NEW: OpenAI o1 Strawberry Is INSANE!🍓🤖 PHD Reasoning, New AI Coding & Business Paradigm Deep Dive

What's going on guys breaking news open AI just released their new model 0 and this is the strawberry model that everybody has been talking about it and anticipating and in today's video we're going to do a deep dive into exactly what you need to know so you can get caught up with this model and know what you need to do in the next few days two weeks two months to actually start implementing this new model and new paradigm of how AI

Is now working let's dive right into it [Music] all right guys so I know there's so many people uploading videos and tweets and all this stuff about strawberry and I would encourage you to continue learning as much as you can now why is this video a little bit different well I spent the last few hours trying to compile all the information I could find as of now on this topic and formulate kind of like a thesis based on we're say seeing

As what this will actually mean for AI as well as for developers business owners and just anyone that's using this uh these types of models or really just anyone in the world because this is going to affect everybody right so in today's video we're going to talk about what is a one why it's important how does it perform in terms of versus other other models as well as you know previous models with open AI how it works we're going to talk

About its pricing and the video may be a little bit long but I'm going to try to my best to speedrun through everything that you should know and only go through the most important information that will help you right now and this in-depth document that I made and compiled over the last few hours with pretty much everything I could find around this topic at up to this point is going to be in the description down below absolutely for free for you

To take a look at download learn from use whatever the case may be um so yeah if you got some get some value here then just hit that subscribe button hit that like butt but let's Dive Right In also guys for certain parts where you may need to just passively listen you can put it to two times or 1.5 times speed and maybe for parts where it's a little bit more technical or you want to uh learn things and fully digest

Things then you can put it to normal speed okay okay so before I actually go through this document let's take a look at the official open AI blog post which uh is called learning to reason with llm so not going to go through it all I'm just going to go through some of the most important parts but let's start right now so we are introducing open AI o a new large language model trained with reinforcement learning to perform complex reasoning 01

Thinks before it answers it can produce a long internal Chain of Thought before responding to the user open AI 01 ranks in the 89 percenti on competitive programming questions code forces places among the top 500 students in the US in a qualifier for the USA math Olympiad aim e and exceeds human PhD level accuracy on a benchmark of physics biology chemistry GP QA while the work needed to make this new model as easy to use as current models is still ongoing

We are releasing an early version of this model open aai 01 preview for immediate use in chat GPT and to trusted API users okay so if you're on tier five you can use it in the API and uh yeah I'm definitely going to be doing some videos using cursor with it in the Noto distant future so our large language or our large scale reinforcement learning algorithm teaches the model how to think productively using its Chain of Thought in a highly data

And efficient training process we have found that the performance of 01 consistently improves with more reinforcement learning train time compute and with more time spent thinking test time compute the constraints on scale this approach differ substantially from those of llm pre-training and we are continuing to investigate them so this is a completely new paradigm of actually how you can improve llms so here's the different um graphs right here here's some EV vales so to highlight the reasoning improvement over GPT 40

We tested our models on a diverse set of human EV Valves and ml benchmarks we show that 01 significantly outperforms gp4 o on the vast majority of these reasoning heavy tasks unless otherwise specified we evaluated 01 on the maximal test time Computing setting okay so 01 greatly improves over GPT 40 on challenging reason reging benchmarks solid bars show pass at one accuracy and the Shaded regions show the performance of majority vote consensus with 64 samples okay so we can see competitive

Math right here competitive code code and then PhD level science questions and then here we see the ml benchmarks as well as the PHD level science questions and we see GPT 40 right here and then we see 01 Improvement okay so pretty cool you can see that it is way better at the math questions we can see the chemistry it's way better physics biology um elsat it's way better so pretty interesting stuff in many reason heavy benchmarks oh Rivals the performance

Of human exports recent Frontier models do so well on math and gm8 K that these benchmarks are no longer effective at differentiating models they had to make new benchmarks because lm's AI is getting so smart and we evaluated math performance on AIM an exam designed to challenge the brightest high school math students in America under 2024 AI exams GPT 40 only solved an average of 12% so 1.5 out of 15 of the problems 01 averaged 74% 11.1 out of 15 with

A single sample per problem 83% 12.5 out of 15 with consensus among 64 examples and 99% 13.9 out of five when re ranking 1,000 samples with a learned scoring function a score of 13.9 places it among the top 500 students nationally and above the cut off for the USA mathematical Olympiad he also evaluated 01 on GP QA Diamond a difficult intelligence Benchmark which tests for expertise in chemistry physics and biology in order to compare models to humans we recruited experts with

