people always ask what I do or how I got into AI… my background - over the past 6 years, have built AI systems for startups, mid-sized companies and global enterprises - youngest to get a phd in biostats from uc san diego - previously led AI/ML research initiatives at Harvard - recruited by the CIA at 18 and yeah on paper it seems like im this 200 IQ mf but funny enough I actually got rejected by pretty much ever college out of high school... and was a straight dumbass growing up (still am lol) but from that moment on I told myself theres no way in fuck ima be a bum… and just went head deep and pretty much taught myself how to code and theoretical math (which is what every AI model is built off of) and shit didn’t really come naturally to me either honestly... it really was just a fuckton of time spent throwing shit at the wall and slowly becoming less and less ass until one day I somehow became kinda decent at this stuff haha then out of nowhere through a professor at school, I got in contact with the CIA and later Harvard and ngl all that put my life into a whole different trajectory that I never envisioned for myself (note I never worked for the CIA thank god in hindsight getting a security clearance is prob one of the most painful processes in the world but was able to leverage that to get other opportunities like working at harvard and shit)

obviously with the rise of AI, it has become 100x for the everyday person to push out content but something important to keep in mind is the difference between content that’s like wow that’s so real I can’t believe AI made that vs content that actually converts and makes money bc the biggest trap people are falling into it seems like with AI content is that yeah it may look good and you cant even tell it’s AI but a video looking realistic obviously doesn't determine if an ad will print or not as there are obviously tons of ads made by human people that are 100% real that dont make shit and this isn't me hating on AI ads as I literally run a ton of them myself but it's also important to understand there exist much more important levers to pull such as the messaging/angle, emotional triggers, etc so essentially just DR/marketing fundamentals

this is actually so sick especially the fact it allows for mixed-inputs meaning you can embed video text together and get a singular embedding out a good use case for this would be building out an AI swipe file for ads like for a specific ad you can take the creative, script, hook type, angle, stage of awareness, USP, whatever you want really and create a singular embedding out of them using this new Gemini model and makes it 100x easier now do searches like "I wanna see supplement ads that focus on gut health" as the search will be done on the text layer and the creative layer now where as before it could be only done via the text so if you wanted search on the creative layer you'd probably would've had to get an analysis of the actual creative via GPT or Gemini to convert into a text based explanation and then append everything all together

if you want the highest quality AI UGC or talking head ads, Sora 2 Kling Motion Control is probably the best combo... as they are extremely complementary for each other where Sora struggles is they have this distinct plastic/mannequin look to them where you can just tell its made from Sora and you don't really get true character consistency either bc they don’t allow photos of people to be used in their reference pics but in terms of natural movements and pacing which is extremely important obviously, they're by far the best so use Sora 2 to create your base video so that you get the natural movements and the pacing of the character… then throw that into Kling Motion Control where you use a realistic looking reference photo from Nano Banana Pro and now you get the benefits of Sora’s natural movements and pacing... but now it also looks more realistic and you also get true character consistency to create long form videos as Kling allows for reference photos

if you're making any AI UGC, all video generation models have like this distinct static esque voice to them I'm not sure how to articulate it tbh especially if you watch it on an mobile it's usually more apparent there compared to watching it on my mac regardless, would highly recommend throwing any AI videos through Adobe Podcast's Enhance Speech at the bare minimum if you're not using any other voice changer as it's super easy to use... like it's just drag & drop and then the tool will just clean up the audio from there

read this shit if you’re doing anything AI related

random observation as for whatever reason much of my work lately has involved image analysis of some sorts with LLMs but despite GPT being ass af lately... when it comes to extracting visual details from images into a summary or a json, it has been absolutely shitting on every other model

