Some links: Original cube: codepen.io/wonderwhy-er/full/dyxMJJd ChatGPT refactoring to WebGL chat: chatgpt.com/share/66ffcb1b-1854-800f-be6a-716f1527eeab ChatGPT Canvas produced cube: codepen.io/wonderwhy-er/full/bGXpQMv Claude artifact link: claude.site/artifacts/b5c144a0-2990-4aa6-81e3-6f6e4e69c75c
Hey, thank you for the video, very interesting! I feel we are almost there. Building software requires a lot of planning, and the ability to iterate through a few cycles of review and feedback. With a good agentic system, let’s say an architect agent, product owner, devs, QA; that can go back and forth through many feedback loop and iteration; these systems should be able to build fairly complex software.
GPT4 was not good enough, I tried it before. GPT4 is like an intern. Capable but confused and without wisdom. When you put a bunch of them as a team it still becomes a mess, like if you put a lot of interns on a team. Now o1 model is curious here. It seems it's more capable to be an architect or senior who plans and helps the team to get unstuck. I still suspect that it will be not enough to get to what you are describing, but in 2 years I think we will start getting there. By then we will have infrastructure that will be using gpt5/o1 level models with mild success, GPT6 will arrive and give it a final boost to push it over the edge of working. It's kind of like WebSim was here for a while but using GPT4 and other models just for fun. Then Sonnet 3.5 arrived and when plugged in it suddenly became not a toy but a useful tool. Wild 4-6 years ahead like switching from horses to cars :)
what extension are you using to speak to your computer, looks cool. seems you have a cool extension list. May be make a video on extensions that you use for developers
I use Voice In It's not superb, multiple other solutions are not good too chromewebstore.google.com/detail/voice-in-speech-to-text-d/pjnefijmagpdjfhhkpljicbbpicelgko The best one is in ChatGPT mobile and desktop apps. But that can not be used out of them. I am thinking of making my own extension that uses webgpu accelerated whisper Based on this huggingface.co/spaces/Xenova/realtime-whisper-webgpu
Hey thanks for the comparison - Am I correct i asuming that ChatgPt Canvas currently does not provide the option to preview/articfacts to show the output of the code?
That is a very good advice, I will pin you for some time. I would still love to see a diff between two version, for large file its hard to see where it changed what.
@@cpaps8500 its coming I think you can alrwady start with mini to plan things out and contine in canvas I suspect that gpt5 as mixture of experts will include o1 models. So when planing and investigation is needed uts used, and when applying changes to canvas or custom gpts are needed this canvas model is used
I’ll have to give it a go, I used GPT to learn how to create wireframes and react content and then once I understood that, I started to understand the code and now I’m modifying code snippets with GPT. If this was available then I probably wouldn’t have wasted my time. Thanks for the insights! I can’t wait to explore them
At this stage what should beginners do.Do they continue learning current programming language or wait for some more abstracted programming language to come? In five years time things will get change,so what should we do as of today ?
I think we all should get used to things changing all the time. My suggestion from last year that I think still holds... If something takes you more than 6 months to get results do something else, by the time you get there things will change. And this is not about programming only, rather I find it plausible that we will have AGI-like things in 5 years, by that point intelligence, in general, will be "abundant" But all in all, find what you can do here and now in the next months that is valuable for others. Do that. Already with these tools, one can do useful things quickly. Learn them. I mean I just over last weeks made a game that uses webcam and AI models to track if a kid brushes his teeth. Am thinking of launching it soon.
Basically, if you always wanted to make a game, focus on making it using tools that get you to results the fastest. Don't bother about tools themselves, they will be changing a lot in the next few years. Just look for tools that get you productive the fastest.
@@EduardsRuzga @EduardsDIYLab yes you are right the ai tools will keeping evolving getting better and better but do you have this feeling that From all over the years we have observed that technology has evolved to eventually reduce the human effort but this time it feels like it has started to eliminate the human effort.The only effort human can make is to think and give the command I mean we have not yet reached at that level autonomicity but eventually it will within 5-6 years.
@@GalaxyHomeA9 Yes, that part is weird. The unanswered question is how much depth there is we did not uncover. I mean, the richness of expression today with 3d games and movies and other experiences and what could we do in the Stone Age is already super large. Reaching perfection is impossible, we operate within the boundaries of resources, time, or material ones. And that is one of the things. However powerful AI is it's not magic, it's not god, it can't break laws of physics, it is still constrained by time, space, energy, speed of light. The only future I see where we are not at odds with AI is one where we are augmented by it. So each of us will just get smarter in some way, whether through collaboration with personal versions or direct augmentation like Elon Musk Nurallink. Now imagine a world of humans augmented by such AIs? It still will be a world of competition, just at a higher level. I mean, imagine taking two AIs and telling them to compete to make a better image, they will still compete, just at a different level than we can now.
