Training Your Own AI Model Is Not As Hard As You (Probably) Think

[@Steve8708]

Read more in my latest blog post: www.builder.io/blog/train-ai

[@balachandar3360]

Hi Steve I am building a project on digitizing catalog for a hackathon in your recent video you told you trained a model on detecting images from webpages. if you can open source that it will mean a lot because we are having very less time to complete the project.

[@saadshajy3849]

We created a number of GPT powered personal AI assistants for businesess like law, media industry and pretty much for corporate companies where they can integrate internal corporate resources and data securely to retrieve useful information and save 70% cost

[@Osirisdigitalagency]

​@@saadshajy3849 thats nice, could you explain the steps that was taken to train these personal models

[@tribulusfox3816]

Hi Steve, How can I reach you?

[@TinS0lder]

Did you make "not a hot dog".

[@JakeLuden]

This channel just feels like it’s on another level than the rest dude, in every way. The way you turn intimidating concepts into very easy to follow videos is invaluable. Quickly becoming my absolute favorite dev channel.

[@amoledon]

This channel is combination of business, technology and sales training anyone needs to build a start up. Kudos to Steve.

[@rdrjr4129]

These types of videos where you give a general approach to a common problem rather than a specific solution is so much more valuable.

[@ReflectionOcean]

- Explore pre-existing models for solutions at 0:53. - Break down complex problems into smaller parts at 0:55. - Test general models to understand their limitations at 1:19. - Consider the costs and feasibility of training large models at 2:17. - Solve as much of the problem as possible without AI at 3:13. - Identify the right model type for your needs at 4:00. - Generate high-quality example data for training at 4:44. - Use tools like Google's Vertex AI for model training at 5:54. - Deploy and test your model with real-world data at 6:46. - Combine specialized AI models and plain code for best results at 8:35. - Utilize LLMs for final customization steps when necessary at 8:47.

[@samueldominicmani2577]

this was helpfull thank you

[@PotatoMagnet]

Dude, you’re soothing to listen to and it’s nice to hear your chain of thought for approaching a problem. Subbed!

[@PlsDontSuuue]

for real, the moment i open other videos and hear an indian accent its an instant nope for me

[@impossibletechie8059]

Would be happy to pay for a complete course on this explain each step in detail

[@Antoine-l9m]

I was getting stuck on training a chatbot agent using off the shelves LLMs. This video came out at the right time and saved me at least a few days of monkering around. I really like the thought process breakdown instead of just showing jumbo codes like a tutorial channel. You got yourself a sub!

[@Sindoku]

Man Steve, you are QUICKLY becoming my new favorite person to learn advanced coding techniques from. Please keep up the amazing work!

[@saadshajy3849]

We created a number of GPT powered personal AI assistants for businesess like law, media industry and pretty much for corporate companies where they can integrate internal corporate resources and data securely to retrieve useful information and save 70% cost

[@pedroaquino5368]

Simple video, complex concept explained beautifully. Thanks Steve!

[@MrThryler444]

Between turning to basic code and utilizing an LLM (a large "super" model) for generating customized code, I'd suggest storing both the input and output. Over time, you can train your own smaller LLM with fewer parameters, perhaps 7 or 13 billion, specifically tailored for the language/framework. By collecting user feedback through edits to the code and storing this feedback, you can achieve better results than relying on an expensive "super" model like GPT-4.

[@pollywops9242]

Amazing I was looking for exactly this info I was super overwhelmed

[@samuelparry5736]

I'm interested in the layout hierarchy approach too. What type of model did you use for the layout hierarchy? Do you give it an image and get it to return some HTML / JSON Structure? Or do you transform the existing Figma nodes into some text representation? How did you generate the training data for this step? My initial thoughts would be to provide the model an image (the same screenshot used for the object identification step), and get it to return HTML. Then for the training data, when you take the screenshots of the webpage, you also save off the HTML and provide that as training data for the hierarchy model (maybe after some processing). If you can share, I'd love to learn more!

