Thin wrappers around LLMs will not survive. I like Cursor, but at the same time I don't like that they build on top of Open Source montize it and not giving something back.
Yea there's no money in it right now cause it's super easy to compete with a flimsy product. Maybe it's just a timeline thing, but real products, if they ever happen, will not be chat interfaces with obvious AI prompting.
@@Varadiio Real AI products that can't be rapidly replicated as you say will have to be complex agent based systems. Anybody building a 'chat' based AI product is in for a rough time.
yup they basically built a plugin/fork. You'll never beat the big players like that. You need to build something actually new, like IntelliJ which beat Eclipse.
This is why competition is healthy, these guys can fight it out trying to one up each other and in the end we get a great editor experience either way. No doubt MS would be “slower” to push this stuff if cursor hadn’t raised the bar etc
@@MrR3set at that point I'm faster writing it myself. I have to spend time explaining myself to a bot that might not even process what I've said correctly and will produce utter rubbish. Why spend the time trying to exlain what you want to do to a bot, when you could just do it yourself.
@@melkorbane Agreed, the human aspect is overlooked and only working on metrics of accuracy but not measuring human evaluation if its giving good results. Its still in beta and its microsoft tool. So we know how it always goes.
"The point about test is they assert the HUMAN'S expectations." Very well put. I would extend the same to TYPES. You should be the one defining the interfaces and therefore the RELATIONSHIPS between modules and functions. THEN start tabbing some auto complete with firm guard rails in place.
@@korbpw too risky, most of the times it's better to do things yourself, otherwise you'll be stuck asking copilot to fix the already messed up generated code which will only make it worst lmfao
@@yassir-eh using it wrong, you can be imperative with the prompts, so it will produce the code you want, if something's wrong, it's your fault, the level of freedom you give it is controlled by you, disconsidering the hallucinations.
@@korbpw 1. Using it wrong? I'm sorry but there is no definitive right way to prompt an AI, yes there are things that can make the prompt better, be there is no certainity that a prompt will give you a 100% correct answer. But fixing something using the correct logic yourself has better certainity, if both take the same amount of time becuase you are stuck trying to figure out what to ask the AI then why not just fix it yourself and Use AI for things that are just mundane and tedious. 2. Accuracy, right now a really good programmer can match a LLM model's accuracy in fixing bugs if the task is hard to explain to the AI. So its better to just do it yourself
Manual code reviews as a team are highly valuable. We used to do this when I was a junior developer and they were fantastic. We would get together as our small team and review the changes made throughout the week. We literally reviewed every change. Now, our commits were not always the smallest (before the days of DevOps and microservices, though we were mostly doing DevOps and single it web development... it was already microservices... because web pages were their own code). It could still be done today, but perhaps picking a subset of changes to review would work, as well. Being able to discuss the changes, the patterns, etc. that were implemented made for highly quality code in the future.
Sublime text has the remarkable new feature of staying the same forever. It’s like a 90s pickup truck. It does what you want it to do and not much else
@@ichigo_huskyWell if you have time, you can configure anything you want, and it's easy thanks to the neovim ecosystem, you literally create your own editing environment, for free.
4:40 vitest allows for in source testing, but like you mentioned shortly before the timestamp… Tests sometimes have way more code than the functions they test. The size of files can really blow up. But you can have vite prune it in production and avoid exporting everything. I often find I only want to test exported code anyway, though. Anything that isn’t exported will be used in the exported function, and their behaviour is tested in a more integrated way which I prefer all around. I’m not a 100% coverage kind of person.
This video perfectly summarized what I've been dealing with writing C# using Copilot or Codeium, partly the plugins don't work, and partly their answers are just complete bullshit or don't write in the style they should write in - and then you post your code into ChatGPT instead and that works way better for some reason
1000% agree on your comment about testing. Thank you! 😍This is why human testers / QA Engineers / SDETs are still going to be needed for quite some time.
It is becoming increasingly hard to pick the right AI assistant. There are already dozens of them and you cannot fully try them all to decide which one suits you the best.
Cursor has some awesome UX, and that makes difference for me. Copilot team even though very active they seem to be focusing more on expanding features rather than polishing base experience.
