How DeepSeek is Changing the World of AI

all right we got to talk about it deep seek AI is here it is a chinese-based company that has been making waves in the last week that is rivaling open AI model 01 I do have to say this looks a little bit similar to the chat gbt interface but what we do know inside of this is that it does have a lot of similarities to be honest all of these large Legos models are just copying off of each other when it comes to the interface so I'm not going to get too caught up on that but it does look eerily Sim similar to chat GPT hi my name is Demitri panichi and I'm a content creator agency owner and AI Enthusiast you're listening to the AI agents podcast brought to you by jotform and featuring our very own CEO and founder idin tank this is

the show where artificial intelligence meets Innovation productivity and the tools shaping the future of work enjoy the show all right we got to talk about it deep seek AI is here it is a chinese-based company that has been making waves in the last week that is rivaling open AIS model 01 as you can see on the homepage of this website it breaks down its capabilities and Compares it to different models and at the moment it's actually 100% free and they also have an API that's available as well and it also seems to have an API that's in the works uh you do get this error when you go on the website but I'm very interested to see how this all plays out so we're going to play around with it in this video and show you how it works so if you're listening to the

audio version of this podcast make sure to go to our YouTube channel the AI agents podcast and no matter where you're listening please subscribe to the YouTube channel hit it with a like and do not forget to leave us a review on Apple podcast as we are trying to rank higher in the world of AI and podcasts however first I just want to talk a little bit about how deep seek seem to have come to the Forefront here uh it's not necessarily something that uh you experience every day you know a new product just popping out of the blue uh seemingly 100% free from a country that genuinely has seen some pretty rapid growth in Tech in the last little while and all I did was literally signed in with my Google account and next thing you know you have the ability to use search

and the ability to use deep think which is an more advanced uh deep logic and reasoning solving problems here which uh unfortunately when it what works with you know chbt on the base level you only get so many different uses but this is starting already with tech extraction from documents uh searches I think I could probably give it a screenshot if I gave it a try in comparison to chat GPT I'm curious if we can do that the answer is yes I can take a screenshot and I could just ask what is this a screenshot of and as you can see when I ask that a question pretty quick response yeah it is it is literally the deep cki and I do have to say this looks a little bit similar to the chat gbt similarities but to be honest all of these large Legos

models are just copying off of each other when it comes to the interface so I'm not going to get too caught up on that but it does look eerily similar to chat GPT just a little bit of background on the product though it is a AI model based out of China the model was developed by the Chinese AI startup deep seek which claims that the R1 matches or even surpasses say CH GPT 01 on multiple key benchmarks but operates at a fraction of the cost this is kind of a key differentiator here the cost is obviously something with AI everyone's concerned about deep seek success is even more remarkable given the constraints facing Chinese AI companies in the form of increasing us export controls on Cutting Edge chips but early evidence shows that these measures are not working as intended rather than weakening China's AI

capabilities the sanctions appear to actually be driving startups like deepsea innovative ways to prioritize efficiency resource pooling and collaboration so the old adage you know you make with what you have I don't know if that's an actually old adage but you know the concept of you limit someone's resources they try to get more Innovative that's actually what seems to have happened here and there actually been reports that the founder of this ended up finding a way to basically Jerry rig a bunch of uh 10,000 uh Nvidia gpus together now that's just a rough estimate that's been out there but essentially what what we found is that the product requires a significant number of gpus and high performance networking to keep its lights on and allegedly they're actually using unreleased versions of Nvidia gpus out there this is the specific thing that I read the Chinese

media Outlet 36kr estimates that the company has over 10,000 units in stock of a Nvidia a100 chips a type that's now banned from export to CH but Dylan Patel the founder of the AI research consultancy semi analysis estimates that it has at least 50,000 recognizing the potential of the stockpile for AI training is what led leang to establish deep seek which was able to use them in combination with the lower power chips to develop its models so essentially the US tried to starve out this innovation in China and what's happening is they're just making these models already very capable so if I if I I'm going to test it on some like logic based stuff so with deep think Ron I'm going to like propose it with with an interesting thought experiment so um like I'm going to get philosophical here for a second but

it explain in depth how you can have epistemic justification uh for reasoning via um something outside your uh sense data AKA how can you justify the existence of transcendental categories I'm just curious if it can answer like a philosophy question okay okay and it's interesting here so it's actually showing the thinking so I try to get it I tried to give it like a advanced thinking question as you can see it's already talking about Kant and K's philosophy and transcendental categor so this has took 16 seconds here okay so con synthesis is that um he argued that transcendental categories are innate structures of the mind not derived from experience but necessary to organize it they act as a framework that allows raw sensory data to be synthesized these categories are justifies Justified a priori independently of experience because they are prerequisites for any possible existence

for example the category of causality is not inferred from observing events but is required to perceive sequences of yeah that's that's how that would work um without transcendental category sensory input would be a chaotic flux this is all aligned with what how it works con transcendental deduction posits that categories must exist because they are the conditions that make experience possible itself and he's claiming that it's not circular reasoning but a logical necessity categories are Justified because they are the rules without experience would lack order and intelligibility and then gets into some more um explanations yeah so this is like a a decently complex um philosophic iCal thing so I figured breaking this down in in this deep seek would make sense um so as you can see it can think pretty well um it's at least able to spit out uh responses to philosophical questions

