my name is David Andre and here is how to build anything with deep seek R1 now open AI is in trouble because deep seek R1 is an open source AI model that's just as powerful as o01 and open AI 01 was released only 46 days ago which means the progress in AI is speeding up fast so in this video I'll show you how to build anything with dcr1 even if you aren't a programmer real quick I'm hiring a full stack developer with lots of front end and design experience to help me build vle so if that's you make sure to apply the link is going to be below now here is how deep seek R1 compares to open A1 now as you can see these models are very close but the main takeaway is that this model comes just month and a half after opening I released their and it's not like this is just deeps benchmarks this guy on Twitter created his own eval where deep seek R1 destroyed all other models that's not all this is just the beginning Because deep seek the company they're insane they also dropped six small models which are distilled versions of R1 ranging from 70 billion all the way down to 1 . 5 billion which you could technically run on your phone so if you want me to make a video on these deeps models and how to run them locally on your machine make sure to subscribe now what about the cost well let's compare it to open A1 for 01 input tokens cost $15 per 1 million for output it costs $60 per million tokens so o1 is a very expensive model deep seek R1 on the other hand costs half a dollar 55 cents per a million input tokens and when it comes to Output it's $2. 2 per million tokens this means that deeps R1 is roughly 27 times cheaper than 01 at the same level of performance this is just insane and not only that we can also see inside of the model's reasoning this is the first super powerful reasoning model that lets us actually see the chain of fought we can see that it fought for 101 seconds and we can see the mistakes it made and so on open AI of course doesn't show any of that even though you're paying for the reasoning tokens you cannot see them if you're using o1 or o1 there is just something magical about seeing the model think through a problem just like a human would now how did deep seek do this like how did they pull this off well they spend less time on safety instead of like wasting 3 to six months like open AI they just ship also they have an absolutely crack team and a big part of that is because China has one of the highest average IQs in the world and also they're trying novel approaches such as deep seek R1 Z in this case Zero means for no supervised training data so instead of the model learning from Human labeled examples it starts from scratch AKA from zero only using reinforcement learning by the way this is the same as Alpha zero that mastered the game of go in that wonderful documentary from Google de mind now what's truly amazing is that this purely reinforcement learning approach resulted in increasing thinking time without anybody telling this to the model the model figured out on its own that if it spent more time on thinking and reasoning it will have better outcomes this is a clear example of emergent property now this was just one of many aha moments this model has created during this phase deep seek learns to allocate more thinking time to a problem by re-evaluating its initial approach it underscores the power and beauty of reinforcement learning rather than explicitly teaching the model on how to solve a problem we simply provide it with the right incentives and it autonomously develops Advanced problem solving strategies now you might be thinking but David how is this good for people like us well models like deeps R1 make it possible for people like you and me to actually compete with massive companies like open AI or Google for example open AI recently announced chbd tasks which is a direct compet to my startup vect and if you know anything about open Ai and Sam ultimate they like to destroy startups which is why I think the analogy of David versus golf is perfect here I mean my name is literally David and openai is a $200 billion Goliath with infinite funding bagged by Microsoft thousands of employees hundreds and millions of monthly active users and I'm a solo developer which makes competing with almost impossible however open AI cannot add deeps models into their products but I can and I will which means deeps R1 is coming soon to vectal AI now I think gy fan said it the best we are living in a timeline when a non us company is keeping the original Mission of open AI alive truly open Frontier research that empowers all of us it makes no sense and as Elon often says the most entertaining outcome is the most likely with that being said let's get to building all right so the first thing we have to do is go to deep seek platform and create an account I'm going to link this in the description that way it's easier for you but in here you can either use email and password or just Google Boom once you log in it should look something like this now if this is your first time using deep seek you will obviously have $0 right here in that case go to top up and charge up $2 I don't really think you need to go any higher than that because of how insanely cheap the models are so even $2 this will last you for weeks it's crazy right so once you charge up go to API keys and we to create a new API key now actually before we do that let's create an empty project inside of cursor so boom there is cursor let's create an empty python file called Deep seek r1.