Great video Eduards! can you also do this with multiple files? i mean many web projects need js, css, html combined. If yes, which platform is better with that?
I would say that online in the browser I tried many, many are hard to use and not complete solutions. replit agent, create.xyz come to mind. Also, gptengineer.app/ But they all have issues and limitations. I am working on one myself too :D There are also offline solutions like Cursor, Aider, Claude-Dev, more developer-oriented and do not help with hosting/sharing the results of your work. Claude and ChatGPT in that sense are also limited but can be used... Just feels incomplete in that sense.
Hm, I gave it a try with this chat chatgpt.com/share/66ffce05-e41c-800f-a84a-72815d1febfb And ChatGPT can work and switch between multiple Canvas files, in that chat I was working back and forth on article and code
@@EduardsRuzga Nice! I wonder if it can merge two canvases? As in: Write a program that does x in one canvas and a snippet that does function y in another canvas, then merge y into x in the appropriate place and amend x to do the correct calls etc... might have a play around with that
Great video that clearly shows that AI has a long way to go to replace programmers. PhD level Intelligence is the biggest laugh the AI produced recently. Don't get me wrong. I love the advances AI made. And if used by skilled people it can be very supportive. But the cost benefit ratio is still very bad. Current AI is good for pseudo spectacular demos... but light years away from replacing relevant tasks. It is very good for information retrieval and also the voice assistents are pretty stunning these days. But pretty far away from leaving them alone with my bank credentials or health parameters!!!
Well, so far I have seen people who are not developers building small apps and automatisations. The moment you need larger, multiple features, complex apps, yes they fall short. However, the speed of change and the amount of untapped potential are also huge. I mean, we are in September of 2024. ChatGPT was released in November 2022. It failed at coding even dozens of lines that work. Then in March of 2023, they released GPT4. and then Code Interpreter. I think that was a moment of large shift. It felt like we switched from assembly language or punch cards to higher-level language for small things. The question remains of whether it's possible to build larger things with it. I remember playing with AutoGen a year ago building agentic flows and it was BAD. It felt like putting a bunch of middle schoolers to create a software project. So the issue is twofold. Foundation models are not at the agentic level yet, the correct way to integrate them with systems to allow making large complex software is also not really nailed yet. But it is progressing all the time. It feels like AI abilities are growing 4x in depth and breath every 6 months as far as coding goes. From 5 lines with GPT3.5 in November of 2022, to A4 list of code in March of 2023 to 4 pages of working code this summer with Sonent 3.5 and o1 models. Next year it will be what? 32 pages? 64 pages of working code in 2026? The codebases I work at are thousands of pages long with complex interconnections that humans fail to grasp. But I bet that there will be larger unlocks. I do think that in the next 4 years, we will get to systems that will be able to make very complex software. Both because models will get better and ways to integrate them with code execution and testing will get better.
Just a practical example. A decade ago I was a Flash developer, landing pages and social games. I was a DeviantArt member back then, largely publishing creative Flahs toys like music visualization toys. Flash died, the code sits on my hard drive, too much effort to rewrite to JavaScript. When WebSim and Sonnet 3.5 came out I gave it my old code and it rewrote it to working JavaScript, that is not trivial, there is no direct one-to-one mapping, especially on a level of Flash APIs. And yet it did that... Without my assistance... If that is not a wakeup call then I do not know what is. We may end up in a situation where the tech dept is a thing of the past, you write tests and criteria and make the model run for weeks making a perfect solution, optimal, well documented, and passing all tests. We are not there yet but we are inching closer.
@@EduardsRuzga I wish instead of going more and more broad (ChatGPT claims to support ~100 human languages and 20-30 programming languages) they would rather focus. Just one human language is enough for me. And some few programming languages. But no matter how good the systems get. In the end for serious tasks a human must inspect the generated code. Because to this day no generated code can be regarded as error free and also assure that there are no back doors in the generated code. And the morality of AI is the same as that of any other business. GREED is the driver. Already the oh so moralistic developers do not bother to let their bots block public sites by producing too much traffic in their greed to suck all info they can get. And also bots can never estimate good versus bad annotation. And just a source is not a really good evaluation esp. in these days where noone can say whether content has been generated or not.