not that you would ever do this but it's always good for a quick logic test to see if this stuff works but I'd like to get into uh another thing real quick which is probably more the code based stuff so um write a basic uh Chrome extension code for a product that can capture markdown text and serve as a basic to-do list curious what it's going to spit out and I love how it actually shows it uh usually like it's showing like this in chpt but this actually just like walks through all of it in front of you now you could just keep it like this but I find that interesting and by the way what is interesting too is you can use the Deep think and search at the same time which you cannot do in chat GPT you can only do uh 01

uh without search as you can see here as you can see here if I go to 01 there is no search capability but in 40 there's a search capability now there's got to be some issues with the fact that it's copying the same logos and stuff but probably haven't copy marked I guess chbt probably hasn't copyrighted that in a UI like this but I'm not sure as you can see it gave me a bunch of code which is cool um I do want to kind of talk about this though because it doesn't seem to give me any canvas cap capabilities unlike chat gbt right now there's no scheduling of tasks um or anything to that effect so not exactly great there but I do like that it made this in the same sort of copy paste format that tra gbt has so it's pretty solid

also just to point out if you haven't noticed it is free right now right the annoying thing is that all the Articles you want to look at about this product are not free for the most part but anyways um the capabilities of this product are being talked about a lot on the internet but there is no cost to it at the moment and some other minor notes are that if you're a benchmark person you can absolutely compare this to the other products out there they don't have a direct comparison to gp01 but they do have one in comparison to 40 Claude 3.5 llama 3.1 and the ones that are in bold are like the highest rank out of the list and it seems like it's kind of killing it at code and math the product also does have a deep seek AI uh mobile app

which is interesting and it pretty much works in the exact same way as the desktop app now when it comes to the API let's talk about that obviously you probably want to see some of that so if I go to the API docs they do exist here so someone already built this out which is great um and people are actually a little bit concerned on the American and uh chaty PT side at the cost of this it is cheap cheap cheap cheap so if we take a look at the cost per million tokens right even for like chat GPT 40 $250 for a million input tokens right 1 million input tokens Price look at that price do you see the difference 2.5 Bucks versus 1 and5 cents is and then for the Deep seek reason version it would be .14 so this is an incredible

different price difference so we got input tokens versus output tokens obviously for the reason but like let's go to 01 here cost of 01 is there an 01 cost oh they don't really have 01 out it's more in like oh here it is so we got 01 $15 for a million input tokens look at this look at that price difference for the Reasoner this is in this is an insane difference this is this is 10x difference so you got inputs versus outputs right so we got for example 1 million output tokens for $60 for one okay output 1 million tokens $2 that's a 30X difference this pricing difference is insane they have gotten really really interesting uh stuff here I I don't know quite how they're going to have this continue but what I do know is I guess the API will have a smaller

context window in comparison I know that 64,000 is not quite what Chachi BT has regards to context length but uh it doesn't matter the total output is exceptionally large um so if you're looking to to utilize the API right this is this is going to be way cheaper than whatever you're working on with chat GPT so if you have chat GPT raer based products and you find that in your testing this output similar information as to what chat GPT does uh you might want to use this CU it is way cheaper and will literally reduce your costs by more than I can even you can even express um I mean .14 for the input and then 0.28 for the output let's compare that to 40 real quick one more time because I'm just trying to wrap my head around this $10 2.5 so 122 so

basically 12 and A2 bucks for both of those versus 40 C like I said like a 30 times cost difference that's not okay I mean I'm not saying it's not okay in the in the sense that it's not like this isn't somewhat reasonable but the difference is astronomical and obviously you can get lower pricing with price batching but even with the price batching it's barely it's I mean it's like half but that would only make it a 15x cost difference and obviously like I said the context window is 128k for 40 um and 200k for 01 but I'm sure the context window is going to get bigger for deep seek and if the pricing is already this cheap it's definitely something to consider um if you want to get into more philosophy with it um you can ask it about tag which is a a

transcendental argument for justification of thought um getting philosophical with AI is hilarious but regardless of that yeah I mean this doesn't really uh have I don't really know what else to say like can I ask it a question like uh can you make can you make a canvas of the code see if I can do something with canvas because I don't think I can but the search and the deep thinking capability is very intriguing to me I did get a response saying the server is busy please try again later um because of the overwhelm of this but uh long story short it is a disruptor in the space if you want to try it out I would recommend giving it a sh a shot but here's here's something that I will have to say to to put some context together here it has some bias

just as the US has some bias they both have different opinions on world events that is the complete uh comment I will make on it one has a certain perspective that might be more aligned with what the United States says one be might more aligned with um where this country of origin is from uh I'm not going to make a comment as to the truth claims of either of those I'm just going to say that they do have different opinions on this type of thing and it is a little bit unstable in the sense that you are getting a um you know a couple different like errors here saying oh we're busy but now it works fine took me a second but for the most part pretty much chat GPT without as much cost and um puts a big twist into the world that we

live in because I was not expecting something like this to come out and uh the United States definitely may have attempted to prevent something like this from occurring ing but evidently uh this is innovation that excites so if anyone's interested in trying it out check out deep seek right now there are a bunch of different cool tutorials about the product out there that I would recommend you just scour the internet for thank you so much for watching and we'll see you in the next one bye