Nah. Claude is the secret sauce of cursor and a lot of impressive seeming ai products that are just Claude wrappers. It’s significantly better than chat gpt and therefore anything Msft has. Chat gpt has a serious model drift problem
Better to control than to own (tax reasons). Why all the super wealthy put all their money in Trusts and Foundations, such as the Bill and Melinda Gates foundation. Has nothing to do with charity… same with Warren Buffet 😅
Also restricting it to not use certain things or saying "use the packages I already have in the project". But you can be more specific bout how you expect it to make things type safe. Give it examples of good and bad, etc. Also ask it to explain the answers so you can tell it what to not do. Cursor also has the "include all files as context" which is interesting (CTRL+Enter when you send a chat in Cursor). The inline spinners imply that it's doing some kind of chain of thought pre-processing of all the documents before it reads your instructions.
About AI written unit tests : Totally agree their point is to ensure the code does what a human would expect. But I also think there is a use case for ai in it. The same way a dev should understand the code written by an AI, if an AI writes tests well enough for an human to understand and review, the human can assert that the test is testing properly. What makes unit test writing such a pain is the repetitivenes of the syntax that is needed to make them feel comprehensible. Using AI would be a huge plus to make it easier. One may fear that in the long run devs might forget the proper use of unit tests entirely, but if we can get AI to be explicit enough it could also make a positive feedback loop to better understand unit tests as we write them.
Місяць тому+13
Telling an open source project to use a closed source project to review their PR's is kinda off to me..
Are you allowed to work on open source projects on windows or does that being closed source compromise everything?
Місяць тому
@@MerkieAE Windows isn't a SAS system where all the data is stored in the cloud. Further Windows could be used by a developer but wouldn't be a dependency directly, you can just use something else without loosing any data.
Honestly, I’m okay with AI/LLMs sucking at writing or rewriting chunks of code. I can do those things myself and most of the time those kind of things aren’t really a time sink. Where I want AI/LLM tools to be good at is doing the mundane stuff: autocompleting things I was already planning to write (like comments or rewriting a bunch of same-y code), filling in mock data and checking if I could write something more efficiently. And that’s already been quite good and reliable with most of these tools (if you keep their pitfalls in mind) in my opinion. When it tries to do more advanced stuff and I have no idea whether or not the answer is reliable or efficient, I might as well not use it. It takes as much time to google if my LLM tool screwed up as googling how to solve it on my own.
This seems like a normal coding session, when working with AI. It’s equal parts masochistic and alluring to spend an hour to trying to teach the AI to write a 3 line method. 50 shades of AI
I'm more lenient toward the new Copilot features since they are still in the preview stage. Once it's out of preview, I will try it and then decide whether to switch from Cursor. The half-price for Copilot is not a minor benefit for me.
I feel like the 'generate tests' will be really useful for meeting BigOrg code coverage expectations, but as you hint, not actually helpful in testing user expectations.
8:40 "And it's the exact same code" That has been my experience when trying to make it do something I 'really' needed to get done quickly and didn't know what to do. AI is, very seriously, dumb as hell. And the people who think it's just about to take all our jerbs are on a similar level of regurgitating information.
See my comment too.. Things like copilot has gotten SO much worse too. Now its not even usable as a oh rewrite this to take a object instead of a array of object, or oh i missed checking errors after these calls, add it
As a well experienced dev I don't need a cursor. I write a few lines of code that do many great things. Just learn clean code. Then you just need to learn new things like firebase or any API service integration while working with tasks. ❤
AI coding tech will be as commoditized as calculators in the next 5 years. Every product remotely related to apps will have it as a feature. Every product that could benefit from such "AI extensions" but does not risk becoming not competitive. It's kind of like mouse input. Imagine having a Windows or Mac app that can only be used with the keyboard? Last year one term under which that was coined was "Polimorphic apps" which rewrite themselves based on user input. I do belive that the end destination of this kind of thing is integration into OS, OS provides GUI, and it will provide AI friendly app layer on the operating system level removing the need for the majority of small apps and possibly allowing customization of large apps. That is what I suspect will be slowly happening, will take a while, a lot of "tech dept" to untangle both on software and human habits level.
I got told to use cursor… but from my view point I just wait until there more stable tool come out many years before I start using it. But I don’t know why people get mad t me. I just want stable not new feature
My thoughts exactly! I see everyone talking about cursor and it was tempting to use it, but honestly Im not gonna waste time to learn it if in a couple of months vscode(or some of the other main ones) come with and incredible update that'd make cursor obsolete... which Im pretty sure it's gonna happen to cursor next year.