when it comes LLM benchmarks analyses I rarely ever look at them bc at the end of the day the best models are the ones that YOU like the most there is simply no benchmark that can capture this bc everyone structures their prompts and uses LLMs in so many different ways and plus everyone has so many different and nuanced use cases with AI (even if you subset to something as simple just plain writing) that it's impossible for a benchmark to a make blanket statement on which model is superior a great example of this is I actually prefer Gemini for writing ad scripts especially if I want the best result in one shot and I have no reason why other than I just simply like the outputs better and again this isn’t even me saying Gemini is the best LLM for writing ad scripts... it’s simply my choice of model for creating ad scripts that seem the best to ME based off the way MY PROMPTS are written and structured again everyone has different preferences in what a good ad script even looks like and everyone has different ways in which they provide context and different ways they write and structure prompts the key thing to understand is you only come to these conclusions on which model feels best to you not by making assumptions based off what other people say or what benchmarks… you come to these conclusions from actually testing shit out yourself and seeing whats good vs whats not good

had to go on a small break from posting was just locked in af other projects and honestly kinda burnt out a bit from posting I think from just relentlessly only tweeting about strictly AI for months I just got kinda bored of it and also kinda forgot that this supposed to be a PERSONAL brand not just an AI account especially considering I don’t really shill anything on here and I have many other interests business related as well outside of just AI regardless still crazy af to me that 10k people actually follow me and even remotely give a fuck about what I have to say but my break is over now and my creativity and desire to write shit has returned I will be posting more consistently (but no where near at the rate I was doing before) luv u all

people ask me all the time what AI automations/systems to sell to companies 90% of all my AI related client work is listed by my dawg here

There's a lot of BS in the AI space. Lot of hoopla and over-complication by non-software engineers to try and bully you into buying their course. Forget that noise. If you want to make money with AI on the side, there's only 5 things you need to know. - which models work best for which tasks - cost/benefit analysis across models - when to start new chats vs rely on existing context - how to get consistent output formats for a workflow (this is the basis of automation) - how to string together models (output as input, which is the basis of agents) There are a few people that keep it honest, that I really recommend following to learn the ropes across concepts: @eptwts @EXM7777 @eyad_khrais @VibeMarketer_ @paoloanzn @tomcrawshaw01 @tyler_agg @lazycorp

I think a common mistake people make when writing prompts is the idea of throwing in tons of examples and expecting an LLM to magically learn from them for example let’s say you want an LLM to write ad scripts for you so the common way I see people go about is just find a bunch of high converting ad scripts and dump it in there as context and let the LLM go to work by giving info on your product - right? but let me explain why not be so smart... first you’re leaving it to the LLM to evaluate what makes these scripts good. which it’s smart enough to do to formulate it’s own opinion on what these scripts have in common… but does that line up with what you specifically want? because if you had 10 different marketers evaluate those batches of scripts you will get a variety of different reasons for why there are a good as obviously everyone has their own unique opinions. so you’re essentially gambling that the way the LLM is interpreting these examples you gave it line up with how you want it to look which is very slim haha so before considering examples what you want to do is explain in a step by step manner how you want it to go about writing this ad script... along with the logic and reasoning behind each step and I don’t mean step 1 write the hook saying write the hook is just telling the action on what it needs to do it’s not different than saying “write me an ad script” bc its like alright cool bro but how? you need to define intricately the thought process/behaviors/logic it should follow as it is doing this process so for example step 1 could be ____________________ read through the product/offer description and identify the market sophistication level as it pertains to this target market based on the following criteria Unaware Problem aware Solution Aware Product Aware Most Aware ____________________ im not going to do it here but then you’d provide definitions for each of the market sophistication levels and then here you can provide examples of products/offers and what market sophistication level they are and why you see here the examples supplement the understanding of the specific step rather than the end all be all and then you’d proceed with the rest of the steps overall though the best way to understand all this is you know when you use any LLM with a thinking feature... how you can click on the chain of thought to see how it’s going about it? we are essentially just guiding that part - the chain of thought and when you do it this way correctly you'll find that providing a shitton of examples will actually be overkill and you'll get much better outputs given the step by step instructions you provide mirror your internal chain of thought when you do the task P.S. this is also a great example on why you actually have to be good at said skill to use AI with it bc how tf are you going to come up with the logic and these instructions if you dont know wtf you're doing