[@elliottchong]

This is awesome dude! I watched your last video about developing proper AI apps. Your content on AI is top notch, you have opened my eyes to a whole new world of AI development. Much appreciation, keep up with the good job! Thank you :)

[@samuelparry5736]

I see Steve post.... I click...

[@constantinelinardakis8394]

5:17 quality of model is dependent on quality if the data

[@cagdasucar3932]

I don't understand. Why would you need to run image detection for importing Figma designs? They already have node/element structure. Image block detection is not needed for an importer plugin like this at all. It would be useful if you wanted to build a webpage from mock-ups, not Figma.

en mantıklı yorum

[@samsunforev1846]

I was copycatting others project using Openai wrapper, and it didn't feel right to me, but I didn't know what to do instead. Until I came across your channel and now I have a whole new perspective on creating an SaaS AI product. Thanks Steve!

[@ZaVeg-p5s]

Love this. It truly shows how much goes on behind a single button we click on a page

[@ericschrdr]

Thank you for simplifying and proposing a best solution - great stuff! Subscribed 🤘🤘🤘

[@mat-tr-ez]

Thank you for putting out these type of videos. Each video has covered advanced topic in a simple manner and made me curious about building something simular!

[@yarapolana]

First time seeing a video by you! Instant follow, really well broken down.

[@LukeOtwell]

Love the bullet points, and break down of what each task implemented. I would also add that to get things up and running quicker, see if there is an API for certain steps in your workflow i.e. dynamic scraping or object detection. Eden AI is expensive and slow, but if you haven't nailed down the workflow, it can help you try tools and compare services. Great vid!

[@Thosch_io]

Very interesting! The area of opportunity within the AI sector now is niche AI's that does one specific task very well as it is impossible to compete with general models as OpenAI, Anthropic, Google and other large companies has almost infinite money and resources. This means that quality of the data is as you say essential and one of the best ways to build a moat against competitors. I will definitely keep in mind that I should just write normal code to solve as much as possible before creating my own model. Do you use any open source models or have you made one from scratch? Been fiddling with fine-tuning LLMs such as llama-2 7B myself and I am now considering trying to find a model I can train instead with my own data.

[@Failedsociety998]

I would like to know if it is possible to develop or use an artificial intelligence capable of: 1. Reading and interpreting various types of complex data codes, including hexadecimal or binary formats. 2. Identifying and isolating specific parameters within this data. 3. Modifying these parameters based on requests or predefined settings, while ensuring the security and integrity of the modified data.

[@dynastyduck3799]

So quick and clean, thank you for taking the time to do this; I was trying to understand "Training", and now I know it is a button

[@arttupakarinen7642]

sooo... in a way. to automate computer use, you you need to specify tasks like "click input with label name", and this would automatically classify you an "input detection AI" which is only searching for input boxes. Different small tasks are all different AI:s OR system functions ("type text "ABC"" would call a system function) to make it efficient. Then there would be another AI for making image into TEXT and passing it to LLM for analyzing. etc. etc. And to me it does not even sound too much of trouble to train these "small task" AI:s like "Search input with label "name" " or "click red button on the center of the panel with text "OK"". .. Maybe the hard part is the TASK-delivery AI which'es function would be to turn natural language into system functions or AI models - to - system function kind of thing. would be a lot of fun!

[@normanschulz708]

your content is very clever and feels enlightning for me who is discovering machine learning and AI models step by step now. Not at all specialist on the topic. Thanks for that dude. Big up !!

[@lauri2806]

I appreciate this a ton. Since seeing a nice, unique project a guy had with regarding AI, I only thought about applying multiple AI's together trying to get what I need. Now I have some new insight on what I just might be able to do by myself, with my own skills.

[@blender_wiki]

In the broader context, investing in large-scale language models (LLMs) for natural language understanding may be deemed inefficient and resource-intensive if such capabilities are not a primary requirement. A conventional coding strategy that involves breaking down workflows into modular functions with standardized algorithms, coupled with the integration of smaller machine learning models, is not only more maintainable but also facilitates better evolution. This approach is not only more realistic but also cost-effective in comparison to the utilization of extensive LLMs.

