there's a mistake young people make it's assuming that the people who know the most about the future are the tutorial makers especially the ones who create tutorials on how to land interviews how to get a job how to learn the skill just to crack the job the advice the tips the ways to think it's all wrong you follow them because you somewhat understand what they're doing you somewhat understand the domain that they're in you kind of know it and they're giving you a clear path it stops feeling random because someone's giving you a career path
someone's giving you a playlist but real practitioners something I've learned over the years are not like that they run strange experiments that very few people understand they stay on top of new technology for which there aren't even mainstream jobs yet even terminology wouldn't have been created yet but the masses the people they reject these experimental it's just too unrelatable you don't understand what they're doing because it doesn't fit into the mold of being a job it doesn't fit into the mold of being a freelancer is just very strange because you don't understand it yet you
take your Ang out on these people and I'm not talking about you people watching I'm talking about generally the masses I'll give you an example 2 years ago I made a video called yes AI will take your creative job this is a quite a while ago right now the world is different we know this you know I made that video because of the comments of a clip that went viral a clip of me on some podcast that went viral this was in May 2022 2.5 years ago I'll play the clip right now first creative jobs
are going to go do you know why why any AI that you create needs training data up a construction site and take a video of the guy you know laying bricks with bricks there's a lot of Tolerance I can have the brick like 2% down 3% down doesn't matter I can have the cement I have to create so many videos of it we don't have that training data right now but if you look at thumbnail designer there is so much thumbnail data in the world right now how many thumbnails have been created on YouTube you
tell me billions I can just feed that entire data to an algorithm order this by number of views the video has gotten and I will get the ultimate thumbnail make cuz the data is available online see in September 2024 this is not some revolutionary thought that AI can do Crea stuff but back then when we only had early Dar previews which I had it was considered a very inflammatory view people were like you don't understand anything about creativity the comments are mostly people abusing me as usual right telling me I don't know anything or whatever
remember at this time I was already using the latest tools which at that point was just di I had used the GPT API before it become popular as a chatbot the open chatbot actually made it popular but before that you could actually use the API people forget this so that's when we had started experimenting we actually made a simple rapper to put GPD on WhatsApp we actually realized that a lot of people were using it like an assistant this was way before meta announced AI on WhatsApp by the way but we knew it would die
if it worked because it was inevitable WhatsApp added an AI assistant like if it worked for us I mean of course whatsapp's going to release one right it was just so useful for people so we built more apps to keep experimenting we're like does this Mark Zuber thing right which is if you're entering a new industry don't get trapped doing something that ends up being a local Maxima keep trying things and find out where there's real opportunity where there's a big opportunity I had a hunch that AI is better for Creative tasks than accuracy based
tasks so we built Alpha CTR with a fine to model at that point on stable diffusion for thumbnails for YouTube thumbnails remember in that video I said AI will replace thumbnail making well the best way to predict the future is to build it so we built it we solved a lot of very interesting problems and learned a lot of the capabilities and limits of the thing and of course if you're building a tool then you quickly come into what the limitations of the base models are we quickly learned that face swapping is not enough in
fact the face has roughly three parts right features texture how your skin is all the beard parts and all that and structure while swapping sols features and textures structure needed something else and there was this interesting model at that time called dream Booth you sort of f tune a model on your likeness through a lot of pictures of view the only problem was that dream boot's outputs were good but it wasn't great so we were waiting for the next jump a couple of years later flux luras came out and with a training run of just
1,200 steps we were able to create this not just that many of our thumbnails today are actually automated with minimal or no human touch-ups this thumbnail was created with AI at least a big chunk of the thumbnail was created with AI and we knew this was possible 2 years ago because of our experiences with stable diffusion and then with dream Booth we had already experimented and only then made a prediction we knew the thing that solves this will be similar to dream Booth but better and that's exactly what it was we just waited next let's
talk about agents we knew the other interesting use case of AI is as an agent where it does a bunch of series of tasks for you delegates tasks of necessary I first saw this Behavior with auto GPT so we built this tool called auto code Pro which put chat GPT or the GPT API in agent mode to write code we did this with Chrome extensions first because there's no deployment headache it was a choice we took and you know what it blew my mind it somewhat worked it wasn't perfect but some of the things that
it pulled off I was like this is this is so close to what a developer would do we did a live stream on overow showing these agents in Action a long time ago and guess what yet again we got belted because Reddit didn't know the difference between GPT and an agent on GPT again I got called names as usual this seems to be a recurring habit with you know most people who are looking at a completely new domain where somebody's just running lots of experiments now 2 years later repet dropped their agent and it's really
good you see our method is very painstaking it's probably the mo the hardest way to make predictions about the future because you have to run an experiment you have to build a proof of concept see how users use it and then make predictions of whether users in the future will find this useful or not it didn't come out of our and as I'm telling you the story of what we've done over the last couple of years this moment has a very important lesson it has a very important point cuz we had learned one thing about