@@Heisenberg2097 that code inspection part. I don't think that is the case. I mean even in software development test driven development is the king. What if there are performance, memory, and functional and unit tests that are compact and understood by humans. Do you care about the code then? It's kinda like with 3rd party libraries. Do you care to inspect the code as long as API works? I am interested in test-driven development with AI in that sense and working towards such UX for people who use it. I have a video on my server commander GitHub project. In it, I gave it HTML page, asked it to extract and make it runnable in the node part of the javascript, made it inspect outputs of running the code, brainstorm on how to improve, then rewrote, ran, and inspected the results loop until it made new code. Then integrated it back into HTML page. This is where I want to get with what I am working on but in a cloud-like WebSim, Replit, gptengineer, or new bolt.new So you do not need to install and setup things locally.
@@EduardsRuzga I want the AI to write the tests itself. Just the other day I let Perplexity do an FMEA for an MRI-Thermomentry software for monitoring regional deep hyperthermia treatments and the results were amazing. A single prompt yielded a thorough FMEA. But it's not black magic. In the end it is 'ONLY' good data mapping.! There are many so reasonable uses for AI. Esp. diagnostics and reporting will increase in stability and accuracy as machines have no daily form. But we are far away from the One-Click-Stock-Market millionaire app. Also look at the reliability of weather forecasting or pandemia predicitions and solutions... at first AI eperts have to thoroughly explain why their tech does what it does... and also from that we are far away.
I think Claude was ahead on quality of code - the small cubes rotated whereas they did not with ChatGPT. Can you instruct ChatGPT when to do a pinpoint edit? Also, can it insert comments into the code itself?
I would agree that Claude seemed to reach better results faster here. But I did the same experiment yesterday evening and here ChatGPT won over. I do think they are very close now. About asking for pinpoint edits and not rewrites, I need to test more. And yes, it can insert comments.
Calde? You mean Claude right. As far as knowladge of libraries, which ones Claude knows that ChatGPT does not? I showed before that ChatGPT can use internet search for libraries it never seen.
2 місяці тому+1
most missing feature is running the code in the browser
I agree, am actually working towards having a platform that allows ChatGPT to run code in the browser. But there is more than that and it's not the main part of it.
doesn't seem like a difference big enough to buy gpt premium, i'm doing fine with small unlimited models and in ide copilots and claude sonnet free daily queries for now. The stuff i'm doing is simple, maybe i'll change my mind when i need something more compex. Breaking down problems and asking for each step makes it easier for them to do it and lets you have control on what you are doing. I'd like to hear if others with more experience think i'm wrong.
@@WaveOfDestiny yeah, could be. I am heavy user, use it for writing, brainstorming, summarising, making image ideas. But there are so many free tools out there that if you are occasional user you can do with using free tiers of a dozen of offerings.
This was first impressions video and here it did quite well. I am testing it more and there are issues. It does fail with complex files. I also tried one workflow that did not work for me anywhere else. Give it a UA-cam transcript in a canvas, ask to think of good places to split it to chapters in chat, then add them in appropriate places of transcript as comments. If you ask to do the mall at once it fails, but one by one it succeeds and Claude or Gemini can't do it well... And one can imagine similar tasks for code. You can ask it to make comments, and then selectively apply them. That is different from Claude that does everything in one shot and also fails spectacularly depending on complexity of the task. I really hate how Claude always rewrites and as task becomes more complex it rewrites with more and more mistakes... Will see over next weeks as I find more of productive workflows with it Anyways, that is my experience, yours is different, care to share more? What tools do you use?
@@tomascoox yeah, tried it once so far. I would compare it to gptengineer, replit agent, getlazy though. It is solely focused on coding and has deploy options. First impression of bolt was good, better then alternatives. What is yours?
Some links:
Original cube:
codepen.io/wonderwhy-er/full/dyxMJJd
ChatGPT refactoring to WebGL chat:
chatgpt.com/share/66ffcb1b-1854-800f-be6a-716f1527eeab
ChatGPT Canvas produced cube:
codepen.io/wonderwhy-er/full/bGXpQMv
Claude artifact link:
claude.site/artifacts/b5c144a0-2990-4aa6-81e3-6f6e4e69c75c
Hey, thank you for the video, very interesting! I feel we are almost there. Building software requires a lot of planning, and the ability to iterate through a few cycles of review and feedback. With a good agentic system, let’s say an architect agent, product owner, devs, QA; that can go back and forth through many feedback loop and iteration; these systems should be able to build fairly complex software.
