By popular demand, I created a new video showing detailed steps for installing the tools and getting the API keys required for this tutorial. That guide is linked below the description.
What an excellent and clear video with high signal and no noise. So easy to follow and get the concepts without getting lost in the technicalities. Thanks - Subbed!
This is phenomenal! I've been ripping my hair out for hours on end trying to build something kind of similar, but now I can see light at the end of the tunnel... Thankyou, looking forward to more of your tutorials.
Thank you so much @willwhitehouse2750! I really appreciate your feedback! It's always a great feeling to hear I've helped someone remove an obstacle so they can get done what they're trying to achieve. Keep on going! I'd love to hear what you build and how it goes.
After watching this video, I immediately subscribed to your channel. This tutorial has exactly the right abstraction level, if aimed at developers. I am so tired to fast forward 30 minutes of instructions about how to get an API Key or how to install a tool, and then get a 2 minutes "and here happens the AI magic" lesson. Please, stick with your style. There are already so many channels with non-coders or beginners as the target audience, it is good to have your channel with a more advanced audience as a target group. It is also nice to see, that you are providing additional how-to-setup videos as a seperate service for those who are just start getting their toes wet.
I really appreciate this @steffeneichenberg1833! This is really valuable feedback! After having several people suggest I include every detailed step for installing every tool, I created another video showing how to install all this 😆I understand that we'll have folks with varying levels of experience - and, I'd love to help them all, because I think AI can help newbies really get inspired to learn to code. But I agree with you - I personally don't want to watch a video that shows me how to click on a link to go to a website, then click on a link to download the installer, then click through an installer wizard. That's basic computer literacy - nothing to do with software dev, IMHO. Curious whether you think the way I chose to handle that - the separate "how to install" video - is the way to address this. I then link that video to other tutorials that actually hopefully provide value?
Thank you for that @PiakBot! Yep, Claude artifacts is super cool! Do you have some ideas about how an AI coding assistant could be used alongside Artifacts?
I really appreciate all your feedback! I wanted to let you know that I'm seeing some common themes in the comments and I plan to try to address many of them in the coming weeks. Here's a short list of some recurring questions/suggestions: - Costs related to the LLMs being used (e.g. Claude, GPT 4, etc.) - Detailed instructions for setting up the development environment for these tutorials. - Best practices for prompting the coding assistant to get the best results. I hear you. Look for more information on all this in the coming weeks.
good content! best tutorial i've seen so far.👍👍 So prompt it to generate empty structure -> fill out the key implementation -> change the ui And he touches upon how to understand ai generated code which is superb!!
very useful! I'm glad to see a working way to do the Chain of Thought pattern. I've had challenges with my AI rewriting other parts of my code. Going step by step really helped.
@asemerci... sorry, but.. huh? Gonna assume you were kidding? That's standard procedure. And every person who releases any content that uses any sort of API key had better do the same.
Video was very informative, with great information and detail. Could you consider using a larger cursor? It might be less distracting than the big yellow dot, we cant see whats behind it. Additionally, increasing the font size would make everything a bit more readable.
I really appreciate your feedback @jaapjob! Yes, I'll be ditching the big yellow dot in future videos. I'll also try increasing font sizes wherever practical.
Hey Tim, this was excellent on so many levels! ( useful app, pace of explanation, video documentation ...) I am having a blast working on projects with the claude sonnet / aider combo as well! This is -- in my eyes -- the first time the LLM reasoning capability and context window is good enough to code entire projects. I gave up coding with LLM's for almost a year since there occured just too much hallucination ( too frustrating ). Question: The Claude context is so superb in the Projects space. I wonder if you could create something similar when you want to use your API Key instead. Basically a simple web ui with the ability to upload documents and also implement your own RAG system to get more information from the web when necessary. Also, the Claude Projects space is limited to a few files. Would be great to extend this a little more within your own app. This way you could have a great knowledge hub for your personal projects/interests. Might be attempting something like this in the next few days but maybe this could also be a great project/tutorial video for this channel! Greetings from Germany! - Dustin
I really appreciate your feedback @duhai1836! Your first paragraph corresponds exactly to my own journey with all this. Until Claude 3.5 was released, I was having mixed results - maybe GPT 4 Turbo would get things right 70% of the time. But, C 3.5 change the game! Very excited to see what we all do with this! And, who knows what the next generation of LLMs will be capable of here. I have yet to dig into Claude Projects. I will soon. Just so much going on in this space, as you know. When you run the experiment you mention, I'd love to hear what you find. Please share.
will definitely try this once i get home... I want to understand, is this more of an aider beginner tutorial or is the technique of making it stop/continue very useful? Is the web scraping function different than browser 'Save As'?