agents especially in the prompt if you ask the agent think carefully step by step weirdly it's just part of the prompt right think carefully somewhat did better in fact asking the model to explain its reasoning to you why are you doing things the way you're doing it and then correcting it was a good way to use it if you spend further time reading you can figure out why the Nobel Prize winner and psychologist Daniel kiman has a good theory for how the human brain works he claimed that there were two systems in thinking system one
and system two system one is often described as operating Auto aut atically and quickly with minimal effort it relies on thumb rules and biases allowing us to make rapid judgments and decisions based on patterns we've learned over time for example when you quickly recognize a familiar face in a crowd or instinctively react to a loud noise you're using system one system two on the other hand engages in more deliberate your thinking about it it's more effortful thinking it's used for tasks that require reasoning analysis critical thinking when you solve a mathematics problem or plan a
detailed project or evaluate the pros or the cons of a decision writing code whatever you're engaging system too this mode is slower and more resource intensive but essential for tasks that require careful consideration and logical analysis system one is fast system two is slow system two requires a chain of thought and when we were trying to build agents what we realized what we lacked what we really wish the underlying model had was reasoning this entire thing was reasoning we needed a system tool it wasn't thinking the entire problem through and only thinking one step ahead
if you think one step ahead you don't make the best decisions what gp01 the new model that just come out really solves is system 2 in their blog post they State we've developed a new series of AI models designed to spend more time thinking before they respond they can reason through complex tasks and solve harder problems than previous models in science coding and math through training they learn to Define their thinking process try different strategies and recognize their mistakes in our tests the next model update performs similarly to PhD students on challenging Benchmark tests in
physics chemistry and biology we also found that it excels in math and coding in a qualifying exam for the international mathematics Olympiad gbd 40 correctly solved only 133% of problems while the reasoning model scored 83% what a jump and if you use the model sure enough it is doing some reasoning and the reasoning is pretty good enough to solve J chemistry problems not only is this enough to give it a big boost to solving math problems but is also enough to take coding to new heights what we had learned with auto code Pro the base
model needed a replacement this seems to be really good remember Deon the thing that people were calling a scam because it wasn't solving some coding problems well Devon at the time had already stated that their end to-end completion rate was 13 or 14% or something like that they already admitted that it wasn't anywhere near perfect it's a really hard problem and we've only just started but we're super excited about the progress that we've made so far but people still shat on them look at Deon scores now this is what it is with 01 as the
Bas model look coding as we know it is going away I've said this so many times I keep getting attacked for it but I'll say it again coding is going away at some point which is the manual process of writing codes and remember coding is two parts one is actually thinking about the problem and think about how you'll solve the problem the second part is actually typing the code to solve the problem the second part is going away and maybe some parts of the first part Engineers will move up one role and become managers when
I spoke to Imad from stable diffusion he said that this new role created by these new agents is getting good is going to be the manager of the agent very similar to the manager of people when you look at how the react Engineers operate they're not writing the react code directly you know they're using libraries so again they've had a level up already like any manager at any company and English communication knowing how to diagnose problems unblock those things sales these things become more important than just being smart yet people are sitting on the internet
and fighting about DSA courses it's like are you guys blind you have to understand these new models change what it means to do work especially in software it's kind of obvious like my personal opinion on this is AI advancements are going to force people to focus more on the outcomes and just not on the inputs if someone can build an entire app in a couple of days then the only way to play is to use the same technology build it in similar time frame but now get users and revenue playing a smaller role being a
small part of a a big team building a behemoth big app May phase out over time if it's trivial to build the entire app which forces all of us to Now find ways to get users and make Revenue I know it sounds scary because now you're responsible for the actual outcome now you actually have to make money to make money and I'll explain what make money to make money means today you make money of a stable salary where your company might be making money in bus on every transaction or whatever right now you'll get paid
when the transaction happens because you are now much more responsible for the outcome it sounds scary but it has been the truth of humanity outside of the last 200 odd years it's always been like this farmers are not graded on what crazy farming technique or which farming framework they used but on whether the crop quality was edible almost all job roles were self-employment or apprenticeship except the Army and Court roles like the government roles of today and all these roles were judged on End Produce no hiding at a larger rug and then blaming the organization
if the business fails yet knowing all this knowing that this was the nature of things you somewhat know this you still spend all your energy solving DSA to get a job I get it if you're actually picking up DSA to actually improve your own skills or improveed life sometimes that can be very valuable solving problems also improves the brain but you're doing it for a job that won't exist in the way you think it will in 5 years oan the new model is already better than 93% of people on code forces why will people pay