Phds to answer GP QA Diamond questions we found that 01 surpassed the performance of those human experts becoming the first model to do so on this Benchmark these results do not imply that 01 is more capable than a PhD in all respects only that the model is more proficient in solving some problems than a PhD would be expected to solve on several other ml benchmarks 01 improved over the State ofthe art with its vision perception capabilities enabled 01 scored 78.2% on

Mmm making it the first model to be competitive with human experts it also outperformed GPT 40 on 54 out of 57 MML U subcategories Chain of Thought similar to how a human may think for a long time before responding to a difficult question 01 uses Chain of Thought when attempting to solve a problem through reinforcement learning owan learns to hone its Chain of Thought and refine these strategies it uses it learns to recognize and correct its mistakes it learns to break

Down tricky steps into simpler ones it learns to try a different approach when the current one isn't working this process dramatically improves the model's ability to reason to illustrate this Leap Forward we showcase The Chain of Thought from 01 preview on several difficult problems below so I would definitely suggest checking some of these out you have Cipher coding math crossword English science safety Health Science and you can really see GPT 40 how it answers versus how o1 answers and you can

See the Chain of Thought right here and really how indepth it goes through U planning this out you know taking the input overall tasks constraints approach implementation steps uh understand potential issues write the skeleton and then boom then it you know generates the answer so a lot more in depth and uh yeah pretty crazy stuff so coding for all the devs out there this is definitely something you're going to want to listen to so we trained a model that scores 213

Points and ranked in the 49th percentile in the 2024 International Olympiad in informatics by initializing 01 in training to further improve programming skills this model competed in the 20 24 ioi under the same conditions as the human contestants it had 10 hours to solve six challenging algorithmic problems and was allowed 50 submissions per problem for each problem our system sampled many candidate submissions and submitted 50 of them based on a test time selection strategy submissions were selected based on performance on

The ioi public test cases model generated test cases and learned scoring function if we had instead submitted at random we would have only scored 156 points on average suggesting that this strategy was worth nearly 60 points under competitive constraints with the relaxed submission constraint we found that the model performance improved significantly when allowed 10,000 submissions per problem the model achieved a score of of the gold metal threshold even without any test time selection strategy finally we simulated competitive programming contests hosted

By code forces to demonstrate this model's coding skills our evaluations closely match competition rules and allowed for 10 submissions GPT 40 achieved an ELO rating of 808 while in the 11th percentile of human competitors this model far exceeded both GPT 40 and 01 it achieved an ELO rating of 1807 performing better than 93% of competitors so we see code forces ELO percentile we see GPT 40 01 preview 01 and then 01 ioi so pretty crazy stuff all right so here we

Have about safety hiding the chains of thought so we believe that a hidden Chain of Thought presents a unique opportunity for monitoring models assuming it is faithful and legible the hidden Chain of Thought allows us to read the mind of the model and understand its thought process for example in the future we may wish to monitor the Chain of Thought for signs of manipulating the user however for this to work the model must have freedom to express its thoughts in unaltered

Form so we cannot train any policy compliance or its user preference onto the Chain of Thought we also do not want to make an unaligned Chain of Thought directly visible to users therefore after weighing multiple factors including the user experience competitive advantage and the option to pursue the Chain of Thought monitoring we have decided not to show the raw chains of thought to users we acknowledge this decision has disadvantages we strive to particularly make it up by teaching the model to

Reproduce any useful ideas from The Chain of Thought in the answer from the 01 model series we show a model generated summary of The Chain of Thought all right so they won't be allowing you to see that actual Chain of Thought and then conclusion 01 significantly advances the state-of-the-art and AI reasoning we plan to release improved versions of this model as we continue iterating we expect these new reasoning capabilities will improve our ability to align models to human values and principles

We Believe 01 and its successors will unlock many new cases for AI in science coding math and related fields we are excited for users and API devs to discover how to improve their work daily so pretty interesting stuff all that's going to be linked down below as well as it's linked in this document so you can just check out the document and you'll get access to all that you can check it out too on their site take a look at some

Of those examples I was going through and let's take a look at what people are saying on Twitter and kind of my thoughts on this as well so here we have Sam Alman you know basically just tweeting out here is o1 a series of our most capable and aligned models yet Boom the article we just read through b is still flawed still limited and it seems more impressive on the first use than than it does after you spend more time with

It okay so you know kind of downplaying not really downplaying because it isn't perfect but you know this uh definitely there is this is just a start you know there it's this is like still the dumbest it's ever going to be it's going to continue getting better and better but here he goes on to say but also it's the beginning of a new paradigm AI that can do General purposing general purpose complex reasoning o 01 preview and 01 mini are available

