Your videos are awesome. No fluff, no silly intros, no stock videos for the sake of having stock videos. Just stuff I can learn and apply. Never change.
"Let's talk about it so I can get some views" That is the most honest answer I've seen from a RUclipsr xD. I'll subscribe purely from the standpoint of him not trying to bull shit anyone.
@@FirstYokai Doesn't changes anything in reality and practice xd, the product is hardly maintable doesn't matter if its written by a human or " a robot" xd
This uses Claude which tends to not rewrite long segments of code so if you're not carefully you could end up clearing all the core you had initially and good luck getting that back in the same structure
facts i messed up and it feel utter all the code after being basically done lol. learned my lesson and started putting the code in my notes and just put its changes in my code in my notes then posted the entire thing back in the file
It looks more and more to me like we won't get replaced by AI anytime soon. Especially in a field that is constantly changing and evolving it just can't keep up. These AI tools still shine the most when used as assistants that help you instead of letting them do everything. That's why copilot and cursor will be in demand where this product seems like it's trying to be more autonomous which I'm not sure will work with the current state of AI. Honestly it looks more like a cool gimmick than a real product you can use.
I don't really get how it can seem more and more that the AI won't replace us when each new iteration of it is more capable than the last. Just because the current version is not yet a viable software engineer can't you just look at the trajectory? Humans aren't magic, there is no reason why an AI can learn to manage a Next.js project (something that just a year ago a bunch of people claimed it could never do) but not databases or cloud infrastructure or whatever else people are gonna claim next is the sole purview of the human mind.
@@phsopher The biggest problem with LLMs is they need to be trained to stay relevant and need a huge amount of context to be useful in big projects. Just see that this one is using experimental app directory for NextJS because it’s not trained with the latest version. The package version can be 14 or 15 for all it cares it still knows the version it was trained on and will start to hallucinate at some point and this goes for any library or framework. Im not saying we wont get to a point where AIs will be able to replace us Im just not seeing this happening with LLMs in the current state they are. Yes they will get better but for now I’ll stick to my belief that they wont go further than being very good assistants.
@@eduardstefan6833 Unless you include latest documentation in the context which nowadays can be millions of tokens, literally several text books' worth, far beyond what a human brain can hope to hold. Nothing forbids an implementation where a model goes to the Next.js web page, grabs the latest documentation and implements that. In fact even this model was able to use the newer version once prompted to do so. This is just another one of those insurmountable obstacles that are surmounted and forgotten six months later.
Also basically every LLM coding demo I've seen is just web development even though there are so many more applications/fields in software development. In most of these, AI fails miserably. The training set is just the largest with web dev
@@eduardstefan6833 Unless you include latest documentation in the context which nowadays can be millions of tokens, literally several text books'worth, far beyond what a human brain can hope to hold. Nothing forbids an implementation where a model goes to the Next.js web page, grabs the latest documentation and implements it. Even this model was able to use the newer version once prompted to do so. This is another one of those insurmountable obstacles that are surmounted and forgotten six months later.
I think it is still the best to use it as a tool, not for the whole process start to finish. I for example just managed to implement a whole new page and API route that uses Replicate into already existing app, using bolt and v0. Definitely speed up the process since the app was already using tailwind, shadcn, zod etc. Bolt helped me with the API route, I just pasted in the Replicate schema and docs, and then I redefined the UI using v0. Ofcourse it needed some manual work, to get it just right and how I wanted it :D
But why not just use cursor directly in your editor? I don’t see a good reason to use v0 over cursor. Cursor can figure out shadcn components just fine
Thank you for testing this more in depth. I was hyped on the announcement but had my reservations. You highlighting the obvious difference between web containers and the desktop environments cleared my doubts and I know this won't fit my use-case but can definitely serve the use-case for my mentees who are learning basics and they want to see things working before diving deeper.
All comments saying that coders will never be replaced, you’re actually right stop whining. It’s just a great tool for ppl who are not good at coding to make prototypes fast
👀 The problem with generating code is usually that when things start going wrong (which they inevitably do), AI will start going in circles and desperately try different solutions and make an even bigger mess of it. Great for generating boilerplate and starter code though, saves a lot of time.