I've been using cursor for a little bit. And while I like a lot of it. Sometimes I find that I spend more time fighting with it's autocomplete than working with it. When it gets it wrong, it so much harder to ignore it's suggestions. And sometimes if feels like it's making changes without me confirming them.
The insane feature of cursor is "Cursor lets you breeze through changes by predicting your next edit.". All other feature is not important, I will not go back to Copilot until they have the feature
Does it generate code on the fly based on your previous code like cursor? Writting and endpoint the way your team does it and have it auto complete or generate the rest of the endpoints for you is amazing. (Based on your models or services as well). Having the AI adapt to your code
As of now, I'm often disabling the AI helpers. By default, they keep distracting me so much that I forget what I wanted to do, by behaving like a chatty person looking over my shoulder and talking bullshit. It's mostly noise obfuscating useful code completions. Mostly because it's missing context I guess. When I'm about to write a comment, I do this because there's some non-obvious reason. I don't want an AI to guess obvious things and talk over me, but that's what happens right now. But I use AI from time to time to help writing tests. That's so much boilerplate code to write, and I keep forgetting the nuances required for writing tests. However, this also only works well if there's not too much context needed.
ai assistant driven development is impressive but still not hacked. you switch from a designer into an assembler. I wouldn’t recommend new people just dive into it as you do need some experience to stitch things together in a way that is functional and maintainable
The no unchecked index typescript error thing is bullshit, the array length was checked to be 1 so reading 0 should not return undefined. I can totally understand the copilot messing it since I would've too.
Same. I was looking at that and I couldn't find the issue. The array is known to have exactly one item and the type prevents it from being null/undefined, so what's the problem?
@@besizzo My other comment got blocked because I linked to their website but yeah, they raised $12M for it and the wording seems to point to a new editor.
@@solivagrey1707 My comments went to the mod queue because I put an URL to their website, but they raised $12M "to build a text editor" so it looks like it's a new one.
lol I respect that "I'm gonna leave this here so that it fucks me the next time I open VS Code", a man who values comedy over productivity even at his own expense
Hmm... New Ads... Maybe add a more obvious transition ("And now, a word from our sponsors") - the flow needs to be smoothed out a little somehow. Otherwise, cool idea... May have been good to go slightly deeper on the Review App suggestion, and how it works.
Would be great if they could use the whole folder as context, not just selected files. That way you could make updates to the whole codebase. But what do i know. Im not even a dev.
I think using AI to write unit tests is fine as long as you are actively reviewing the tests that it spews out in considerable detail, otherwise there's no point writing tests. Unfortunately I suspect the reality will be that too many devs will implicitly trust the AI generated tests without any consideration.
Do you have a .cursorrules ? You should try to finetune your cursor to your project. You can add examples and instructions focused on your style, or have always in mind the TYPESAFETY of everything
I believe the inline chat/commants and chat are separate settings, and the instructions you gave it were set in the Chat setting, not in the inline command setting.
It's funny how programmers see the writing on the wall yet keep looking forward to better AI assistants but Figma doesn't have the spine to release Make Designs. Really hoping there's a Cursor-like Figma killer in the works. v0 doesn't really solve the rapid ideation and prototyping plain designs do
With Cursor I can chose which AI API is used. In this video here you use co-pilot which uses the despicable OpenAI API. Tons of bad feeling from me for everything-OpenAi. vscode is Microsoft-owned, and they of course want to push co-pilot hard since they've invested tons in OpenAI.
Zed with Supermaven and Claude AI is just AWESOME My laptop have 16gb ram and as a react-native dev, I need atleast a browser with some tabs and an emulator always opened Vscode is good but it's SUCH A RAM HOG I wish a better font and easier prettier integration in zed but still better overall than vscode
I auto like your videos before watching btw 🙂 I highly recommend you check someone really good with Cursor to see how you could use it to it's full potential. You're using the "Composer" feature a lot, where it's not the best to use. Also, you biased the Cursor solution with the VScode solution. That's why it was bad, i tried the same thing and it was way better. In your last video about Cursor, I saw that you'were not using the context and the CMD+L at all, which is the most powerful at current moment.