py okay this is funny deep seek is actually recommending some Chinese characters for me um definitely don't need to do that if you don't know how to use cursor or how to set up an empty project just ask vectal or CH and they will give you the step-by-step instructions all right so now that we've done that we have to go to the Deep seek dogs and again I'll link this Below in the video but but basically we need this API call and in here we want to find deep seek Reasoner right here this is the new reasoning model and this is the name in the API let's click on this specific link that way we're looking at the documentation created specifically for this model so first we need to install open AI okay so let's go back into cursor and let's open the terminal you can do that command J or you can go up terminal and new terminal and in here we should run this command now before we do that it's good to First select the environment you want to use so I'm going to use test which is my test environment again if you don't know anything about K environment just ask how do I set up K now the advantage of using vectal for this is that it knows all of your active tasks it knows everything you're working on so it has that essential context and it can help you complete those tasks anyways once you activate the project you want let's simply copy this boom pip install update open AI all right so I've already have it installed next we need to decide on our API parameters and actually we can probably do that after we copy their code so documentation is actually your friend the the common misconception among beginners is that documentation is only for programmers actually the the more beginner you are the more you should be using recomendation because it literally tells you the right syntax but anyways let's look at this right might be overwhelming let's look at it one step at a time so I'll break this down into stages so that you don't have to worry right so here we import man this these Chinese tokens are so funny we import the open AI package and then we initialize the client with the API key but as you can see the API key is wrong here it's empty right so we need to go back to the Deep seek platform and generate a new one so click on create new API key I'm going to name it vl. a create new API key copy switch but I make sure to never share your API keys I will delete mine before uploading the video okay so save the file so now we've put in the correct API key and we have the correct URL here that the next step is to Define our messages array how should an individual prepare himself for the Post AGI world and Society what are the highest leverage actions one can take in the Years leading up to AGI okay this is a really interesting question I think then we send the message to the Deep seek Reasoner model to the deeps API okay so as you can see we construct the message list and then we pass it as the prompt and we make sure you set the model as the deeps Reasoner because otherwise it will default to the deeps V3 model which is still really powerful but nowhere near as powerful as the R1 model so Reasoner is R1 just for clarity right let's copy all of this and I'm going to utilize the cursor agent so let me do command L boom composer I'm going to say update deep seek R1 by adding more parameters okay boom here's the documentation above is the official docks follow it and actually I put in wait actually I need to switch to agent we want to use a agents right okay so obviously this is a simple task so even the normal composer could do it but let's see so yeah this is something that can happen with composer especially cursor agent it does more than you ask it to right so as you can see it added a bunch of stuff okay I'm just going to reject everything and just add Max tokens manually so if we don't want it to take forever let's make it like 2048 or maybe even 10 24 right now what's interesting is that This only affects the final response for the reasoning there's going to be a new parameter which is coming soon called reasoning effort and this can be up to 32,000 tokens so realistically most of the time when the model is responding is going to be in the reasoning part not in the actual final response so yeah we cannot control that right now but that's fine okay so let's see we send the message and we get the first and then rounds two okay so this is just a second example so let's try with the first example right let's check my vectal okay empty project streaming all right so we're going to set this set this up in a bit but let's parse the response okay and click on run do we run into any errors or not yeah this is this is the disadvantage of not having token streaming you're just like completely in the blind right for example if I open up terminal I'm going to give you a little sneak peek into the future video and if I run o Lama Run Deep seek R1 70b but I'm running this locally on my MacBook so again if you want me to make a tutorial on that how to run these distilled deep seek R1 models the smaller ones on your machine or your computer make sure to subscribe and comment below so if I say what is bigger 9. 