Today ramping up some number of hours in chat GPT for plus and team users and our API for tier five users all right he goes on to say a couple more tweets the link to that Twitter thread is here he go then he also goes on to say no more patient Jimmy which is reference to Jimmy apples if You' been in the AI space or AI news space for the last month for last year Jimmy's basically been talking about strawberry and

You know basically referencing exactly when this uh model is going to drop so he's like a certified leaker that always has kind of insights regards to open AI um here we have open ai's tweet when they're releas basically saying they're releasing a preview of open AI then we have Jason way he was also in some of the demo videos for 01 that I'll play later on here but he he goes on and says super excited to finally share what I've been

Working on at open AI 01 is a model that thinks before giving the final answer and my words here are the biggest updates to the field of AI see the blog post for more details one don't do Chain of Thought purely via prompting train models do better Chain of Thought through RL uh reinforcement learning two in the history of deep learning we have always uh tried to scale training compute but Chain of Thought is a form of adaptive compute that can

Also be scaled at inference time results of aim and GP QA are really strong but that doesn't necessarily translate to something that a user can feel even as someone working in science it's not easy to find the slice of prompts where GPT 40 fails 01 does well and I can grade the answer but when you do find such prompts 01 feels totally magical we all need to find harder prompts and then four AI model's Chain of Thought using human language is

Great in so many ways the model does a lot of humanlike things like breaking down tricky steps into simpler ones recognizing and correcting mistakes and trying different approaches would highly encourage everyone to take a look at the Chain of Thought examples in the blog post which we just covered the game has been completely redefined link to that tweet right here and then Andrew M I've had access to open AI 01 for several weeks my advice on us again one don't think

Of it like a traditional chat model frame 01 in your mind as a really smart friend you're going to send a DM to solve a problem she'll answer back with a very wellth thought out explanation that walks you through the steps two write your prompt in notepad plan out what you want explain all the steps provide even more detail than you're used to three use 01 mini for tasks that don't require as much world as much World Knowledge but following things

Through step by step four 01 tends to give me the parts of the answer and the complete response while 01 mini will give me the steps five in my antidotal and highly subjective experience when you make Corrections or adjustments use patience in explaining what you need changed models at reason respond well to reasoned responses and then we have this tweet from Amar basically saying just combine open AI 01 and cursor composer to create an IOS app in under 10 minutes 01

Mini kicks off the project 01 was talking too long to think then switch to 01 to finish off the details and Boom full weather app for iOS with animations in under 10 minutes and the video sped up you can take a look at it but it's basically just showing him building that out so guys this is going to be huge for coding all right this is like the perfect thing when it comes to coding very hard C uh reasoning tasks or

Issues with code and it's going to be able to do this right so I'm very excited to start using this with cursor here we have congrats to open AI for releasing 01 economics we have Tyler asked 01 basically to write a college essay genetics uh cat Brown asked A1 to help her reason through n of One cases medical cases that nobody has ever seen physics Mario used 01 to draft and reason through complex Quantum uh physics equations code Ren hungu uh

Prompted a full snake game and it was generated zero shot working perfectly and Beyond an obeyed instructions to add obstacles so this is just referencing some of the demo videos and we'll take a look at some of these later but here's a tweet to that F also says 01 is the most persuasive language model ever and then he references Sam Alman basically saying I expect AI to be capable of super superhuman persuasion well before it is superhuman at general intelligence which

May lead to some very strange outcomes so I've mentioned this before in videos if uh AI gets super persuasive I mean it's going to be kind of crazy if you could talk to something that's essentially like a Salesman for anything that can sell you and change your ideas on anything I mean it's going to be really crazy how you know you could use AI you could like use it to convince you of things or convince other people of things right so

That's there's pros and cons to that but um also two there was uh information saying that how open AI you know Sam Alman and whatnot had to basically show Strawberry to the government like a a few weeks ago before they actually released it so um pretty interesting stuff then we have Greg Brockman his tweet open a i1 our first model pretty much just showing gratitude and going over his um you know opinion on this whole model okay now within this document

I gave basically a summary of some of the key features and performance benchmarks that we covered in that actual um blog so I'm not going to read through it all but I will highlight some of the key points and also things that I want to talk to you guys about so here's just what is A1 some key features so reason through complex cast solve harder problems in science coding math refine their thinking process try different strategies recognize their mistakes to Performance

