transformers scale so well? Or do they have some other special properties that allow huge models like gpt4 to be trained?
The favorable scaling of embedding is indeed what allows transformers to scale intelligence so well. However, the sacrifice to do transformers is that compute needed for context length grows quadratically, so other architectures like Mamba could become useful for long context tasks, however those architectures do not scale intelligence as easily as transformers, so it is always a tradeoff.
@@12344man Thank you for the reply. I understand the tradeoff now, so how do you think the new 10M context window in the gemini models is possible? Do you think its just brute force computation, or have they included aspects outside of transformers in their models to facilitate the higher context lengths?
What is the purpose of the values component in the key-values-queries? I assume queries and keys help understand the amount of attention to give to each token in the data for a given token's query, but what does the consequent calculation with values do? Does it act as a sort of resdiual connection?
The value acts purely as a residual layer.
@@12344man I don't know if this is stupid or not, but then why does the value need its own weight? Can we not directly add the query-key products into the original embedding and have the same effect? Is it just to increase the model's learning capabilities?
nice explanations!
good stuff!
Here before 1 million views 🔥🔥🔥
This is amazing, will more backend lines make it run faster?
The main performance limitation of this is the rate limit of the API endpoint I am using. The easiest way to make this faster is to just give cohere a call and ask for a higher rate limit.
Thanks for your contribution Sanjay! Great insights
Awesome. Can you pls share us the github repo of the code and any Readme page to implement this ourselves?
Yup, its here:github.com/sanjaycal/mindMapper
You are a smart man!
$ cat main.py import ollama messages = [] while True: query = input(">>>") messages.append({'role':'user', 'content':query}) response = ollama.chat(model = "llama3", messages = messages) responseContent = response['message']['content'] messages.append({'role':'assistant', "content":responseContent}) print(responseContent)
Would be great to get more about the setup. Nice video, thanks.
which claude model are you using?
I am using Claude-3 Sonnet
First
🔥🔥🔥
No idea what you're talking aboit but wonderful video!
Interesting
Here before 1M subs! 🔥🔥🔥
Rice 🔥🔥🔥
Awesome idea!!
You fou d my channel as well?
Found
@@12344man it’s because I got a notification that you subbed to my channel so I checked out yours
@@AliYazbeck1 got it
Nice invention Sanjay👏👏 Expecting more and more from you
Thanks for the encouragement!
Bava wow, amazing , keep it on. True example of how young minds can really contribute to make it a better place. Really proud of you.
Thanks!