GPT4 was not good enough, I tried it before. GPT4 is like an intern. Capable but confused and without wisdom. When you put a bunch of them as a team it still becomes a mess, like if you put a lot of interns on a team.
Now o1 model is curious here. It seems it's more capable to be an architect or senior who plans and helps the team to get unstuck. I still suspect that it will be not enough to get to what you are describing, but in 2 years I think we will start getting there. By then we will have infrastructure that will be using gpt5/o1 level models with mild success, GPT6 will arrive and give it a final boost to push it over the edge of working.
It's kind of like WebSim was here for a while but using GPT4 and other models just for fun. Then Sonnet 3.5 arrived and when plugged in it suddenly became not a toy but a useful tool.
Wild 4-6 years ahead like switching from horses to cars :)
what extension are you using to speak to your computer, looks cool.
seems you have a cool extension list. May be make a video on extensions that you use for developers
I use Voice In
It's not superb, multiple other solutions are not good too
chromewebstore.google.com/detail/voice-in-speech-to-text-d/pjnefijmagpdjfhhkpljicbbpicelgko
The best one is in ChatGPT mobile and desktop apps.
But that can not be used out of them.
I am thinking of making my own extension that uses webgpu accelerated whisper
Based on this huggingface.co/spaces/Xenova/realtime-whisper-webgpu
That voice command you did, was it a plugin?
@@NDIDIAHIAKWO yeah, chrome extension called voice in.
@@EduardsRuzga Ok! Thanks alot.
Hey thanks for the comparison - Am I correct i asuming that ChatgPt Canvas currently does not provide the option to preview/articfacts to show the output of the code?
Yes, you are correct, and I comment on video that this is major Claude Artifacts advantage.
Use the back and forward buttons to see the edits.
That is a very good advice, I will pin you for some time.
I would still love to see a diff between two version, for large file its hard to see where it changed what.
I love these comparison videos! I wish it was o1-mini with canvas because I find o1-mini the best coder
@@cpaps8500 its coming
I think you can alrwady start with mini to plan things out and contine in canvas
I suspect that gpt5 as mixture of experts will include o1 models.
So when planing and investigation is needed uts used, and when applying changes to canvas or custom gpts are needed this canvas model is used
I’ll have to give it a go, I used GPT to learn how to create wireframes and react content and then once I understood that, I started to understand the code and now I’m modifying code snippets with GPT.
If this was available then I probably wouldn’t have wasted my time. Thanks for the insights! I can’t wait to explore them
At this stage what should beginners do.Do they continue learning current programming language or wait for some more abstracted programming language to come? In five years time things will get change,so what should we do as of today ?
I think we all should get used to things changing all the time.
My suggestion from last year that I think still holds...
If something takes you more than 6 months to get results do something else, by the time you get there things will change.
And this is not about programming only, rather I find it plausible that we will have AGI-like things in 5 years, by that point intelligence, in general, will be "abundant"
But all in all, find what you can do here and now in the next months that is valuable for others. Do that.
Already with these tools, one can do useful things quickly. Learn them.
I mean I just over last weeks made a game that uses webcam and AI models to track if a kid brushes his teeth. Am thinking of launching it soon.
Basically, if you always wanted to make a game, focus on making it using tools that get you to results the fastest. Don't bother about tools themselves, they will be changing a lot in the next few years. Just look for tools that get you productive the fastest.
@@EduardsRuzga @EduardsDIYLab yes you are right the ai tools will keeping evolving getting better and better but do you have this feeling that From all over the years we have observed that technology has evolved to eventually reduce the human effort but this time it feels like it has started to eliminate the human effort.The only effort human can make is to think and give the command I mean we have not yet reached at that level autonomicity but eventually it will within 5-6 years.
@@GalaxyHomeA9 Yes, that part is weird. The unanswered question is how much depth there is we did not uncover. I mean, the richness of expression today with 3d games and movies and other experiences and what could we do in the Stone Age is already super large. Reaching perfection is impossible, we operate within the boundaries of resources, time, or material ones.
And that is one of the things. However powerful AI is it's not magic, it's not god, it can't break laws of physics, it is still constrained by time, space, energy, speed of light.
The only future I see where we are not at odds with AI is one where we are augmented by it. So each of us will just get smarter in some way, whether through collaboration with personal versions or direct augmentation like Elon Musk Nurallink. Now imagine a world of humans augmented by such AIs? It still will be a world of competition, just at a higher level.