Hi @ytubeanon! I didn't necessarily intend this to be an aider beginner tutorial. I tried to strike a balance that I thought would allow folks at many different levels of experience with coding assistants (or AI for software development in general) to get a feel for how aider can be used to quickly bootstrap apps from the ground up. And also build something they can use at the same time. I chose RAG for this one, because I it's useful and lots of people want to know how to build these types of apps. As for the stop/continue part... yes, I've been experimenting with LLMs and AI coding assistants for some time and I consistently find that you have to use them in the same way we human developers code - by breaking the dev flow down into iterative tasks, each building on the last one. And, making sure you provide them enough, but not too much, context. And, not only telling them what you WANT them to do, but often telling them what you do NOT want them to do. I have yet to experience any LLM (even the mighty Claude 3.5) be able to accept a big task and generate a high-quality "chunk" of code that doesn't break the app. But, I'll admit that it takes a lot of experimentation and learning to get to the point it's all inuitive. I do intend to put together some videos that walk through the "why" behind all this and some "simple" tips and techniques. My hope is to initially teach this by demonstration. Lastly, aider's /web function is only different from the browser "Save As" option in that it lets you tell aider to directly pull one or more web pages' content into the context for the convo with the LLM. You could also copy/paste a page's contents directly into the aider chat session and instruct it to use that for context. But, that's far more tedious and error-prone. aider will do a much better job of handling it if you use its built-in scraping function. I said a lot. Please let me know if this was more confusing than enlightening.
@@CodingtheFuture-jg1he thx for the detailed reply! for what it's worth, here are 2 interesting things I read on Aider's Discord: - "I have had great success with making a markdown checklist and aider will work thru the list and even check the items off in the markdown list as it completes each coding block." - "pro tip, Agent your plan split it to agents to break down then cli Aider over and over rather than one session multi tasks...... its less efficent on every task but anything that fails it will end uf faster....We have been making leaps and bounds in coding the last 3 weeks because of some very recent changes we have discovered so I expect theres going to be some interesting announcments in a few days if we can prove our conept on some runpods."
this content is so good. thanks you so much. I'm struggling with optimizing performance of my app. Can you create the more advance content related to use AI assistant to debug and optimize kind of performance issues. Really looking for your content.
Thank you so much for your kind words @frankle3025! Using AI tools to optimize performance is a great idea! I haven't dug into that yet myself. But, I've used AI (specifically, the Claude and GPT-4 models) to debug code. Most of the time when I provide it with a copy of the error output, I get a fix within one or two attempts. Now, it needs more than the error message - it also needs to have information about your code - the files, what the app does, the libraries you're using, etc. - in its context to do a decent job of finding/fixing your bug. I've also used it to find dead code, security vulnerabilities, etc. and it works pretty well. Seems like it should also be able to find hot spots in the app. Although I haven't tried this yet myself, let me suggest this approach: 1. Focus the AI coding assistant on just the files in your codebase you suspect "could" be impacting performance. It's OK to throw in a few that may or may not be relevant. But, try not to just ask the LLM to review a large codebase all at once. If it's a small codebase (let's say less than 2,000 lines of code), you might be able to have it analyze it all at once. Most coding assistants, including aider, let you focus them so they pay more attention to specific files - for aider, that's the /add command. 2. Provide some concrete details on the performance issue you're experiencing: "When the user logs in, it usually takes over 10 seconds before they see their home page. I see THIS error message in my server logs when this occurs: ". 3. Be specific in your request for help: "Analyze the code I gave you and the server logs. Diagnose it for anything that might impact performance. Also, suggest updates I could make and how to test them." I'd love to know whether you try any of this and if you're able to solve your performance issues!
Thank you for your great videos. Could you please share how much this project costed you in terms of Claudie 3.5 Sonnet API usage? Just looking at their pricing page is not very helpful since I don't know how many input / output tokens a project like this one involves.
Thank you for that @olegdragora2557! You're correct. It's really difficult to estimate how much a certain usage of these APIs will cost. I'm going to start including that info in these tutorials in the future - maybe in the video description. It would have been pretty easy to calculate had I used aider's /token command before I closed out my aider session. While in a session, you can run that command and aider will give you the token usage stats and an estimate on your costs. Also, I could have checked my Anthropic API usage page before and after but didn't think of it this time. I'd estimate that I used around $1 for the complete RAG app. Lots of folks asking this question and also setup questions. I'll put together a quick video to review both topics soon.