9 or 9 11 we should see the reasoning tokens in here because AMA is going to be streaming them now hopefully my laptop doesn't overheat because I'm recording and I'm hosting a super powerful model locally but let's see okay so we we didn't we get the response I don't think we printed the response so boom this this is what it looks like it does the think um block and then this these are the reasoning tokens so to determine like this is fascinating you can see how the model is thinking about this to determine which number is larger I first consider the decimal places very clever all right anyways we don't have to go for this example but let's update our code so we actually don't print this out this is crazy you know what let's just impl token streaming right away I want to see the tokens being streamed that's the fascinating part so let's scroll down and then let's look at our examples I'm going to copy this go into cursor I'm going to paste this in this is a good practice by the way make sure to give cursor Cent documentation because even though it's using the son 35 model or even if you're using the deeps model it still doesn't have access to the internet unless you give it that tool access right so okay I'm going to say update our code in deep seek to ensure we actually use token streaming and we print out the response as it's being streamed so honestly in 2025 you don't have to write any lines of code you can just speak to the models in English or in Mandarin for that matter and they will write a code for you okay so let's see all right so lots of changes wow 5507 new lines of code let's go through them at a time so okay at the top it's deleting empty line can accept that importing system not sure why though then it sets up U what do you mean it's literally our prompt anyways for streaming it puts this parameter stream equals true and also it adds some extra print statements for better communication or better you know response inside of the terminal then it initializes empty strings to connect the full response let's accept that but this is a good practice just instead of accepting it all which sometimes you know you can just full send it and see what happens if it's production code and you know if it's important code just go one by one and see so for chunk and response then it prints out the chunks nicely let's do that all right so now we have token streaming implemented and we actually have more feedback in the console so let's run this again reasoning process boom there it is look how much faster it is when it's running on the Deep seek platform than my MacBook right so yeah this is the advantage of Hosting models on the web but I'm actually curious in this response you know what I going to say instead of printing into the console okay I don't know if that's possible but yeah whatever okay this is this is bad so the final answer is being printed with final answer before every time okay that's that's bad okay so now what you what you can do in cursor you can just put add to composer and you can say the reasoning part was show shown correctly I mean the token streaming however for the final answer you messed up messed up and showed final answer before every single token fix this okay so it's trying to add some booleans in here let's see if this is good I'm more interested in seeing the reasoning Pro process to be honest so let's see okay so preparing for a post eegi World requir strategic focus on high leverage actions to enhance adapt adaptability resilience and unique human strengths here is a structure approach cultivate adaptability lifel learning skill development focus on meta learning skills wow this is actually good guys listen up this is good prioritize learning agility over static knowledge I definitely agree with that basic a understanding learn foundational concept I mean this is actually really good advice like basic AI understanding learning the fundamental concepts of AI that's what everybody should do and actually if you're serious about AI then make sure to join the new Society because inside we have tons of exclusive content trainings workshops on how to build AI agents how to master from engineering and anything else you might need in to succeed in a post AI World in fact we actually have how to prepare for AI as well so if you are serious about Ai and you want to be ahead of 99% of people make sure to join the new Society it's going to be linked below all right so now let me show you how to turn this into a team of Agents because they actually support multi multi round conversations in the documentation so in each round of the conversation the model outputs the coot which stands for Chain of Thought reasoning context into this variable and The Final Answer into content okay so what I'm going to do is I'm going to copy this yeah they don't build this into team of multiple agents so we have to do that ourselves so I'm going to copy this going to paste ped into cursor agent as always and I'm going to say okay now let's add a second AI agent that builds on top of the output of the first one above is the official documentation from Deep seek so you know how to do it and we can explicitly say that the content variable should be included so make sure to include the content variable in the prompt of the second AI agent from the output of the first AI agent in the prompt of the second one all right let's see if curser can do this on its own because right now we only have one right but the real power is connecting them into multiple agents that's where you can build teams of agents and especially given the costs are so cheap like you can just have a team of Agents running in the background and even if like nine out of 10 times it's useless it doesn't matter because even that one out of 10 will be worth it give given how cheap these models are I mean 27 times cheaper than 01 same level of performance even better on some benchmarks like it's