here's a list of 10 underrated twitter accounts: @Mho_23 @lamxnt @beechinour @ErnestoSOFTWARE @tomcrawshaw01 @vasuman @VibeMarketer_ @tyler_agg @eyad_khrais @RyanClogg

if you’re trying to sell AI as more of a service say like consulting or AI automations its important to understand there is a huge dichotomy between AI systems that grabs attention on social media vs AI systems that businesses actually pay for for example on social media the shit that rips are like autonomous AI agents that spits out 2000 ads in a second and just automatically made them all this money or that replaced an entire marketing team where as from my personal experience the automations that a lot of businesses need and pay for usually boring ass automations that pertain to the finance/payments, contracts, structuring and cleaning data to improve product, etc and most of these businesses genuinely dont even want a fully autonomous agent doing all these tasks for them bc the idea of having an AI operate without a human in the loop whatsoever sounds risky af to them (and rightfully so) but ironically this is also the shit that grabs the most attention on social media and causes people to think is what businesses need what that's often not the case

to me the biggest difference maker when it comes to building a really fucking good AI system whether if you are technical or not is how much you are willing to obsess and understand the intricate details of the process you are automating for example I recently built a semi automated UGC system and yes there’s the 80/20 that makes good UGC so like the character choice, the script, etc which one would call the high levers that get you 80% of the way there but when you start diving deeper into this shit there are subtle nuances that to make up the last 20% that really make a hugeee difference and I see rarely ever talked about with how many people posting about some bs UGC factory they built haha let’s say I had a magical button that produced a good script and good character for the product and you wanted to build an actual automated AI UGC factory… how would you maintain realism? so they don’t look mannequin like? how would you maintain voice, character and setting consistency across numerous Sora/VEO calls? as a lot of ads im sure span 45 seconds which require numerous calls... what about how you would break up the script? as numerous calls require you to splitting up the script… but what’s the best way to do it? because if you put too much it cuts it off and now your ad when stitched together is missing part of the script and it’s incomplete and then if you put too little yeah you can cut off the end in post production but it also sometimes to ruin the pacing at which the AI avatar speaks (and you still have to use code to identify when the script actually ends to automatically splice off the the correct part of each call) or when VEO or Sora do mess up (because no matter how perfect the prompt is it will) being able to detect that at scale and initiating some of retry logic? and there’s so much more… and these are details that any person would agree is extremely important but is rarely considered obviously this doesn’t apply to just UGC this applies to literally EVERY process that you want to even semi automate with AI especially if it’s more intricate and to me the issue with a lot of these AI automations is not striving for perfectionism in the process they’re trying to automate and getting as ridiculously fucking close as possible as if someone were to manually do it themselves and although even with all this it’s still only 80% of the way there imo this only happens when you really drill down on every tiny detail of the process

my primary use case for gemini has been using it to critique my golf swing

i think a huge problem a lot of businesses have if they want to implement AI systems is the fact that they simply just dont have much or any data collected to provide as context so as a result a lot of times the first thing we implement isn't even AI but its's mapping out how to acquire such data... this could be in the form of writing documentation, external APIs, web scraping, using deep research, product design changes to request info from users, etc but this shit is what's required if you want to build any AI system that's personalized to any business

if you wanna make money selling AI systems especially to colder traffic you are much better off pushing a specific solution to a business problem that just so happens to be done via AI rather than leading with AI automations or like we can help you save 20 hours a week with AI shit like this worked like 2-3 years ago when the idea of AI was novel and cool but now that the market has sophisticated, people are entering that stage where they think yes AI is cool and is obviously the future but now there is a realism/skepticism aspect to it in that it’s not going to be perfect and it’s not in fact magic and since that novelty factor has worn off a considerable amount selling these AI systems goes back to just standard principles of crafting a good offer and shit like that