Benchmark we talked about and one performance Benchmark the blog post didn't talk about was the strawberry test and it passed the test so this is basically it asking how many RS are in Strawberry there are three Rs in the word strawberry so this has kind of been like a meme in the AI space because that GPT couldn't or GPT 40 or what whatnot couldn't get that question right and now GP uh G uh open AI 01 of course did pass it

And they've kind of been moving away from the GPT name as you can see now is open ai1 it's not gp01 it's the first model with this new framework and new paradigm okay so here is the differences from previous models and we touched on this as well but basically it introduces a new approach that's um for AI scaling that focuses on thinking time rather than just model size and training data so the performance improves as it spends more time processing a

Problem similar to how humans often benefit from taking more time to think through complex issues this represents a significant shift in Ai and how you can improve AI models and performance and capabilities so so here's some of the different ways that it's improved and different from the other models a lot of these we talked about Training Method problem solving strategy performance scaling reinforcement learning new scaling Dimension so as uh gnome Brown U mentioned this approach opens a new dimension for scaling

It is no longer bottleneck by tree uh pre-training data as inference compute it can now be scaled as well all right I also include two brief uh descriptions of what Chain of Thought is as well as reinforcement learning for those of you who may not know or aren't maybe you know really want to refresher on it because this is kind of the two main things that we know of that that we know for sure that this new model is really leveraging

Like they haven't fully basically told us how it works um but yeah it's main we do know about Chain of Thought and RL so Chain of Thought is an approach that allows AI models to break down complex problems into a series of immediate intermediate tasks mimicking humanlike reasoning for 01 this means the model generates an internal Chain of Thought before producing a final answer it can reason through multiple steps refine its thinking and correct mistakes along the way this process is

Built into the model rather than requiring explicit prompting from users what does this mean guys well for those of you who may have been building a lot of different agents or different you know prompt engineering like prompt engineering has been such a huge thing we're going in the direction where the everyday user is going to be able to speak in everyday language and it's kind of adding like in code how there's different coding languages and different levels ofra of extractions from

Like say binary coding this is the same thing is going to be happening with AI models where before you know you're going to have to like you know use Chain of Thought prompting and like all these different prompt techniques to get an AI model to give you the desired output as we're seeing now and as we're going to continuously see in the future it's going to get easier and easier the model is going to do the thinking for you the model

Is going to do the reasoning the model is going to handle everything all of that all right every little thing you could think of so here's standard prompting here's train of thought prompting so you know you're not going to as of now like it's going to keep progressing so you won't have to you know be a master prompt engineer per se but I do still think as of now it's really good to know these things and it's still uh this we're

We're not there fully there yet okay so reinforcement learning so is a is a machine learning technique where an agent learns to make decisions by interacting with an environment and receiving feedback in the form of Rewards or punishment in the context of o1 the model is pre- is trained using a large scale RM algorithm it learns to think productively using its Chain of Thought the model performance improves with more RL uh reinforcement learning training train time compute and more time spent

Thinking test time compute there's a diagram there explaining that now applications and use cases we've pretty much covered most of these um but yeah some of them are here cenation which is a big one for you know you guys I know a lot of devs Founders what whatnot are watching this and this is going to you're going to see a huge huge huge boom in in build uh Builders all right people building things all right we'll go over that I have

A section on that what this means for Founders and and devs in just a second advancements and problem solving in stem Fields so basically solving problems in different fields um stem um that you know we just humans can't solve so it's going to I mean we're getting to that point where this is going to start happening and then availability and pricing so we covered this uh chat gbt plus users and um API users on tier five and then op 01 preview

Costs uh $15 per million input tokens and $60 per million output tokens so pretty pricey and then 01 mini is much more uh cost effective price at 80% less all right so the pricing is right here and you can see right here um is the context window Max tokens training data and the information about the U models right here you can find this in the back end side of the open AI API docs and then implications for the future of a

AI so advancements towards AI so this series represents significant Le towards AGI showcasing humanlike reasoning capabilities and we kind of touched upon these I did this video nine months ago on qar um and basically you know everyone's talking about the qar leak you don't know about qar then there was like strawberry there was Orion whatever um I don't know I guess well this one I guess is strawberry I don't know if it's like if this is qar or if this is

Just a a small glimpse of like qar um and maybe they have like a much more powerful model in the back end obviously they definitely have more powerful models or what whatnot but this is what they're releasing to the public I believe um and anyways yeah we kind of talked about like uh qar and like similar to what we're seeing with 01 and like how we different concepts that we talked about in that video regarding to qar it kind of makes