hard to Fathom just how powerful this is like imagine if phones suddenly got 27 times cheaper it would be worldwide news everywhere but with AI nobody seems to care I don't know may maybe we are The Crazy Ones okay so I don't know why it's adding this but harmless change so okay this is the second AI agent so here it constructs a more elaborate okay so it actually does a message history like this okay so user this is interesting so user asks how should an individual prepare for yeah that's my prompt and then the response from assistant and then the user ask the followup based on the previous answer what specific skills okay so maybe I should say what should be three specific actions I can take today all right let's see what deep seek R1 thinks how we should prepare for AGI so then the second all right the second guy is reasoning as well yeah I think this looks good so actually let's change the prompt think about what a post AGI Society might look like what will be important and what will be irrelevant what will the super intelligent AI replace and what will humans still be doing focus on the years right after AGI is created okay and then for the second prompt say based on your previous answer how should a person who wants to prepare himself for the Post AGI World spend his time in the daytoday life in the Years leading up to AGI okay this this should be interesting okay so AGI SCH leave massive automation job displacement yes of course high level actions are those that give the most benefit for the effort okay so Final Answer preparing for post AI World requires strategic focus on adaptability so this this is basically similar to what we've seen I'm interested in what the second AI agent thinks second a agent there it is okay so the user is asking for day-to-day plan based on the previous High strategies they want practical steps someone can take 30 minutes a day in AI literacy 30 minutes a day watching David Onre videos all right weekly activities could be scenario planning attending webinars okay use AI tools to enhance productivity such as vectal I mean if it knew about VL it would surely recommend it maybe I should ask it how to develop vectal okay that's actually not a bad idea using deeps R1 to help me put deeps R1 into vectal I'm definitely going to do that I really want to add it into the vectal chat because it's going to make vectal super powerful and I will do whatever it takes to make vectal the most agentic AI tool in the world it will have the most AI agents running and working for you out of any tool right now I would say perplexity is number one because when when you do web search it you know triggers multiple AI agents but that's still kind of low I mean as I said you can have ai agents running always like at all the tasks in your list there should be e working at all times trying to compete it and I'm going to make that happen so if you don't have act AI you're missing out all right so yeah I think this is pretty solid okay let's see wow this is this is a pretty good response we have a actual schedule here for a daily update so morning mindfulness read AI news creative practice work launch listen to podcast on AGI such as Lex Freedman work strategic session skill building exercise family call reflect okay so this is this is pretty safe plan maybe if we asked it for a more aggressive plan that would be interesting but yeah now you know how to build AI agents with deeps R1 and you can turn this into any app like just have a bit of imagination okay think what are the implications of it we have o1 not only open source o1 but 27 times cheaper 01 many millionaires will be created from AI in 2025 the question is do you want to be one of them or not if so take action do not postpone it and start building it has never been easier to create an AI startup I'm documenting every single change I'm doing in my own AI startup inside of the new society and I think this is actually probably the only place on Earth where you can see an AI startup being built from scratch and how I'm figuring this out I don't think there's anyone else who've built an entire startup and deploy it to the public and scaled it to 100 plus paying users just using AI I mean I literally knew nothing at the start this I've never made any AI startup before and I've been able to do it just by using AI tools so if you do make that decision to be in the category of people that do want to actually make something happen with AI new Society is a great resource for you to get started because I literally I'm showing everything it takes to build an AI startup so that you don't have to make the same mistakes as I did with that being said thank you guys for watching and if you watch on the end you are one of the rare individuals that actually value their time and are like to be hyper productive and I really do believe that V is going to be for you especially with this new feature that I just added which is notes notes are basically anything else that isn't a task could be a random idea you know a reminder a quote a thought anything you want to write down and remember and the best part is vectal automatically sorts each new item you create so you you don't have to think about it you just press q and you enter the name and if it's an actionable task it will go into your task list if it's a note it will go into your notes timeline and this is the first step towards making vectal a full stack productivity app going Beyond just a to-do list so if you want to experience the future of productivity go to vl.