@@CodingtheFuture-jg1he Thank you for a detailed answer! Looks like $5 free credits from Anthropic should be enough to get started and evaluate the approximate costs for own projects.
@@olegdragora2557 I know that money does not grow on trees, but I think you will find working with AI coding assistance is worth the money. And I am not talking theoretically. I am a software engineer with more than 40 years of experience. I am using AI coding assistance for more than a year now in a professional environment. Invaluable! And also for private projects where you try to learn something. You will learn so much faster with AI assistance. It is like having you own tutor. I have the money to spend on AI. But even if I had not, I would rather spend the time saved by using AI, to work an extra minimum wage job to get that money, than not using AI.
@@CodingtheFuture-jg1hebtw I uploaded screenshots to Sonnet 3.5 of my API usage reports from Anthropic's site. I told it to add up the 'Cost' column and it worked
Thank you @yurijmikhassiak7342! I can't comment on Cursor yet. Haven't tried it, but I keep hearing great things about it, so I'll be giving it a whirl soon. Have you used Cursor? If so, what do you think so far?
@@CodingtheFuture-jg1he I love it. I am not a developer but was able to build a simple app using next.js vercel posgre hasura refine. But it takes time to learn the correct prompts as this is not a typical setup. But after you set it up if can help you write apps much faster with much less code as this is almost low code development. Less code - less AI errors, faster development.
aider creates a map of your repo - the filenames, function/method signatures, etc. - to provide as context to the LLM as it generates code. Here's an explanation: aider.chat/2023/10/22/repomap.html. Haven't looked into what the "Repo-map: using XXX tokens" output line means, but I suspect it's telling you the number of tokens it's estimating that just the current map of your codebase context will consume on the next LLM request.
Hi @chronicallychill9979! I've only used Perplexity once or twice. But, unless I'm mistaken, you don't setup a Claude API key for Perplexity to use - I think it just uses its own API keys and lets you select Claude from its LLM options. So, you don't really get a Claude key from Perplexity. You have to create an Anthropic account and create your own key. But, that's really easy. You can see the "Step-by-Step Development Environment" tutorial I published just a couple of days ago. Hope this helps.
Thanks. Wondering whether it is currently faster to code a program with an AI Assistant by 1) Prompting Aider (with ChatGPT integration) to automate building code, troublshooting within terminal 2) Asking ChatGPT for code using that code, then copying any errors back to ChatGPT, pasting back fixed code (repeating this process n times) Assuming that AI assistants are built for trying to be faster than the second option, but are they?
Hey @michaelmorgenstern5544! IMHO, option 1 is by far the most productive workflow! I've tried both. I'm a huge believer in reducing context switching. PLUS, aider adds a ton of functionality and smarts on top of the LLM that no AI chatbot (yes, even Claude.ai) simply cannot provide. Copy/paste of code between ANY chatbot, when you have a tool that's hyper-tuned specifically for coding tasks, can auto-create and update any code in place inside your project AND handle the git commits for you... I mean, I'm not really sure why anyone is still bothering with the chatbot generate/copy/paste/back to chatbot/copy/paste/... thing at this point. Chatbots, like ChatGPT, Claude.ai and meta.ai are really useful for a lot of code-related tasks - but, when it comes to actually generating of updating code to be used in your development repo, when you really look at it and then try both, I think you'll see - it's like night and day. It's simply impossible to work as fast, with as few errors and as seamlessly using chatbots to generate your code.
Great video, thanks! I am wondering if TDD and AI-supported coding go well together. I really like the structured thinking and the high-quality unit-tests that ensure long-term maintainability that result from TDD done well. I noticed the steps you do here are bigger than the steps you‘d usually take when using TDD. Could you maybe demo building a feature using aider and TDD?
I really appreciate your feedback @pascalgugenberger2116! I've been thinking about how using AI coding assistants - especially when we're generating large blocks of code at once, as in this tutorial - jives with TDD. I'll admit that I haven't been a TDD practitioner (at least not consistently) and so I haven't really built that natural intuition that would be required to do the topic justice at this point. But it's important and needs to be addressed. I'm adding it to my backlog. If you have any concreate ideas for a POC to start to experiment with this idea, please do share. There's obviously more to software engineering than scaffolding a new project, as I do in this video 😉
I really appreciate your feedback @T-rotten! If you'd like to try out any of the aider tutorials with llama, that's super easy. Doesn't change any of the tutorial steps. It will probably change the quality of the code generated however, because right now, no LLM beats Clauder 3.5 Sonnet when it comes to coding. Currently the best way to connect aider to llama is using it on Groq (groq.com/). For now, they let you use their APIs for free - that will likely change at some point. You go there, create an API key, then set the GROQ_API_KEY environment variable in your aider terminal BEFORE starting up aider. This page shows how to set this: aider.chat/docs/llms/groq.html.
