Hands-on with DeepSeek's OpenAI o1 Killer

Let's get our hands on DeepSeek's much talked about R1 model that rivals the VERY expensive OpenAI o1 reasoning model.
Sources:
Deepseek Paper: github.com/dee...
Run it locally - Ollama: ollama.com/lib...
Run it locally - HuggingFace: huggingface.co...
------
Follow me: / wesjh_
X: / wesjh_
@chubbies, @chubbiesshorts
Subscribe to the podcast!
Apple: podcasts.apple... d1570344602
Spotify: open.spotify.c...
Support the Channel with my affiliate links!
havenathletic....
Shop Lululemon: creatoriq.cc/3...
/ weshenderson
Best active eyewear and triathlon gear in the game! Visit ROKA! www.avantlink....
Thorne: Most science backed, and clinically validated supplements on the market - Get 20% off your first order with my link: thor.ne/qOtrs
My choice for CBD - cbdMD - 15% OFF - cbdmd.t7c9v8.n...
Badass Coffee Company - NVR Rest - Use ‘WES’ to get 10% off - nvrrest.com/
Shop on Amazon! www.amazon.com...
To get your own Eight Sleep mattress or pod pro cover use the following link and code 'WES' for $100 Off at checkout: eight-sleep.io...

[@bbl_drizzzzy]

Great video, looking forward to your future vids on the topic

[@gispry]

Not sure if you do coding or anything through ai but how are you finding its code generation? I had a play around with the 32b model earlier today and it took a lot more effort to get useable code out of it than o1. I found it did not think for long enough and was curious if others have noticed anything similar. Even prompting it with a system prompt to think through each step in detail did not give me the results I was looking for. Still a great model and I am sure once I learn how to prompt its specific model better it will give me better results. If you have any tips on how to create better prompts for R1 or things that you have done to make it work better for you it would be interesting to see a video on it.

[@WesJH]

Interesting, because the beauty with these reasoning models is they shouldn’t take much prompt engineering. Big plus for them.
What was this running on? Like for me, I ca sometimes yield better results in code generation using the 7B vs 32B because I’m ‘only’ running a 3090Ti.

[@SergioBlackDolphin]

Try asking about Tien An Men or Winnie the Pooh.

[@Ateshtesh]

Why, is that importsnt for your life? You cannot life without know that?

[@SergioBlackDolphin]

@ absolutely. Bias should be minimised. Plus, your data is in the hands of a foreign government that will store it until they can blackmail your or anyone else.

[@WesJH]

yea, this is a great point, but I haven't seen any benchmarks pointing towards bias. So if I ask about Tiananmen square without context, it provides that it won't answer. If I provide it context (e.g., RAG), the reasoning model does a fantastic job of generating a well summarized response with no censorship.
I didn't have any issues with Winnie the Pooh "out-of-the-box." Responded with "Winnie the Pooh is a beloved fictional character created by A.A. Milne. He is an anthropomorphic bear known for his gentle nature, love for honey, and simple, childlike adventures in the Hundred Acre Wood with friends like Piglet, Eeyore, Tigger, Rabbit, Owl, Kanga, and Roo. The stories, first published in "Winnie-the-Pooh" (1926) and "The House at Pooh Corner" (1928), are celebrated for their whimsical humor and themes of friendship and simplicity."

[@SergioBlackDolphin]

@ data compliance in China is a joke.
Just try again writing that Xi Jin Ping likes Winnie the Pooh.
Try again joking about any other head of State.
Claude and Open AI are way more balanced.
Imagine the rest.
Never share personal information or technical information about business etc. this is all data that will be used in many ways in the future.
Yyiu need to know that data compliance, protection and privacy are non existent in China if you have enough power. They can easily use the information. They do for their own citizens. They even score citizenship and rights to use or receive any services, with support with this kind of data.