I mean, imagine taking two AIs and telling them to compete to make a better image, they will still compete, just at a different level than we can now.
Great video Eduards! can you also do this with multiple files? i mean many web projects need js, css, html combined. If yes, which platform is better with that?
I would say that online in the browser I tried many, many are hard to use and not complete solutions. replit agent, create.xyz come to mind. Also, gptengineer.app/
But they all have issues and limitations.
I am working on one myself too :D
There are also offline solutions like Cursor, Aider, Claude-Dev, more developer-oriented and do not help with hosting/sharing the results of your work.
Claude and ChatGPT in that sense are also limited but can be used... Just feels incomplete in that sense.
Hm, I gave it a try with this chat
chatgpt.com/share/66ffce05-e41c-800f-a84a-72815d1febfb
And ChatGPT can work and switch between multiple Canvas files, in that chat I was working back and forth on article and code
@@EduardsRuzga Nice! I wonder if it can merge two canvases? As in: Write a program that does x in one canvas and a snippet that does function y in another canvas, then merge y into x in the appropriate place and amend x to do the correct calls etc... might have a play around with that
That way you could iterate on functionality
Great video that clearly shows that AI has a long way to go to replace programmers. PhD level Intelligence is the biggest laugh the AI produced recently. Don't get me wrong. I love the advances AI made. And if used by skilled people it can be very supportive. But the cost benefit ratio is still very bad. Current AI is good for pseudo spectacular demos... but light years away from replacing relevant tasks. It is very good for information retrieval and also the voice assistents are pretty stunning these days. But pretty far away from leaving them alone with my bank credentials or health parameters!!!
Well, so far I have seen people who are not developers building small apps and automatisations. The moment you need larger, multiple features, complex apps, yes they fall short.
However, the speed of change and the amount of untapped potential are also huge.
I mean, we are in September of 2024. ChatGPT was released in November 2022. It failed at coding even dozens of lines that work. Then in March of 2023, they released GPT4. and then Code Interpreter. I think that was a moment of large shift. It felt like we switched from assembly language or punch cards to higher-level language for small things.
The question remains of whether it's possible to build larger things with it. I remember playing with AutoGen a year ago building agentic flows and it was BAD. It felt like putting a bunch of middle schoolers to create a software project.
So the issue is twofold. Foundation models are not at the agentic level yet, the correct way to integrate them with systems to allow making large complex software is also not really nailed yet. But it is progressing all the time. It feels like AI abilities are growing 4x in depth and breath every 6 months as far as coding goes.
From 5 lines with GPT3.5 in November of 2022, to A4 list of code in March of 2023 to 4 pages of working code this summer with Sonent 3.5 and o1 models.
Next year it will be what? 32 pages? 64 pages of working code in 2026?
The codebases I work at are thousands of pages long with complex interconnections that humans fail to grasp.
But I bet that there will be larger unlocks. I do think that in the next 4 years, we will get to systems that will be able to make very complex software. Both because models will get better and ways to integrate them with code execution and testing will get better.
Just a practical example. A decade ago I was a Flash developer, landing pages and social games.
I was a DeviantArt member back then, largely publishing creative Flahs toys like music visualization toys.
Flash died, the code sits on my hard drive, too much effort to rewrite to JavaScript.
When WebSim and Sonnet 3.5 came out I gave it my old code and it rewrote it to working JavaScript, that is not trivial, there is no direct one-to-one mapping, especially on a level of Flash APIs. And yet it did that... Without my assistance...
If that is not a wakeup call then I do not know what is.
We may end up in a situation where the tech dept is a thing of the past, you write tests and criteria and make the model run for weeks making a perfect solution, optimal, well documented, and passing all tests. We are not there yet but we are inching closer.
@@EduardsRuzga I wish instead of going more and more broad (ChatGPT claims to support ~100 human languages and 20-30 programming languages) they would rather focus. Just one human language is enough for me. And some few programming languages. But no matter how good the systems get. In the end for serious tasks a human must inspect the generated code. Because to this day no generated code can be regarded as error free and also assure that there are no back doors in the generated code. And the morality of AI is the same as that of any other business. GREED is the driver.
Already the oh so moralistic developers do not bother to let their bots block public sites by producing too much traffic in their greed to suck all info they can get. And also bots can never estimate good versus bad annotation. And just a source is not a really good evaluation esp. in these days where noone can say whether content has been generated or not.