Hi @muffawuffaman! Please see my "AI Coding Assistant Costs Revealed: aider and Claude 3.5 Sonnet" video where I showed several before/after Anthropic API billing snapshots. The costs for this video are in there, plus costs for a previous video.
Awesome. So we need to create a backend and then focus on frontend. Is my understanding correct? You are prompting from a developer's perspective but will we get an usable code if we prompt like a normal user perspective as well? Also how much the API costs for this project.
Thank you for that @sivakumarm3569! I'm not suggesting that you have to start with the backend, but I'd highly recommend that you do. I'd be interested in an experiment where we flipped it around - 1) aider build me a UI for a RAG app that does XYZ then 2) now build me an RAG API the UI can consume. What I suspect is that it won't work as well, but who knows. Lots of people asking about the costs. I didn't track them closely this time but will in the future and include that info. I'd estimate that I used around $1 for the complete RAG app.
@@CodingtheFuture-jg1he Thanks for the response. Since I am a non developer, I couldn't think of the functionality of the backend before the front end because that's how we see the app (like just the frontend) lol. I think it's time to experiment with this. Thanks again for the video. Subscribed.
I replicated this and ran into rate limits with GPT-4o. I tried 3.5 turbo and still failed. I got frustrated with this and had Aider build in the option to switch LLM at the beginning. I switched to Anthropic, and finally there were no annoying rate limits. Also,I am making this work on Mac so it is doable. Next will be to add local LLM support and then upload files. After that I will rewrite this to work with JS framework maybe.
Wow @darklen14! Talk about adapting on the fly! Sorry you ran into the rate limiting. On the other hand, I'm sure it will help you in the long run. I'd love to hear/see what you come up with for the JS implementation! As for local LLMs... I had such high hopes for them about 6 months ago for coding. I've tried pretty much every one that's supposedly "good" at coding. The current state is quite disappointing though, once you've used something like Claude 3.5 (talking specifically about integrating the LLM into your coding). Having said that, my new high-powered Mac is arriving any day and I'll certainly be re-testing local models like qwen 2, DeepSeek 2, etc. At the very least, no rate limiting locally (unless your computer starts smoking that is) 😆 The small language models that will suffice on lower end hardware for coding are coming. I'm rooting for them!
Thank you for letting me know @darklen14! Sorry for the delayed response. Anthropic was having some API issues around the time you posted this. I hope you stuck with it while they ironed that out. Please let me know if it's still an issue.
lol I built this a little too easily, I downloaded the transcript from the TH-cam video, asked aider to build what it described... blam, now I have a working app, just had to change `text-davinci-003` to `gpt-4o-mini`... I'll have to slow aider down if I'm going to learn
@@CodingtheFuture-jg1he yeah, because this simply is too good. I actually ended up watching the video on my monitor. But I am sure you will get more viewing time if you produce with the fact that 9 out of 10 watch on their phones.
By popular demand, I created a new video showing detailed steps for installing the tools and getting the API keys required for this tutorial. That guide is linked below the description.
I think I'm just about done with programming.
Only so many needed at this point...
What an excellent and clear video with high signal and no noise. So easy to follow and get the concepts without getting lost in the technicalities. Thanks - Subbed!
This is phenomenal! I've been ripping my hair out for hours on end trying to build something kind of similar, but now I can see light at the end of the tunnel... Thankyou, looking forward to more of your tutorials.
Thank you so much @willwhitehouse2750! I really appreciate your feedback! It's always a great feeling to hear I've helped someone remove an obstacle so they can get done what they're trying to achieve. Keep on going! I'd love to hear what you build and how it goes.
Finally, a TH-camr who uses AI tools in depth and makes projects with them. Keep up the good work! Subscribed!
Thank you so much for your kind words @Dreamslol!
After watching this video, I immediately subscribed to your channel. This tutorial has exactly the right abstraction level, if aimed at developers. I am so tired to fast forward 30 minutes of instructions about how to get an API Key or how to install a tool, and then get a 2 minutes "and here happens the AI magic" lesson. Please, stick with your style. There are already so many channels with non-coders or beginners as the target audience, it is good to have your channel with a more advanced audience as a target group. It is also nice to see, that you are providing additional how-to-setup videos as a seperate service for those who are just start getting their toes wet.
i find that he finds a good middle ground! But yes, please keep exactly the same delivery style! ;)
I really appreciate this @steffeneichenberg1833! This is really valuable feedback! After having several people suggest I include every detailed step for installing every tool, I created another video showing how to install all this 😆I understand that we'll have folks with varying levels of experience - and, I'd love to help them all, because I think AI can help newbies really get inspired to learn to code. But I agree with you - I personally don't want to watch a video that shows me how to click on a link to go to a website, then click on a link to download the installer, then click through an installer wizard. That's basic computer literacy - nothing to do with software dev, IMHO. Curious whether you think the way I chose to handle that - the separate "how to install" video - is the way to address this. I then link that video to other tutorials that actually hopefully provide value?