[@ragtop63]

Here's the kicker: None of this will work for someone who either doesn't know how to code or is a beginner/intermediate coder. If you know just enough programming to know that what a LLM is giving you is either wrong or isn't optimal, but you don't know the solution, there's no way you'll be able to validate the data you want to provide to the custom language model without involving someone (or others) who can help validate that data. At that point, which came first? The chicken or the egg?

[@FlutterPythoncombo]

Steve, I really appreciate the video. Honestly, thank you. It was very informative. God bless you and whatever you build. Thanks.

[@weiguo9887]

Exactly what I’d like to learn! Great content! Subbed

[@stevejones6330]

Excellent video. Clear, concise, to the point.

[@rinorajeti4]

What a lovely video man, the way he talks with all the details you can tell that this is naturally true

[@constantinelinardakis8394]

4:43 example data

[@lost4468yt]

If you were to add an extra network that goes between the smaller ones, it suddenly starts looking (from a high level of course) like the larger scale networks in the brain. The most obvious example would be the brains Default Mode Network.

[@Sylarleft]

We're building an AI solution at my company. And this is the same conclusion we arrived at! Glad to see another company illustrating the same approach cu

[@adicandra9940]

Everything you build is very inspirational, Steve.

[@Focus11493]

not sure its possible cause i dont know much about Figma, but thinking about Figma mirror with the ability to mirror the design in the browser, id imagine you can take the generated html via inspection tools and feed that to the model instead which would kill the image detection model training step and really what you end up with is a HTML to React Component translator. its still along the same lines of what you end goal was which was taking the initial Figma design and translating it into code. anyways just food for thought from one dev to another! great vid!

[@keithchen5467]

Amazing insights. Can you share more on steps 3 and 4?

[@mercedg]

I really liked this video, I’m commenting to boost algo and hopefully get more vids like this

[@SebastianBauer-k4h]

Thanks a lot for this great advice! Cool to see, that the good old "divide and conquer" still is true in the brave new world of AI 😊

[@pixelperfectpravin]

I appreciate your content and your efforts

[@lololoololdudusoejdhdjswkk347]

My question is why are you using object detection model if you can use figmas API for identifying their layout hierarchy? Sure it costs money, but I doubt it would do much and you would probably get much more comprehensive data.(like document nodes). Also I wonder if there is such as improvement to what fine tuning the LLMs do for this use case( aside from a certain framework you want to be improved(svelte)). But if you’re trying to fine tune by identifying divs and stuff and making better code choices, that must’ve taken quite a very long time(since you guys have a team it makes sense for this decision). However much of the newer LLMs like Claude do this without any help either. All in all it seems just a send more data to the “fine tuned” LLM sort of solution which honestly is how most products in this case are right now.

[@noduslabs]

Your tutorials are great. I also make trainings on a text analysis tool I created so I know when I see good stuff! Thank you! 💪🏼🙏🏼

[@debasishtripathy530]

Sure, here's a table summarizing the key differences between using off-the-shelf large language models (LLMs) and training your own specialized AI model: | **Characteristic** | **Off-the-Shelf LLMs** | **Specialized AI Model** | |---|---|---| | **Cost** | Generally more expensive due to high training and usage costs | Can be more cost-effective, especially for specific use cases | | **Speed** | Can be slower due to model complexity and inference time | Can be faster due to smaller size and focused training | | **Accuracy & Predictability** | May be less accurate and predictable for specialized tasks | Can be more accurate and predictable for specific use cases | | **Customization** | Limited customization options | Highly customizable to specific requirements | | **Data Requirements** | Require large amounts of data for training | May require less data, especially if using transfer learning | | **Training Time** | Can take days or weeks | Can be faster, especially for smaller models | | **Expertise Required** | Requires specialized expertise in AI and ML | Can be more accessible with basic development skills | | **Use Cases** | Suitable for general-purpose tasks | Ideal for specific, well-defined problems | In general, off-the-shelf LLMs are a good starting point for exploring AI solutions and for tasks that require general knowledge and understanding. However, if you have a specific problem that requires high accuracy, predictability, and customization, training your own specialized AI model may be a better option. It's important to carefully consider the trade-offs and requirements of your project before deciding which approach to take.