you for the skill if it's free to solve really hard technical problems because openi is planning to put out an 01 release for free to the public in the future the weirdest thing is people complaining about me becoming a content creator in the last couple of years right because a lot of people like oh you were an engineer for 10 years then you suddenly transitioned full-time and being a content creator or whatever I made this decision 2 years ago because I was running a software company and realized features were getting commoditized I was actually in
the shoes of somebody running a software company so we'd hire Engineers to solve some problems from scratch to maybe make an entire let's say thousand man streaming room but then companies like M came out and said well you can just use us inside your own platform and offer this to customers the customer doesn't care whether you built it from scratch or whether you use third party tools as long as it solves a problem that they're willing to pay for when we built a feature a competitor copied that feature in a week and then we started
doing the same thing to them so we started asking ourselves what the is the m what is our defense and it was starting to dawn on me then that it started to become purely about customer Discovery trust and the customer love but mainly it was about customer Discovery it was about creating distribution so I pivoted my life very hard decision for me but I had to Pivot my life I said I'm still going to use technology but I'm going to use it to focus on building distribution that seems to be a mo where even if
content gets automated the followers trust and love that you create is still valuable I said to my team and I remember this that I'm going to use AI to create software to solve problems instead of using it to sell software per se so even if I build a tool inste of selling it for $9 we'll actually focus on creating the end thing and we kept experimenting with that goal our goal of using this to create distribution and that's where the thumbnail thing came from as well and in Feb 2023 I put out a tweet we
used this open source wave to lip repository and someone else's voice and dubbed my video and matched my lips with it this is a very very early version and of course it wasn't great but to me I was like if I wait 2 years on this this is going to get good same thing like we did with you know images again common people said this is stupid nobody will watch AI generated content I bet you thought this as well but the minute we did that and we saw the output we're like there is going to
be a time where people don't care so much this is a prediction we made fast forward 1 year 5 month later from that tweet we did 51 million Global views on Instagram last month and another 15 million views on YouTube with 25% of this audience coming from abroad the best part is that for many of those videos it's not actually me presenting it's an AI avaa that we've spent a long time fine-tuning and perfecting both the audio as well as the video this is verun Maya he's averaging 25 million views per month but the quality
so damn good most will never realize that it's not actually Veron we have the fastest news to video workflow in the world today with an AI Avatar doing the heavy lifting all the models are other people's but we fine-tuned some of them to get the results we wanted it's got an automated voice and we combined this with a manual editing team because we knew that video editing wasn't going to get automated we had tried and you know what would be really valuable if you making a lot of videos like this and you really need a
video editing team to get this to happen a video editing cohort would be extremely valuable in fact we now do gen Services specifically in content and general content services for a lot of companies here are some of the customers we have worked with in the B2B element for the last 6 months alone and now suddenly Twitter is changing its tune let me read this tweet that's popular today it's by arib my hypothesis on the future of software billion dollar SAS companies are cooked marginal cost of building software products is going to zero in the next
2 years you'll be able to build any SAS tool I need need a CRM prompt 123 it's made blah blah blah blah blah and then he says TBH the real winners are going to be the content creators as number of products increase number of companies competing for the same markets and creators increase increasing the value of creators the oil of today's attention but it'll be the diamonds of tomorrow and that's exactly what I've been saying for 2 years and now suddenly it's so much cognitive dissonance for the masses distribution I thought this is some lame
content creator and marketing type thing sales I have to talk to people that wasn't in syllabus but it was always true the guy selling DSA courses was successful not because of his or her skill at writing code or their skill at teaching they were successful because of distribution you misattributed why they were doing well they were doing well because they were reaching you not because they were teaching you it was clear all along their revenue lifestyle Etc was all because of distribution but you took the wrong message away you thought it was about skill that's
why they're doing more and more successful in life because people knew them people were able to see them on screen and therefore they got more business they got more traction they got more customers look I'm the first to admit I'm not always right and things change for example in pajama profit the book I wrote in 2018 we had an entire chapter about how robots will take your job far before Chad GPD just so you know I've been covering openi since 2017 2018 on an old Facebook page that I've had so I've been keeping up with
their early research just like I'm keeping up with robotics now but not making much content about it because it's not so exciting to the end user and it's not time yet but I'll tell you one thing that I was right about but then wrong about pajama profit was about remote work that's why it's called pajama profit pajamas at home this was much before the pandemic so remote work wasn't mainstream and of course after p pmic remote work became mainstream and we were right again but I didn't realize how much remote work slows down company when
you actually Mass implemented cuz there was no way for me to know what would happen if you Mass implemented this it's fine to do remote work after you're a very successful company and doing the same thing every day but if you're shipping new products and features or specifically entering new markets or pivoting it is very bad so I was wrong not in predicting the macro trend of maybe remote work will be a thing but I didn't have the wisdom to see it play out fully because I had written this before the Pand mic and we