The AI tooling we have right now it's pretty good, if it continue to get cheaper and small when you can run locally or send millions of requests per cents I think the future will be pretty interesting.
Cody, you can acutally just open dev tools -> right click the html tag -> and save as a screenshot. With that you get the entire site, no matter how height, as a screenshot.
I hammered on it quite a bit, and in the end asked for a refund. I felt it needs more time in the oven. Wish there were save stages for when it works well, before your next prompt drives your work off the cliff. Hope you are well amigo.
Replit has a feature where you can rollback, but you still end up in circular hell. Webapps are not easily devd out but things like webpages are much easier with AI. But then after you build your simple webpage, who's gonna change the SEO objectives on a monthly basis? Not non-programmers.
I have tried several AI tools now that claim they can “build from scratch “ none of them are there yet. My experience has been a lot like what is demonstrated in this video. But I think we are close. Within 2 years we will have sophisticated agents that can go from idea to MVP very quickly. Which will allow us devs/indie hackers to fail fast hopefully into success
00:00 AI generates apps from prompts. 01:09 Initial code attempts often fail. 02:17 Best to start with templates. 03:36 AI can refactor existing code. 05:14 Long prompts increase error likelihood. 06:34 AI versions can be outdated. 07:52 Application startup times are slow. 08:40 Generated UI includes error handling. 10:22 Database interactions often fail. 12:21 Environment compatibility hinders development. 13:59 Image processing fails in web containers. 15:27 Socket errors block app progress. 16:11 Local coding preferred over web containers. 16:55 Diffs needed for code understanding. Summary by GPT Breeze
Thanks Cody yes grab some of the code and paste locally. We still need a deep understanding of what is going on. So we still have jobs. I personally dont think they will get it ever.
I don't want to sound negative or what! But this is totally not worth the hype. The developer is greedy af to keep increasing the monetary scheme, from 9 usd package to topping up with 10$ now with up to 100 USD/ month. What do they think when releasing an idiot agentic app which can not simply apply new buttons' colors; having issues with getting lost in a hierarchical of similar folders created by itself or having a tendency of instead of fixing stuff, the AI delete a whole bunch of components without thinking. And what's more? There is no going back to the point before it deletes stuff.
I tried it right after vite conf and it broke on first prompt.... started blasting code directly into the chat thread instead of the editor. It got stuck in a loop trying to fix that first problem, wrote a few more bugs while doing that, and then I was out of the free 150k tokens. At that rate you'd run out of the 10M tokens it gives you in the monthly free tier in about a week of fulltime work days.
I had similar thoughts. It's good for beginners but once you to get to the potatoes, you end up running into some issues. Curious to find out what you think of Replit Agent. Been messing around with that a bit lately
> try new ai tool > ask it to do something > get to a point where nothing works > get angry and spend an hour or two trying to fix it That is Replit. Amateur night.
@@philamavikane9423 on the first project, a web app, Replit suggested that we create user accounts with credentials. It couldn't complete that task and wasted literally hours of my time. I wasn't looking at the code at all, I was just letting Replit grind itself into the ground. It couldn't get past the login screen. Then I started over and gave it a very detailed prompt with a different web app project and it did much better...for awhile. Two weeks later the Replit is now permanently hung up. It somehow killed itself where it can't even answer prompts. Tech support was worthless. But here's a huge pitfall of Replit: it runs into the same token problems as whatever are at the limits of the LLM underpinning it. Ergo, if your project on Replit is complicated/larger than something trivial, you are likely to run into the token limit without Replit telling you that. The end result are hallucinations or Replit repeating the same task over and over and "can't fix" whatever bug you have found. The lack of self-examination by the Replit--it just churns through tokens trying to solve a problem without any realization that it will run out of tokens--creates a shitstorm of problems for obvious reasons. I first noticed this problem with Claude 3.5 with a web app I was building. As you approach token limit with Claude in a single chat when you are using Artifact, you will start to get warnings and eventually you will be screwed. LLMs like Claude absolutely rely on being able to read the entire chat for every prompt, so the longer the single chat is (by tokens!), the more you risk literally running into a wall and your chat ends in the middle of a coding experience. I don't remember the token limit for a single chat, but it's probably somewhere. And when you are uploading screenshots in LLMs to debug, imagine how many more tokens those use up. Running Replit or Claude or Bolt or any LLM to try to code out a production level, multi featured application is like starting a race with one tank of gas that you really can't refill.