@@Heisenberg2097 that code inspection part. I don't think that is the case. I mean even in software development test driven development is the king.
What if there are performance, memory, and functional and unit tests that are compact and understood by humans. Do you care about the code then?
It's kinda like with 3rd party libraries. Do you care to inspect the code as long as API works?
I am interested in test-driven development with AI in that sense and working towards such UX for people who use it.
I have a video on my server commander GitHub project.
In it, I gave it HTML page, asked it to extract and make it runnable in the node part of the javascript, made it inspect outputs of running the code, brainstorm on how to improve, then rewrote, ran, and inspected the results loop until it made new code. Then integrated it back into HTML page.
This is where I want to get with what I am working on but in a cloud-like WebSim, Replit, gptengineer, or new bolt.new
So you do not need to install and setup things locally.
@@EduardsRuzga I want the AI to write the tests itself. Just the other day I let Perplexity do an FMEA for an MRI-Thermomentry software for monitoring regional deep hyperthermia treatments and the results were amazing. A single prompt yielded a thorough FMEA. But it's not black magic. In the end it is 'ONLY' good data mapping.! There are many so reasonable uses for AI. Esp. diagnostics and reporting will increase in stability and accuracy as machines have no daily form. But we are far away from the One-Click-Stock-Market millionaire app. Also look at the reliability of weather forecasting or pandemia predicitions and solutions... at first AI eperts have to thoroughly explain why their tech does what it does... and also from that we are far away.
I think Claude was ahead on quality of code - the small cubes rotated whereas they did not with ChatGPT. Can you instruct ChatGPT when to do a pinpoint edit? Also, can it insert comments into the code itself?
I would agree that Claude seemed to reach better results faster here. But I did the same experiment yesterday evening and here ChatGPT won over. I do think they are very close now.
About asking for pinpoint edits and not rewrites, I need to test more.
And yes, it can insert comments.
I just had situation where it decided to do a full rewrite, I stopped it and asked for pinpoint edit and it did!
Canvas has 'Logs' to track changes.
@@deep2mixer you mean back forward version control or what logs?
@@EduardsRuzga Yes...keeping track of revisions or copies of the work. That's what I meant. Thanks for your videos. Keep up the good work.
@@deep2mixer trying, thanks man :)
Any topics you are interested in?
Calde hands down. Mostly because of the reasoning capabilities and knowledge of updates libraries.
Calde? You mean Claude right.
As far as knowladge of libraries, which ones Claude knows that ChatGPT does not?
I showed before that ChatGPT can use internet search for libraries it never seen.
most missing feature is running the code in the browser
I agree, am actually working towards having a platform that allows ChatGPT to run code in the browser. But there is more than that and it's not the main part of it.
doesn't seem like a difference big enough to buy gpt premium, i'm doing fine with small unlimited models and in ide copilots and claude sonnet free daily queries for now. The stuff i'm doing is simple, maybe i'll change my mind when i need something more compex. Breaking down problems and asking for each step makes it easier for them to do it and lets you have control on what you are doing. I'd like to hear if others with more experience think i'm wrong.
@@WaveOfDestiny yeah, could be. I am heavy user, use it for writing, brainstorming, summarising, making image ideas. But there are so many free tools out there that if you are occasional user you can do with using free tiers of a dozen of offerings.
It doesn’t use the newest 4o 1 models, so it’s a pass for now
Chat is very very bad. It has you running around in circles with useless code
This was first impressions video and here it did quite well. I am testing it more and there are issues. It does fail with complex files.
I also tried one workflow that did not work for me anywhere else.
Give it a UA-cam transcript in a canvas, ask to think of good places to split it to chapters in chat, then add them in appropriate places of transcript as comments.
If you ask to do the mall at once it fails, but one by one it succeeds and Claude or Gemini can't do it well...
And one can imagine similar tasks for code. You can ask it to make comments, and then selectively apply them.
That is different from Claude that does everything in one shot and also fails spectacularly depending on complexity of the task.
I really hate how Claude always rewrites and as task becomes more complex it rewrites with more and more mistakes...
Will see over next weeks as I find more of productive workflows with it
Anyways, that is my experience, yours is different, care to share more? What tools do you use?
bolt.new does
@@tomascoox yeah, tried it once so far.
I would compare it to gptengineer, replit agent, getlazy though. It is solely focused on coding and has deploy options.
First impression of bolt was good, better then alternatives.
What is yours?