Nice going!
Claude 3.5 artifacts is a next game changer.
I hope your new video maybe apply Aider ai with Claude artifacts.
Thank you for that @PiakBot!
Yep, Claude artifacts is super cool! Do you have some ideas about how an AI coding assistant could be used alongside Artifacts?
I really appreciate all your feedback! I wanted to let you know that I'm seeing some common themes in the comments and I plan to try to address many of them in the coming weeks. Here's a short list of some recurring questions/suggestions:
- Costs related to the LLMs being used (e.g. Claude, GPT 4, etc.)
- Detailed instructions for setting up the development environment for these tutorials.
- Best practices for prompting the coding assistant to get the best results.
I hear you. Look for more information on all this in the coming weeks.
good content! best tutorial i've seen so far.👍👍
So prompt it to generate empty structure -> fill out the key implementation -> change the ui
And he touches upon how to understand ai generated code which is superb!!
Thank you so much for your kind words @poisonza!
very useful! I'm glad to see a working way to do the Chain of Thought pattern. I've had challenges with my AI rewriting other parts of my code. Going step by step really helped.
Thank you for that @JohnMitchellCalif!
YES!! :-). keep the momentum Brother. you are going to dominate this niche (AI coding).
Thank you so much for your kind words @programmingsiri5007!
Great work, keep going and educate us more with those tutorials 👏
Thank you for that @ahmedzanaty8249! Will do!
Nice job, great intro and you pulled it together nicely.
Thank you so much @DaveFriedel!
My new favorite channel 🙏🏽💎 you earned a sub sir
Thank you so much for your kind words @J3R3MI6!
"Don't worry I'll revoke this API key before this video goes live" 😢 Thanks for the video!
@asemerci... sorry, but.. huh? Gonna assume you were kidding? That's standard procedure. And every person who releases any content that uses any sort of API key had better do the same.
@@CodingtheFuture-jg1heYou are right. I was just kidding.
Great work 👏
Thank you so much @MuhanadAbulHusn!
Awesome video.. I was looking for something exactly like this for a project of mine. Now i have a bunch of ideas of other projects 🙂
Thank you for that @Bugge67! I hope it helps you execute your ideas 🙂
I really like this kind of videos, more would be appreciated :D
Thank you so much @BorisHrzenjak!
Video was very informative, with great information and detail. Could you consider using a larger cursor? It might be less distracting than the big yellow dot, we cant see whats behind it. Additionally, increasing the font size would make everything a bit more readable.
I really appreciate your feedback @jaapjob! Yes, I'll be ditching the big yellow dot in future videos. I'll also try increasing font sizes wherever practical.
Thanks for the video. Keep up the good content.
I really appreciate your feedback @hannesg.3586!
Hey Tim, this was excellent on so many levels! ( useful app, pace of explanation, video documentation ...)
I am having a blast working on projects with the claude sonnet / aider combo as well! This is -- in my eyes -- the first time the LLM reasoning capability and context window is good enough to code entire projects. I gave up coding with LLM's for almost a year since there occured just too much hallucination ( too frustrating ).
Question: The Claude context is so superb in the Projects space. I wonder if you could create something similar when you want to use your API Key instead. Basically a simple web ui with the ability to upload documents and also implement your own RAG system to get more information from the web when necessary. Also, the Claude Projects space is limited to a few files. Would be great to extend this a little more within your own app. This way you could have a great knowledge hub for your personal projects/interests.
Might be attempting something like this in the next few days but maybe this could also be a great project/tutorial video for this channel!
Greetings from Germany! - Dustin
I really appreciate your feedback @duhai1836! Your first paragraph corresponds exactly to my own journey with all this. Until Claude 3.5 was released, I was having mixed results - maybe GPT 4 Turbo would get things right 70% of the time. But, C 3.5 change the game! Very excited to see what we all do with this! And, who knows what the next generation of LLMs will be capable of here.