[@milanp7]

Was this an Ad?

[@shakapaker]

I'm so grateful for that type of content

[@Emmanmanis]

My friend this was a very clear and easy to comprehend video! I know nothing about coding and would like to learn more. Hypothetically if I wanted to develop a AI app. What route would you lay out based on your experience?

[@TheHeroIsRisingUp]

YT please recommend to me these kinds of videos.

[@meisherenow]

Something like...using a dumptruck to deliver a small package is okay if you're doing it once or twice, but if you're starting a delivery service, a little effort to reduce cost is a good idea.

[@JetSoftProHQ]

Thanks that showed that training your own AI model is simpler than we might think. We, at JetSoftPro, a software development service, also think that custom AI development can achieve faster, cheaper, and better results than using off-the-shelf large models like OpenAI's GPT-3 or GPT-4.

[@constantinelinardakis8394]

3:17 try to complete the problem without ai. It helps to break the code down in pieces and write simple code

[@franpalokaj]

Great stuff Steve, thanks a lot!

[@gabrielcebria1887]

Thanks, really cool video!! If you don't mind me asking, what sw are you using for the blackboard presentation? sleek!!

[@felipeblin8616]

Great! Thank for the video. One question. How do you train a help desk with this?

[@TshDev]

Amazing video, can’t express enough the amount of value it has. Thank you🙇‍♂️😊

[@lololoololdudusoejdhdjswkk347]

Very cool, I was wondering,how were you guys able to make your layout hierarchy model and what did the image detection model help with that? Did you guys just use the code of the test data and train a prebuilt model off of that?

[@vpakarinen]

I been wondering this, there's so many general models that are just too big and would love to have smaller model that do specific task.

[@bennguyen1313]

Any thoughts on tinygrad? Also, do you think that eventually the commercial models will be so much better than an open source model that training your data on your own will be for very niche cases?

[@the_saucisson]

Surely Figma could be able to do this completely without AI, since the design is technically already code? Like, the design is being displayed in a browser, so it should just be a matter of optimising it and saving it to a file?

[@chucklohn522]

When you say "insanely expensive", what kind of numbers are you talking about?

[@luizosorio]

Insane numbers.

[@michaelbubnov3306]

You messed up a bit your general formula: quality of the model=data quality/censorship. I remember AI Dungeon going from good enough quality to nearly unusable even on paid plan in one update that added censorship.

[@Collect-AI]

Thanks for shearing. Question , how much was the total cost to build your models?

[@iMplimentation]

Thank you for this Steve, you helped me rethink my initial solutiion to a problem I planned to use AI to solve.

[@olegmikulskiy]

Steve, you mentioned that it's easy to use some python library to run free LLM image model on PC. Can you show it?

[@jonan.gueorguiev]

Thanks for sharing that! Actually even more interesting is the other specialized model - layout generating one. Can you share something about it as well? 🙂

[@shubhamgarg5007]

Tried the Vertex AI a few times for different projects. The issue is that the endpoint always returns the same prediction irrespective of the input. This happened to me for 3 different projects over the course of last 1.5 years after which I entirely gave up on GCP.

[@constantinelinardakis8394]

4:00 to train a model

[@IvanLesnov]

How tune on offline (local)?

[@gemini22581]

Wat model should I use to train a model using my own data for conversational purposes?

[@Ma-pz5kl]

YEAP data set being clean is the raw quality ingredient for a 3 star michelin receipe. architecture comes second.

[@deviladvoc]

is this ideal for studying especially in an engineering field (specifically electrical)? i noticed even gemini advance have difficulty understanding and solving intermediate problems.

[@yoyoyoaliman]

Ngl, within the first 3 minutes, i already had no idea what you were talking about. I feel like this is the problem with the way information is provided today; we talk to others as if they already know everyhing we do, so when we say things are easier than expected, we arent accounting for the fact that it's relative to OUR understanding and may or may not be easier than expected or even understandable by other people.