I wonder what's going to happen 5 years from now when there's a bunch of new ways to do things with like updated versions of react, and then there's the old ways of doing things that nobody uses anymore. Will the AI get confused and give obsolete suggestions? Also, the LLMs will probably have to start to learn using AI generated code, and that will be a mess lol
I think often you need to provide documentation for specific versions of you want accurate output, but honestly if the model has been trained on an older model, it might be hard to have it unlearn
Offtopic: lucia auth author just announced deprecation of lucia-auth package. Are you planning to rewrite your auth using custom API (as he suggests) or switch to using smth like better-auth? In any case - It would be great if you would share your journey, since I was planning to use lucia auth in prod and now I think I will use guides provided by it's author to roll my own auth with oslo and arctic (basically lucia).
Cursor is unusable if you have a workload. It will time out non stop. Company won't tell you they have everyone capped. They say "experiencing issues". It's only for casual use.
Cody hi. Tbh it's NOT ready for primetime. It can't even generate a next.js 15 app even when told to do so explicitely. It instead just goes off and build a 13.5 one ugh. Keeps on stumbling over and over again in errors it creates on its own. Really good project that integrates well with stackblitz but the main AI is horrendously broken
man and you didnt try to deploy and then come back and try to edit that project. idk if it was me but I was not able to do that without starting the project again. I personally don't think is better than cursor yet
> try new ai tool > ask it to do something > get to a point where nothing works > get angry and spend an hour or two trying to fix it > give up and chastise myself for not just coding it myself
Yeah it's pretty rubbish unless you want to build a sticky notes app or something, wasted a months credits just trying to fix it's own errors in building a simple audio player app, cancelling my subscription and going back to actual hand coding, job done in 3 hours and it works
Your videos are awesome. No fluff, no silly intros, no stock videos for the sake of having stock videos. Just stuff I can learn and apply. Never change.
"Let's talk about it so I can get some views"
That is the most honest answer I've seen from a RUclipsr xD. I'll subscribe purely from the standpoint of him not trying to bull shit anyone.
I'm afraid of the future, where you have tons of rotten codebases managed by no code llm kiddies
That future is already here, there are tons of code bases written by humans who don’t understand how to write maintainable code.
@@WebDevCody Might be true, but they at least somewhat know how to write code :D
@@FirstYokai Wrong. They know how to mash on keyboard until error messages stop appearing or appear in non-critical amount.
If you had worked in a append only codebase you know the future is now.
At least we can force the AI to document the code
@@FirstYokai Doesn't changes anything in reality and practice xd, the product is hardly maintable doesn't matter if its written by a human or " a robot" xd
maybe bolt should use bolt to add diff feature and see how it goes when reviewing.
lmao I love how straight up you are at the start 😂 the amount of new AI tools coming out every week at this point
I knew you will do a review about it! Thank you so much!
This uses Claude which tends to not rewrite long segments of code so if you're not carefully you could end up clearing all the core you had initially and good luck getting that back in the same structure
facts i messed up and it feel utter all the code after being basically done lol. learned my lesson and started putting the code in my notes and just put its changes in my code in my notes then posted the entire thing back in the file
It looks more and more to me like we won't get replaced by AI anytime soon. Especially in a field that is constantly changing and evolving it just can't keep up. These AI tools still shine the most when used as assistants that help you instead of letting them do everything. That's why copilot and cursor will be in demand where this product seems like it's trying to be more autonomous which I'm not sure will work with the current state of AI. Honestly it looks more like a cool gimmick than a real product you can use.
I don't really get how it can seem more and more that the AI won't replace us when each new iteration of it is more capable than the last. Just because the current version is not yet a viable software engineer can't you just look at the trajectory? Humans aren't magic, there is no reason why an AI can learn to manage a Next.js project (something that just a year ago a bunch of people claimed it could never do) but not databases or cloud infrastructure or whatever else people are gonna claim next is the sole purview of the human mind.