I have yet to dig into Claude Projects. I will soon. Just so much going on in this space, as you know. When you run the experiment you mention, I'd love to hear what you find. Please share.
will definitely try this once i get home... I want to understand, is this more of an aider beginner tutorial or is the technique of making it stop/continue very useful? Is the web scraping function different than browser 'Save As'?
Hi @ytubeanon! I didn't necessarily intend this to be an aider beginner tutorial. I tried to strike a balance that I thought would allow folks at many different levels of experience with coding assistants (or AI for software development in general) to get a feel for how aider can be used to quickly bootstrap apps from the ground up. And also build something they can use at the same time.
I chose RAG for this one, because I it's useful and lots of people want to know how to build these types of apps.
As for the stop/continue part... yes, I've been experimenting with LLMs and AI coding assistants for some time and I consistently find that you have to use them in the same way we human developers code - by breaking the dev flow down into iterative tasks, each building on the last one. And, making sure you provide them enough, but not too much, context. And, not only telling them what you WANT them to do, but often telling them what you do NOT want them to do.
I have yet to experience any LLM (even the mighty Claude 3.5) be able to accept a big task and generate a high-quality "chunk" of code that doesn't break the app. But, I'll admit that it takes a lot of experimentation and learning to get to the point it's all inuitive.
I do intend to put together some videos that walk through the "why" behind all this and some "simple" tips and techniques. My hope is to initially teach this by demonstration.
Lastly, aider's /web function is only different from the browser "Save As" option in that it lets you tell aider to directly pull one or more web pages' content into the context for the convo with the LLM. You could also copy/paste a page's contents directly into the aider chat session and instruct it to use that for context. But, that's far more tedious and error-prone. aider will do a much better job of handling it if you use its built-in scraping function.
I said a lot. Please let me know if this was more confusing than enlightening.
@@CodingtheFuture-jg1he thx for the detailed reply! for what it's worth, here are 2 interesting things I read on Aider's Discord:
- "I have had great success with making a markdown checklist and aider will work thru the list and even check the items off in the markdown list as it completes each coding block."
- "pro tip, Agent your plan split it to agents to break down then cli Aider over and over rather than one session multi tasks......
its less efficent on every task but anything that fails it will end uf faster....We have been making leaps and bounds in coding the last 3 weeks because of some very recent changes we have discovered so I expect theres going to be some interesting announcments in a few days if we can prove our conept on some runpods."
this content is so good. thanks you so much. I'm struggling with optimizing performance of my app. Can you create the more advance content related to use AI assistant to debug and optimize kind of performance issues. Really looking for your content.
Thank you so much for your kind words @frankle3025! Using AI tools to optimize performance is a great idea! I haven't dug into that yet myself. But, I've used AI (specifically, the Claude and GPT-4 models) to debug code. Most of the time when I provide it with a copy of the error output, I get a fix within one or two attempts. Now, it needs more than the error message - it also needs to have information about your code - the files, what the app does, the libraries you're using, etc. - in its context to do a decent job of finding/fixing your bug.
I've also used it to find dead code, security vulnerabilities, etc. and it works pretty well. Seems like it should also be able to find hot spots in the app.
Although I haven't tried this yet myself, let me suggest this approach:
1. Focus the AI coding assistant on just the files in your codebase you suspect "could" be impacting performance. It's OK to throw in a few that may or may not be relevant. But, try not to just ask the LLM to review a large codebase all at once. If it's a small codebase (let's say less than 2,000 lines of code), you might be able to have it analyze it all at once. Most coding assistants, including aider, let you focus them so they pay more attention to specific files - for aider, that's the /add command.
2. Provide some concrete details on the performance issue you're experiencing: "When the user logs in, it usually takes over 10 seconds before they see their home page. I see THIS error message in my server logs when this occurs: ".
3. Be specific in your request for help: "Analyze the code I gave you and the server logs. Diagnose it for anything that might impact performance. Also, suggest updates I could make and how to test them."
I'd love to know whether you try any of this and if you're able to solve your performance issues!
Thank you for your great videos.
Could you please share how much this project costed you in terms of Claudie 3.5 Sonnet API usage?
Just looking at their pricing page is not very helpful since I don't know how many input / output tokens a project like this one involves.
Thank you for that @olegdragora2557!
You're correct. It's really difficult to estimate how much a certain usage of these APIs will cost. I'm going to start including that info in these tutorials in the future - maybe in the video description. It would have been pretty easy to calculate had I used aider's /token command before I closed out my aider session. While in a session, you can run that command and aider will give you the token usage stats and an estimate on your costs. Also, I could have checked my Anthropic API usage page before and after but didn't think of it this time.
I'd estimate that I used around $1 for the complete RAG app.