[@Geohhh]

No bullshit. Actual experience. Well summed up Thank you!

[@aguiremedia]

So what a solution like this be called or marketed as if one was offering this as a service?

[@guitarhero69]

im new to this kinda thing but what does LLM mean?? im sorry for asking the dumbest question but can someone recommend more videos like this so i can get a better grasp or whats going on in the video.

[@LoucM-h4c]

LLM stands for large language models. I presume its self explanatory what that means but a popular example is ChatGPT 3.5 / 4. It runs off of a LLM that takes text input and returns output based on predicting the 'next text' / 'what text should follow'

[@hidayat1999]

​@user-db9bw5cl1e it also means that it can only chat right? It is not good for logical discussions like creating code?

[@drkeithnewton]

Hi, Steve. This is really helpful. once you use Vertex AI for your specialized LMM, is it still stored at Google or are you able to download and run locally? Are you constantly making API calls for this one use case, for example?

[@n111254789]

Did you ever find out? I can't get a straight answer either.

[@drkeithnewton]

@@n111254789 I did not. I guess I'm going to have to build one and find out myself. Wanna partner up?

[@b3owu1f]

So if you had a couple of computers with say some 3080s sitting around, is it terribly difficult to use those to train on even if it takes a few days or more? I'd be interested in playing around with my one GPU just to see what it can do since it's just sitting around doing nothing. Is python the only way/language that can be used to train models with data? As a go developer, I am rather surprised a much faster runtime language like Go isn't used for something this cpu (and gpu) intensive. With that in mind can you use CPU only if your GPU is low end or does training require GPUs? Subbed btw.. this video was very easy to follow so look forward to more.

[@N7Tonik]

I would say training will be always very much faster with a GPU as GPUs are made for parallel computing

[@JolanXBL]

All this video tells me is to get pre-made code somewhere else, get premade information somewhere else, train using pre-made software. It doesn't tell me how to do it it just says get a job get money and then hire someone else to do it for you. HOW does training work? HOW does it read a model? HOW can code identify pixels on a page?

[@STREETSURFER98]

I have some questions. After Training the Model in Google vertex you have a working object Detector. Nice! But How Do you connect the output of the Model (the coordinates of the bounding boxes) to the Code of the Website?

[@jeffdavidson5601]

Wait, did I miss the part where you show how you got the text and other Figma objects?

[@stevenanderson7046]

Sick dude, agree 100% thanks for the video!

[@aadarsh3301]

a just curious question: how does google photos recognized our face?

[@kayakMike1000]

I want a model that can take video frames from several cameras and identify pathing information.

[@aketo8082]

Thank you. How can I do that with GPT4ALL? Looking forward for an instruction, how I can work with GPT4ALL and with my own documents and information. And how to train and fine tune my data, if this is possible. Thanks.

[@Mohamed_Shokry]

Thank you for sharing this valuable information with us.

[@danielahern3560]

what headphones are you wearing? do they have a bluetooth microphone that you're using in the video?

[@dawid_dahl]

Thanks a lot for posting this. Super interesting and valuable.

[@RakeshMakhija]

Super Helpful. Thank you very much

[@RaphaelAlba]

I was hoping for information on training a LLM or SLM. Kinda disappointed it was a YOLO model. Good for beginners but money would be more well spent on working with future proof multimodal models.

[@raven_glass]

Haflway through this video and dude STILL hasn't said how to actually do this.

[@chrisder1814]

hello could you tell me what you think of the idea I had about using an API of image editing software by artificial intelligence

[@jakemeyer8188]

What company do you work for? My employer has been looking for a way to generate basic sites from Figma's. A vendor solution is just fine with them.

[@Steve8708]

Builder.io

[@jakemeyer8188]

@@Steve8708 Excellent. Hugely appreciated!!

[@costathoughts]

Steve you rock it dude !!! Please create one tool to Java legacy projects lol

[@peter102]

very good video. learnt alot

[@thalesa4715]

golben boy, golden content!

[@comicsans8278]

what do you use to create the flowcharts?

[@trilogen]

Can we run this entire thing offline? Even with the training?