@@phsopher The biggest problem with LLMs is they need to be trained to stay relevant and need a huge amount of context to be useful in big projects. Just see that this one is using experimental app directory for NextJS because it’s not trained with the latest version. The package version can be 14 or 15 for all it cares it still knows the version it was trained on and will start to hallucinate at some point and this goes for any library or framework. Im not saying we wont get to a point where AIs will be able to replace us Im just not seeing this happening with LLMs in the current state they are. Yes they will get better but for now I’ll stick to my belief that they wont go further than being very good assistants.
@@eduardstefan6833 Unless you include latest documentation in the context which nowadays can be millions of tokens, literally several text books' worth, far beyond what a human brain can hope to hold. Nothing forbids an implementation where a model goes to the Next.js web page, grabs the latest documentation and implements that. In fact even this model was able to use the newer version once prompted to do so. This is just another one of those insurmountable obstacles that are surmounted and forgotten six months later.
Also basically every LLM coding demo I've seen is just web development even though there are so many more applications/fields in software development. In most of these, AI fails miserably. The training set is just the largest with web dev
@@eduardstefan6833 Unless you include latest documentation in the context which nowadays can be millions of tokens, literally several text books'worth, far beyond what a human brain can hope to hold. Nothing forbids an implementation where a model goes to the Next.js web page, grabs the latest documentation and implements it. Even this model was able to use the newer version once prompted to do so. This is another one of those insurmountable obstacles that are surmounted and forgotten six months later.
I think it is still the best to use it as a tool, not for the whole process start to finish. I for example just managed to implement a whole new page and API route that uses Replicate into already existing app, using bolt and v0. Definitely speed up the process since the app was already using tailwind, shadcn, zod etc. Bolt helped me with the API route, I just pasted in the Replicate schema and docs, and then I redefined the UI using v0. Ofcourse it needed some manual work, to get it just right and how I wanted it :D
But why not just use cursor directly in your editor? I don’t see a good reason to use v0 over cursor. Cursor can figure out shadcn components just fine
@@WebDevCody True, I guess in my case beacuse v0 and bolt offer decent free usage, haven't tried Cursor yet since I am currently unemployed 😭😭
@@lev1ato by the way what page you created brother
Thank you for testing this more in depth. I was hyped on the announcement but had my reservations.
You highlighting the obvious difference between web containers and the desktop environments cleared my doubts and I know this won't fit my use-case but can definitely serve the use-case for my mentees who are learning basics and they want to see things working before diving deeper.
My favorite youtuber ❤
All comments saying that coders will never be replaced, you’re actually right stop whining. It’s just a great tool for ppl who are not good at coding to make prototypes fast
“I’m gonna have to figure this out manually, which is such a drag.” 😂
we are getting spoiled so fast lol.
👀
The problem with generating code is usually that when things start going wrong (which they inevitably do), AI will start going in circles and desperately try different solutions and make an even bigger mess of it. Great for generating boilerplate and starter code though, saves a lot of time.
Cody is so nice he says please to the ai
When they take over they’ll rememebr
@@WebDevCody lol
The AI tooling we have right now it's pretty good, if it continue to get cheaper and small when you can run locally or send millions of requests per cents I think the future will be pretty interesting.
Cody, you can acutally just open dev tools -> right click the html tag -> and save as a screenshot. With that you get the entire site, no matter how height, as a screenshot.
Nice thanks for that tip!
I hammered on it quite a bit, and in the end asked for a refund. I felt it needs more time in the oven. Wish there were save stages for when it works well, before your next prompt drives your work off the cliff. Hope you are well amigo.
how did you get the refund? I cant find any section where I can request.
@@shdalex I sent in a bug report as that was the only way. I have heard nothing back.
@@24pfilms i found their email on the Stripe invoice page and I sent a refund request, I'll wait and see if someone replies.
Replit has a feature where you can rollback, but you still end up in circular hell. Webapps are not easily devd out but things like webpages are much easier with AI. But then after you build your simple webpage, who's gonna change the SEO objectives on a monthly basis? Not non-programmers.