Lots of folks asking this question and also setup questions. I'll put together a quick video to review both topics soon.
@@CodingtheFuture-jg1he Thank you for a detailed answer!
Looks like $5 free credits from Anthropic should be enough to get started and evaluate the approximate costs for own projects.
@@olegdragora2557 I know that money does not grow on trees, but I think you will find working with AI coding assistance is worth the money. And I am not talking theoretically. I am a software engineer with more than 40 years of experience. I am using AI coding assistance for more than a year now in a professional environment. Invaluable! And also for private projects where you try to learn something. You will learn so much faster with AI assistance. It is like having you own tutor. I have the money to spend on AI. But even if I had not, I would rather spend the time saved by using AI, to work an extra minimum wage job to get that money, than not using AI.
@@CodingtheFuture-jg1hebtw I uploaded screenshots to Sonnet 3.5 of my API usage reports from Anthropic's site. I told it to add up the 'Cost' column and it worked
Thanks. How does it compate to cursor AI for code quality? As I see both tools have similar features.
Thank you @yurijmikhassiak7342!
I can't comment on Cursor yet. Haven't tried it, but I keep hearing great things about it, so I'll be giving it a whirl soon. Have you used Cursor? If so, what do you think so far?
@@CodingtheFuture-jg1he I love it. I am not a developer but was able to build a simple app using next.js vercel posgre hasura refine. But it takes time to learn the correct prompts as this is not a typical setup. But after you set it up if can help you write apps much faster with much less code as this is almost low code development. Less code - less AI errors, faster development.
What's the repo-map in out the aider output mean?
aider creates a map of your repo - the filenames, function/method signatures, etc. - to provide as context to the LLM as it generates code. Here's an explanation: aider.chat/2023/10/22/repomap.html. Haven't looked into what the "Repo-map: using XXX tokens" output line means, but I suspect it's telling you the number of tokens it's estimating that just the current map of your codebase context will consume on the next LLM request.
Could I use my access to claude from perplexity for this?
Hi @chronicallychill9979! I've only used Perplexity once or twice. But, unless I'm mistaken, you don't setup a Claude API key for Perplexity to use - I think it just uses its own API keys and lets you select Claude from its LLM options. So, you don't really get a Claude key from Perplexity. You have to create an Anthropic account and create your own key. But, that's really easy. You can see the "Step-by-Step Development Environment" tutorial I published just a couple of days ago. Hope this helps.
Thanks.
Wondering whether it is currently faster to code a program with an AI Assistant by
1) Prompting Aider (with ChatGPT integration) to automate building code, troublshooting within terminal
2) Asking ChatGPT for code using that code, then copying any errors back to ChatGPT, pasting back fixed code (repeating this process n times)
Assuming that AI assistants are built for trying to be faster than the second option, but are they?
Hey @michaelmorgenstern5544! IMHO, option 1 is by far the most productive workflow! I've tried both. I'm a huge believer in reducing context switching. PLUS, aider adds a ton of functionality and smarts on top of the LLM that no AI chatbot (yes, even Claude.ai) simply cannot provide.
Copy/paste of code between ANY chatbot, when you have a tool that's hyper-tuned specifically for coding tasks, can auto-create and update any code in place inside your project AND handle the git commits for you... I mean, I'm not really sure why anyone is still bothering with the chatbot generate/copy/paste/back to chatbot/copy/paste/... thing at this point.
Chatbots, like ChatGPT, Claude.ai and meta.ai are really useful for a lot of code-related tasks - but, when it comes to actually generating of updating code to be used in your development repo, when you really look at it and then try both, I think you'll see - it's like night and day. It's simply impossible to work as fast, with as few errors and as seamlessly using chatbots to generate your code.
answer?
Great video, thanks!
I am wondering if TDD and AI-supported coding go well together. I really like the structured thinking and the high-quality unit-tests that ensure long-term maintainability that result from TDD done well.
I noticed the steps you do here are bigger than the steps you‘d usually take when using TDD. Could you maybe demo building a feature using aider and TDD?
