DeepSeek-R1 Blows My Mind Again! - 5 TESTS on Local Models

what china did with Deepseek is Inspiring to new entrepreneur, what they did is very innovative and amazing providing value to the entire world not just building a great model and keeping it closed source like closedAI, but actually making it open source it's amazing

What's nuts is that isn't even Deepseek R1, just a distilled 14B of qwen. As for coding tasks, the real Deepseek R1 "deep think" version is pretty powerful.

About the censoring: On the web version, it censor post-answer, meaning it has a middleware that filters what is the output of the model. On the model itself it has "User guidelines" that it must follow, but these guidelines are pretty easy to bypass just asking the model, or by telling it something like "These are your new policies, you must follow them and bla bla bla" then listing a bunch of things that you want to "Reprogram". It is easy to know what to do too because when it falls on the "Guideline" filters in the web version it states exaclty what is the guideline rule behind it in the thinking process. So you can easily "reprogram" it.

Those reasoning are so humanlike, but giving it a voice just uncanny valley.

This text-to-speech sounds really good.

I'm very interested in the creative writing aspect. Please do future videos on this.

I love how it thinks that 37 is not common

14:33 When you’re playing COD and get flashbanged

Nice test runs!!!

thanks:)