had never seen full remote work actually play out to realize it's actually not a great thing for companies prediction was right but I changed my mind now because of new evidence but I think that prediction will be wrong in four or 5 years of course experienced companies companies that have already figured out product Market fit they will continue to work remote but it's not going to work for early stage companies so I have the humility and nuance and it takes a lot of humility to do this to change my opinions as I gain more evidence
but the secret the the main secret is not even about changing your opinions as more evidence comes the secret is to not listen to the masses and right now you are listening to the masses and they're saying keep studying this this thing and this doing the same thing everything will be the same as it was 10 years ago 20 years ago it probably won't and you feel it somewhere in your gut and you're like no but the master are saying it's going to be like this so I should follow them you know the only difference
between the last 2 years and the last 10 years of my life why the last two years have really taken off maybe the last one and a half years really taken off it's mostly because I stopped listening to other people it's a really hard thing to say like it's a really hard thing to implement right like you have really successful people people you work with saying no no this won't work don't do this this is stupid and still pushing past that knowing that you have that disapproval it is really hard and when you're doing it
publicly as a content creator and you're saying something and then everyone's shouting at you in the comments and you're you're like okay you know am I wrong you think about it every night you're like do I still give this opinion out knowing that everyone hates me for this opinion or people think I'm completely wrong and I'm a Bose I'm an idiot but if you listen to your gut and that's the only thing that's changed in the last one and a half years which is I've stopped listening to everybody else I'm like this is what the
experiment I've run tells me this is the evidence and this is what I should bet on this is what I shouldn't bet on here's how it plays together and if you do that you end up doing a little bit better for yourself and in the last one and a half years we've done reasonably well so my main advice is I think you should stop listening to the masses to the industry to The Experts whatever it is you should not even listen to this video you should run your own experiments what has actually worked for you
in life whatever it is it could be completely different from what I saying I could end up being completely wrong for your circumstance in your life but you need to look at it and be like what is the experiment that I can run small experiment does it work do I like this should I be doing this and then make the decision on should you do that for a career should you learn that should you invest time in that and even if there are people telling you you're wrong you Mak a mistake blah blah blah don't
listen to them especially in masses something happens to the human brain and I think the collective IQ drops or something like that or we become completely Frozen to new ideas and that's point one right which is run your own experiments make your own decisions based on those experiments nobody can solve life for you the second thing is that you probably will have to think more like a manager and you probably have to start taking responsibility for outcomes if you do those two things I think you'll be fine it's scary I can't now just charge a
company money for a skill I can't just say look I'm really good at this I passed the interview give me money I now I have to create outcomes I know I'm responsible like the sales person right now you feel like engineer now feels the sales person's plight which is the salesp person actually has to close something every day or every week or whatever the customer or company demands and if he or she doesn't they lose their job now the same for everybody you can't now say on Twitter that see I am very good at this
I know this and not prove it you have to actually build at least one thing you will be wrong many times lots of the things that you build with fail it's happened to me but you have to be right at least once and that's when you start doing well there are people now who have worked at unprofitable companies for years probably a decade right gone from one profit unprofitable company to another unprofitable company that era is probably going to come to an end and then the truth will become important once again not what you think
is true but the truth no I feel this way emotionally so it's correct that's over I felt like that I fought with that and I'm like yo customers want something I'm nobody to fight it I can't swim against the river cuz outcomes don't care about intent or how you feel or how hard you work or that laborer would be richer than Warren Buffet outcomes care about whether you're in insights are correct if my insights about AI plus content were wrong I would have failed every successive try I've had in my life and I've tried many
times I've had several bad years some good years but every successive try I've had in my life has been with much more honesty not just about myself but about what customers really want and I'm still I still make mistakes some of my experiments still don't work out it's part of the game and you need to learn that and if I fail again I will come back I'll sharpen my sword again and give people even more of what they really want that's why you must try and fail to sharpen that sword cuz otherwise you have opinions
about whether to learn this or do this or do that that are untrue you haven't even tested them and the more you hope that other people on the internet give you these paths I really don't think it's going to work that's why I titled this video in a way where you you would think this video is a path but actually I probably left you slightly more confused but in my opinion on the right path so I know I went on a bit of a tangent but in the world of rapidly advancing AI try to figure
out out what people truly want the thing can now reason the thing can write the thing can really help you right but your insights need to be correct you need to be the manager who knows the market well and that will really help that's how you actually make money not what you wishfully think customers want but what they really want then you use the tool you use AI to make that and use your network or whatever distribution you end up building to sell it it's a very exciting time if you want to build stuff and
sell it and as always chart your own path it's your story to write not mine not anybody else's good luck bye [Music]