I really appreciate your feedback @pascalgugenberger2116! I've been thinking about how using AI coding assistants - especially when we're generating large blocks of code at once, as in this tutorial - jives with TDD. I'll admit that I haven't been a TDD practitioner (at least not consistently) and so I haven't really built that natural intuition that would be required to do the topic justice at this point. But it's important and needs to be addressed. I'm adding it to my backlog. If you have any concreate ideas for a POC to start to experiment with this idea, please do share. There's obviously more to software engineering than scaffolding a new project, as I do in this video 😉
Would like to see something that utilizes llama as well
I really appreciate your feedback @T-rotten! If you'd like to try out any of the aider tutorials with llama, that's super easy. Doesn't change any of the tutorial steps. It will probably change the quality of the code generated however, because right now, no LLM beats Clauder 3.5 Sonnet when it comes to coding. Currently the best way to connect aider to llama is using it on Groq (groq.com/). For now, they let you use their APIs for free - that will likely change at some point. You go there, create an API key, then set the GROQ_API_KEY environment variable in your aider terminal BEFORE starting up aider. This page shows how to set this: aider.chat/docs/llms/groq.html.
Thank you for the feedback. I will look into Claud. I’m very new to this and your presentation methods are fantastic. Please keep posting tutorials.
Amazing tutorial 🔥 I’m not religious but when I am, I refer to ppl like you for doing gods work 🫡
Thank you so much for your kind words @marioa6942!
Can you share the total cost of the Anthropic calls that Aider used while developing the software?
Hi @muffawuffaman! Please see my "AI Coding Assistant Costs Revealed: aider and Claude 3.5 Sonnet" video where I showed several before/after Anthropic API billing snapshots. The costs for this video are in there, plus costs for a previous video.
Awesome. So we need to create a backend and then focus on frontend. Is my understanding correct? You are prompting from a developer's perspective but will we get an usable code if we prompt like a normal user perspective as well? Also how much the API costs for this project.
Thank you for that @sivakumarm3569!
I'm not suggesting that you have to start with the backend, but I'd highly recommend that you do. I'd be interested in an experiment where we flipped it around - 1) aider build me a UI for a RAG app that does XYZ then 2) now build me an RAG API the UI can consume. What I suspect is that it won't work as well, but who knows.
Lots of people asking about the costs. I didn't track them closely this time but will in the future and include that info.
I'd estimate that I used around $1 for the complete RAG app.
@@CodingtheFuture-jg1he Thanks for the response. Since I am a non developer, I couldn't think of the functionality of the backend before the front end because that's how we see the app (like just the frontend) lol. I think it's time to experiment with this. Thanks again for the video. Subscribed.
I replicated this and ran into rate limits with GPT-4o. I tried 3.5 turbo and still failed. I got frustrated with this and had Aider build in the option to switch LLM at the beginning. I switched to Anthropic, and finally there were no annoying rate limits. Also,I am making this work on Mac so it is doable. Next will be to add local LLM support and then upload files. After that I will rewrite this to work with JS framework maybe.
Wow @darklen14! Talk about adapting on the fly! Sorry you ran into the rate limiting. On the other hand, I'm sure it will help you in the long run. I'd love to hear/see what you come up with for the JS implementation! As for local LLMs... I had such high hopes for them about 6 months ago for coding. I've tried pretty much every one that's supposedly "good" at coding. The current state is quite disappointing though, once you've used something like Claude 3.5 (talking specifically about integrating the LLM into your coding). Having said that, my new high-powered Mac is arriving any day and I'll certainly be re-testing local models like qwen 2, DeepSeek 2, etc. At the very least, no rate limiting locally (unless your computer starts smoking that is) 😆
The small language models that will suffice on lower end hardware for coding are coming. I'm rooting for them!
Ayaya
I can follow along with this tutorial and have credits for anthropic but every time I submit a prompt I get AnthropicException - Rate exceeded.
Thank you for letting me know @darklen14! Sorry for the delayed response. Anthropic was having some API issues around the time you posted this. I hope you stuck with it while they ironed that out. Please let me know if it's still an issue.
Agents or assistants?
Sorry, but I'm not clear on your question @morespinach9832!
@@CodingtheFuture-jg1he are you speaking of agents as in agentic AI? How are “assistants” different from agents.
lol I built this a little too easily, I downloaded the transcript from the TH-cam video, asked aider to build what it described... blam, now I have a working app, just had to change `text-davinci-003` to `gpt-4o-mini`... I'll have to slow aider down if I'm going to learn
That, my friend, was a BRILLIANT move on your part @ytubeanon! Love it!
Higher resolution & bigger screen... Pleaseeee..... AI for visualization with aide... Please🙏
Thank you for that @islamahmadodeh8732! Please see my response to your same comment on the "AI Coding Assistant Costs" video.
Cant see anything. Remember 90 % are watching yt on our phones
I really appreciate your feedback @claussa! Taking your feedback into account as I learn how to produce better videos.
@@CodingtheFuture-jg1he yeah, because this simply is too good. I actually ended up watching the video on my monitor. But I am sure you will get more viewing time if you produce with the fact that 9 out of 10 watch on their phones.