I have tried several AI tools now that claim they can “build from scratch “ none of them are there yet. My experience has been a lot like what is demonstrated in this video. But I think we are close. Within 2 years we will have sophisticated agents that can go from idea to MVP very quickly. Which will allow us devs/indie hackers to fail fast hopefully into success
Great intro
Exeptional intro
wow ive ignored all this ai stuff , not sure how long it took you with ai but was pretty cool to see it at least generate an icon
Love the clack of that keyboard.
Best intro yet 😂
00:00 AI generates apps from prompts.
01:09 Initial code attempts often fail.
02:17 Best to start with templates.
03:36 AI can refactor existing code.
05:14 Long prompts increase error likelihood.
06:34 AI versions can be outdated.
07:52 Application startup times are slow.
08:40 Generated UI includes error handling.
10:22 Database interactions often fail.
12:21 Environment compatibility hinders development.
13:59 Image processing fails in web containers.
15:27 Socket errors block app progress.
16:11 Local coding preferred over web containers.
16:55 Diffs needed for code understanding.
Summary by GPT Breeze
This is not a replacement. Wait till it gets better, right now I wouldn't waste your time switching platforms.
Best intro ever🤣🤣
Cody looks like Dorean develops with hair and a pair of glasses.
Thanks Cody yes grab some of the code and paste locally. We still need a deep understanding of what is going on. So we still have jobs. I personally dont think they will get it ever.
The sound of your keyboard is addicting 😅 which model is that?
Klack app cream switch
@WebDevCody No way there is an app that simulates the mechanical keyboard sounds 😅
Bolt is what DevinAI wishes to be. 😆
I never knew that you could directly drag the screenshot after taking it on macos lmao. I always go to the screenshot folder.
Glad I helped level up your Mac OS game
Who doesn't want to jump straight to fixing obscure errors?
@WebDevCody do you have any content on how to manage openai credits in my saas?
they using claude in backend
You should give a try to E2B Fragments IDE
I don't want to sound negative or what! But this is totally not worth the hype. The developer is greedy af to keep increasing the monetary scheme, from 9 usd package to topping up with 10$ now with up to 100 USD/ month. What do they think when releasing an idiot agentic app which can not simply apply new buttons' colors; having issues with getting lost in a hierarchical of similar folders created by itself or having a tendency of instead of fixing stuff, the AI delete a whole bunch of components without thinking. And what's more? There is no going back to the point before it deletes stuff.
I tried it right after vite conf and it broke on first prompt.... started blasting code directly into the chat thread instead of the editor.
It got stuck in a loop trying to fix that first problem, wrote a few more bugs while doing that, and then I was out of the free 150k tokens.
At that rate you'd run out of the 10M tokens it gives you in the monthly free tier in about a week of fulltime work days.
You can enable adding imports in cursor in the settings
I do already have that enabled
If you notice all new ai code generatore like v0 claud and bolt are using shadcn components
Please tell me an easy way to download the project source code. I have to manually copy and paste from each file which is a big hassle. Thank you
Just curious how much money are you spending on AI tool subscriptions (not on api calls for your own saas)
Just $20 / month for cursor currently
Nice vid
I wonder if it's good at SvelteKit as well.
How are they use code editor and run in the fly? I am very curious about this thing how can it be build
"I coded almost fully my AI startup, but I need to fix some global state management and core functionality. Could you help me please?"
I had similar thoughts. It's good for beginners but once you to get to the potatoes, you end up running into some issues. Curious to find out what you think of Replit Agent. Been messing around with that a bit lately
> try new ai tool
> ask it to do something
> get to a point where nothing works
> get angry and spend an hour or two trying to fix it
That is Replit. Amateur night.
@@johndavidson3322 I hear you. At what point do you find issues?
@@philamavikane9423 on the first project, a web app, Replit suggested that we create user accounts with credentials. It couldn't complete that task and wasted literally hours of my time. I wasn't looking at the code at all, I was just letting Replit grind itself into the ground. It couldn't get past the login screen. Then I started over and gave it a very detailed prompt with a different web app project and it did much better...for awhile. Two weeks later the Replit is now permanently hung up. It somehow killed itself where it can't even answer prompts. Tech support was worthless.