Sense that with what we're seeing and anyways paradigm shift in AI development so talked about that potential impacts on various Industries science research software Dev uh data analyst and then challenges and AI evaluation so like I said before as models continue to surpass human level performance on certain tasks there needs to be new benchmarks and evals um to accurately assess AI capabilities and then potential for breakthrough discovery the advanced reasoning capabilities of o1 could lead to new Solutions in complex Fields

Like physics chemistry biology Etc if you haven't seen my two uh my video on my thesis on AI coding I would definitely recommend checking it out I also leave a link to the doc right here so you can check those two out and then here's my advice to you guys if for the all the AI uh for all the business Builders coders entrepreneurs whatever you're watching this video you're like how do I start implementing 01 and I'm going to do more

Videos actually showing imple implementation of 01 um and like doing coding doing cool stuff with it so make sure to smash that like button to and smash that subscribe button to stay up to date we upload videos every single day on AI automation business growth Etc but my advice for you right now if the last 2 to three years haven't been enough evidence for you for some of you my advice is go all in on AI you know in your business

In learning it in implementing it Etc like speed is key go all in really um you know maximize your time of the day be productive don't be lazy there's never been a time like this in human history and just anything where you can learn a new skill so quickly you can build things so quickly um things are changing so quickly though so you got to be on top of it or else you're going to be left behind okay and yeah this

Channel is one of the ways that you know I'm one I'm trying to do that but I'm also trying to help you guys do that so if this if you get feel like you get value from this type of content make sure to subscribe like I said to stay up to date but also be always learning um your own stuff watching other content you know be on Twitter alling different uh thought leaders in the space and then implications for coders Builders

And uh business Builders so for Debs and entrepreneurs building businesses or SAS applications the 01 model and its scaling approach offers several potential benefits and considerations so improve code gen the 01 Mini model which is more cost effective is particularly effective at coding tasks this could significantly enhance coding uh code generation and debugging processes especially if you're using like tools like cursor and then complex Lex workflow automation so the models Advanced reasoning capabilities make it suitable for building and executing multi-step

Workflows potentially automating complex business processes and then potential for Innovation so the advanced reasoning capabilities will enable rapid development of more sophisticated AI power tools and applications opening up new business opportunities and markets for those who take advantage of it so like I said guys my big um if you haven't got anything from this video my the big takeaway is speed is key don't sleep on this um like this is just a start but this is just a glimpse into the

New paradigm and the new trajectory removing with AI um strawberry and you know this is also going to potentially well I mean just in general this is going to uh increase competition with other um you know competitors such as Claude anthropic you know I'm I'm guessing probably within the next month or two Max we'll probably see 3.5 Hau and uh Opus come out right so um I guess it breeds you know the competitive nature now that open AI finally finally actually

Dropped something and they're just releasing it directly there's no weight list or anything like that so that's good on the plus side but let me know what your thoughts are on on all this guys let me know what your thoughts are on strawberry 01 have you used this model are you planning on using it how are you planning on using it guys as well as uh yeah that's pretty much it so like I said this whole resource will be in the

Description Down Below guys so definitely check that out you can download it use it reference whatever the case may be and I'll continue to keep you guys posted on the 01 news 01 updates and I'm going to make some tests on this to test 01 see how it actually um you know performs in real world world use cases as well as using it with tools like cursor and other ones to actually see how we can build cool with it all right

Guys so make sure to subscribe to stay tuned also check out our free Discord Channel and Facebook group stri community.com link will be in the description down below as well it's linked right here I will see you in tomorrow's video guys keep hustling keep grinding and of course guys accelerate your stride take care I want to show an example of a coding prompt that 01 preview is able to do but previous models might struggle with and the coding prompt is to

Write the code for a very simple video game called squirrel finder and the reason o preview is better at doing prompts like this is when it wants to write a piece of code it thinks before giving the final answer so it can use the thinking process to plan out the structure of the code make sure it fits the constraints so let's try pasting this in and to give a brief overview of the prompt um the game scroll finder basically has a

Koala that you can move using the arrow keys um strawberries spawn every second and they bounce around and you want to avoid the strawberries after 3 seconds a squirrel icon comes up and you want to find the squirrel to win and there are a few other instructions like um putting open AI in the game screen and displaying instructions before the game starts Etc so first you can see that the model thought for 21 seconds before giving the final answer and you

Could see that during its thinking process it is gathering details on the game's layout mapping out the instructions setting up the screen ET and so here's the code that it gave and I will paste it into a uh into a window and we'll see if it [Music] works so you see there's instructions um and let's try to play the game oh the squirrel came very quickly but oops this time I was hit by a strawberry let's try again you could see

That the strawberries are appearing uh and let's see if I can win by finding the squirrel looks like I won