I'm trying run the Deepseek model on Ollama in a Google Colab notebook with Langsmith, but having a hard time figuring out how to configure Langsmith. Any help appreciated. I've been looking at the docs, but no obvious solution stands out. All the examples use OpenAI.

can you create a video on how to build this step by step for a non-coder?

The most useful and also most interesting part of the deepseek-r1 model is literally what's in between the <think></think> tokens, you wouldn't want to train them away.

#LangChain I feel smarter after watching your content. Thank you for your dense AI stack mentoring. Prior to watching this video I didn’t fully appreciate how ‘in your face’ R1’s performance is until seeing your comparison to o1. Done at a fraction of the training cost.

Have you seen Stanford's Storm?

Thank you, I know this learns from a users tweets. I think it would be cool to create ai agent/bots/personalities (still learning the difference if any) of people not online or who speak differently online, like fictional characters… interacting with captain jack sparrow for example

OpenAI needs to change its name already to ClosedAI

Thanks for the explainer. This is really accessible and easy to follow. I’ve been experimenting with R1:70b on a local server and it’s remarkable what it can do. The explanation it gives of its ‘process’ is extremely useful for working out how best to refine a prompt. The main downside is that the model has biases (ask it about Taiwan!) which limit the ability to trust the model without being extra careful to verify any information it presents.

I would really like a diagram of the connection Deepseek identified. My question was ‘how are the causal factors in obesity interconnected’ would be very grateful for any ideas. By the way, humans are mainly visual so when trying to understand interconnection a diagram is best. Thank you.

Great explanation.

Hi there, thanks for your video. I understand you are going more into the nuts and bolts as to how these local systems can function. My question is: I really like using Notebook LM for a kind of organizing space for developing my research with a multitude of different sources. I am wondering if it is possible to do something similar without having to upload on Notebook LM's network - in other words, at times I use sensitive data which I would like to similarly work with an LLM and research tool it provides , but offline and maybe just communicating on my own computer. I am not a programmer so I don't know coding but would like to learn to make my own closed network to be able to do some more sensitive data research on a closed network. If you or anyone reading this has any resources, I would love to learn. Thanks.

I have watched a lot of videos on AI and LLMs, and I am especially eager to understand how Deepseek can do so much with so little in the way of development costs and runtime resources. This is by far the best explanation I have seen, and you have just earned a loyal subscriber. rather than spend your time telling us how SHOCKED the entire industry is, you provide a clear, detailed description of how this LLM achieve its reasoning results, using graphics to pull everything together. I look forward to watching all your videos, but I do have one suggestion. Don’t superimpose your face on the diagrams and text. It makes me crazy not to see the very thing that you’re describing until after you are done with it. This one small change will make your outstanding video even better. Thanks.

Good video but man please add chapters

Good talk, thank you

idiots keep refraining this stupid notion that theres system 1 reasoning and system 2 reasoning without (giving) any proof that this is how the human brain works. kahnemann introduced his "theory" by giving examples. thats all.

Don't ask deep seek what happened the 3rd of June 1989 in Tiananmen square

Please update the github repo link, I want to access the code but you may move the file to any other location. in addition, during recording, if you zoom the screen, then it would be helpful for those who have weak eye-sight

Hi Mr. Lang ☺ Thank you for the video

we need to have an experiment on groups of people try to learn something from scratch; one group learns from textbooks and classical work and the other uses LLM "summaries". To get a sense of what I'm getting at, is do you want your heart surgeon an alum of using LLM research tools to understand their profession and practice? Maybe it's good at extending your reach, in a sense of like experimental and cutting edge research but without learning solid foundations of established knowledge from primary sources anything gleaned from llm summaries seems like fast junk food yielding superficial learning. This is incredible though, the goal posts keep moving lol

Ngl... He looks like Dr House. Dr House x LLM's 🚨

I also thought it was awesome for DeepSeek to release the training methodology. This is all super frontier - but from my experience it seems like the distilled models are better than their base model (say Qwen 2.5 or Llama 3) at reasoning specifically but not much else.. that is to say the layer of reasoning data didn't magically unlock the ability to generate something profoundly better than the original models, just better at reasoning-based prompts. Still need to test but the 14b model still has a lot of issues/errors from my early exp.

Where is the Python 3.11 interpreter installed when you use the "uvx" command? I want to open this repo in VS Code but need help in figuring out which interpreter to select.

In essence, what DeepSeek R1 shown is that openAI is just doing horizontal and vertical scaling of their models together with some tweaks for fine-tuning, nothing special, nothing "ground breaking " (what Sam Altman was announcing for years now - AGI is close!?! ) . Basically, OpenAI was just making holes in the water and made all their users believe that with the next model, they will release AGI 😂

o1, o1-preview does not have structured output yet?

Great content! Thanks!!

Very good explanation. I'm not a programmer, but I can see how and when I would use this.

Do I understand correctly that to build the e.g., Qwen-R1 you replace the DeepSeek-v3 foundation model with any of the Qwen models? And follow the latter preocedure.

Great timing very topical Thanks!

Excellent

Hey Lance, can you make a video soon about storing memories in a vector DB and reranking etc? I feel like thats the final step right here!

exceptional -- really good explanation and a great notebook . thanks for sharing.

How can I create graph in the studio. And where can I get the Assistant/Graph ID?

Hello, I would to use deep researcher for searching the web for scholarship with criteria that I provided, and search the web deeply for the ones that meet that criteria, can it help with that? what do you say.

Thanks!

Really great workthrough demo and, as always, nice deep discussion of LangGraph's architecture here.

could you please make it in python too

he just showed us how to build our own Perplexity in less than 15 mins !

So far the best explanation I could found. Great video, thanks!

DeepSeek-R1 is nothing but another chinese junk

Awesome 🙌

I was eagerly anticipating your video on this. Thanks

like a brain having sections for reasoning and communication, those sections need to communicate, but should not overwrite each other while training

Amazing

Excellent video. Thank you.

how can we make this update_state as part of the graph for api calls? I am waiting for that sort of implementation.

Nitpicking here: The way you define precision is slightly off from the canonical way you'd use it in an information retrieval setting with multiple documents - setting it to 0 when only 1 document is irrelevant makes it confusing. What would be more clear to use in this setting would be `Precision@k = (Number of relevant items in top k results) / k` or name this check something like `all_documents_relevant` to indicate that you are basically using a custom defined metric not a canonical way to use precision.

🔥

How about inserting in models reviews some history benchmark? It would be nice to see politically brainwashed llms like deepseek-R1 fail miserably.

Excellent content

Release the notionnnnnn (please)