But here's a huge pitfall of Replit: it runs into the same token problems as whatever are at the limits of the LLM underpinning it. Ergo, if your project on Replit is complicated/larger than something trivial, you are likely to run into the token limit without Replit telling you that. The end result are hallucinations or Replit repeating the same task over and over and "can't fix" whatever bug you have found. The lack of self-examination by the Replit--it just churns through tokens trying to solve a problem without any realization that it will run out of tokens--creates a shitstorm of problems for obvious reasons.
I first noticed this problem with Claude 3.5 with a web app I was building. As you approach token limit with Claude in a single chat when you are using Artifact, you will start to get warnings and eventually you will be screwed. LLMs like Claude absolutely rely on being able to read the entire chat for every prompt, so the longer the single chat is (by tokens!), the more you risk literally running into a wall and your chat ends in the middle of a coding experience. I don't remember the token limit for a single chat, but it's probably somewhere. And when you are uploading screenshots in LLMs to debug, imagine how many more tokens those use up.
Running Replit or Claude or Bolt or any LLM to try to code out a production level, multi featured application is like starting a race with one tank of gas that you really can't refill.
lol i was waiting for this
A programmer desparately wanting to replace himself, fails time and time again. This is the funniest situation ever.
PLEASE AI REPLACE ME SO I CAN GET A MORE ENJOYABLE JOB
have you used pythagora?
I wonder what's going to happen 5 years from now when there's a bunch of new ways to do things with like updated versions of react, and then there's the old ways of doing things that nobody uses anymore.
Will the AI get confused and give obsolete suggestions? Also, the LLMs will probably have to start to learn using AI generated code, and that will be a mess lol
I think often you need to provide documentation for specific versions of you want accurate output, but honestly if the model has been trained on an older model, it might be hard to have it unlearn
I was never convinced with this replacing cursor. Theyre fundamentally different, why would I want to code in a browser?
Offtopic: lucia auth author just announced deprecation of lucia-auth package.
Are you planning to rewrite your auth using custom API (as he suggests) or switch to using smth like better-auth?
In any case - It would be great if you would share your journey, since I was planning to use lucia auth in prod and now I think I will use guides provided by it's author to roll my own auth with oslo and arctic (basically lucia).
Rolling your own doesn’t seem hard.
Cursor is unusable if you have a workload. It will time out non stop. Company won't tell you they have everyone capped. They say "experiencing issues". It's only for casual use.
has anyone encountered issues with authentication when you sign in? i seem not to be able to sign in
Cody hi. Tbh it's NOT ready for primetime. It can't even generate a next.js 15 app even when told to do so explicitely. It instead just goes off and build a 13.5 one ugh. Keeps on stumbling over and over again in errors it creates on its own. Really good project that integrates well with stackblitz but the main AI is horrendously broken
Have you gotten your hands on Pythagora?
Good job babe! First!!!!
😅😅
If my wife isn’t hyping me up like this, I’m staying single!
2:22 I had way better results with starting a new project
organic engagement
organic engagement
man and you didnt try to deploy and then come back and try to edit that project. idk if it was me but I was not able to do that without starting the project again. I personally don't think is better than cursor yet
> try new ai tool
> ask it to do something
> get to a point where nothing works
> get angry and spend an hour or two trying to fix it
> give up and chastise myself for not just coding it myself
Does it have a composer like cursor
Thank you it’s neat, but I have like no reason to use that in real life
What is the keyboard sound?
tryklack.com/
I tried to build a landing page simple wait list for an app release. It failed so hard I decided to scrap it. Bolt.new is so bad
Does it create a python application
give appwrite another chance too.
I never gave it a chance to begin with
@@WebDevCody ahahahaha
I see you are taking Sam's warning seriously. Or you are coming with certain esoteric language background
who is sam?
@@WebDevCody Altman?
There’s already a new one 😂
skill issue (proopmting specifically)
Yeah it's pretty rubbish unless you want to build a sticky notes app or something, wasted a months credits just trying to fix it's own errors in building a simple audio player app, cancelling my subscription and going back to actual hand coding, job done in 3 hours and it works
What keyboard do you have? It sounds spicy
klack app virtualized keyboard
Don't write `please` to AI :D
you can tell this guy is high iq
I doubt it
still can't compare to cursor, sry
Third😅😅😅😅
every ai tool is expensive shit