![Cursor vs Bolt vs Replit vs v0 [Best AI Coding Tool]](http://i.ytimg.com/vi/_9OE1KA1Wfs/mqdefault.jpg)
![Cursor vs Bolt vs Replit vs v0 [Best AI Coding Tool]](/img/tr.png)
I'm not using it unless it can generate a video of will smith eating spaghetti.
you demand too much sir
Spaghetti code!
in pig latin
Better luck getting it to generate of video of Will eating p-diddy's aZ#, which is just as likely.
For free, of course 😉😉
Thin wrappers around LLMs will not survive. I like Cursor, but at the same time I don't like that they build on top of Open Source montize it and not giving something back.
Yea there's no money in it right now cause it's super easy to compete with a flimsy product. Maybe it's just a timeline thing, but real products, if they ever happen, will not be chat interfaces with obvious AI prompting.
@@Varadiio Real AI products that can't be rapidly replicated as you say will have to be complex agent based systems. Anybody building a 'chat' based AI product is in for a rough time.
I think AI prompts for generated tests actually does make sense, as he mentioned, tests need to check human developer expectations.
yup they basically built a plugin/fork. You'll never beat the big players like that. You need to build something actually new, like IntelliJ which beat Eclipse.
i thought this was a minor issue, turns out many seem to have this concern with Cursor
This is why competition is healthy, these guys can fight it out trying to one up each other and in the end we get a great editor experience either way. No doubt MS would be “slower” to push this stuff if cursor hadn’t raised the bar etc
exactly!
Are we seriously watching Theo discover that tests are longer than the code?!
What is that?! A circus?!
I think he just doesn't like tests (I don't either, I'm lazy)
I love how vscode was gatekeeping first but now competition camr and suddenly things can be done
AI can only write tests if it knows what a function is supposed to do, not based on what it is currently doing, which may be wrong.
True, but it can also be useful when you are about the extend it's functionality and want to make sure the existing functionality works
I find best results when passing code and giving natural language context.
If you can show example as well is great
@@MrR3set at that point I'm faster writing it myself. I have to spend time explaining myself to a bot that might not even process what I've said correctly and will produce utter rubbish. Why spend the time trying to exlain what you want to do to a bot, when you could just do it yourself.
@@omri9325 True. Actually, that's the main reason i use test sets: to check if I didn't break anything.
Skill Issue. All you have to do is include "make sure you understand all the functions correctly" to your prompt.
CoPilot is like asking a intelligent leetcoder to write production ready application code who only wrote algos but not web apps.
This is because those guys are exactly who trained the models lol.
@@melkorbane Agreed, the human aspect is overlooked and only working on metrics of accuracy but not measuring human evaluation if its giving good results.
Its still in beta and its microsoft tool. So we know how it always goes.
Imagine when the Calculator became mainstream… 😅
The magic about cursor is the combination of Claude + Deep Integration within the code editor.
Yup you can tell cursor is doing chain of thought under the hood.
“what the fuck did that do” made me absolute crease, just caught me off guard and I belly laughed.
Read your comment as he said it 😂 perfect timing
I love that Cursor’s solution to Jest not working is switching to Vitest, that’s based AF
most human LLM response ive ever seen
@@LukePighetti oh hi Luke! also yes, whatever they’re doing with their prompts is working
"The point about test is they assert the HUMAN'S expectations."
Very well put. I would extend the same to TYPES. You should be the one defining the interfaces and therefore the RELATIONSHIPS between modules and functions.
THEN start tabbing some auto complete with firm guard rails in place.
Isn't LLMs copying human behaviour including flaws, biases?
The race who is going to milk the developers the most is heating up.
Cursor gives tremendous value what are you talkinga bout
Time and time again, I get this feeling that I worked and studied to become a Software Developer, not an AI Wrangler.
which is exactly why I stopped using this AI nonsense. It was cool in the beginning
Either adapt to use it, or get left behind? If you know how to use it, can make you very fast.
@@korbpw too risky, most of the times it's better to do things yourself, otherwise you'll be stuck asking copilot to fix the already messed up generated code which will only make it worst lmfao
@@yassir-eh using it wrong, you can be imperative with the prompts, so it will produce the code you want, if something's wrong, it's your fault, the level of freedom you give it is controlled by you, disconsidering the hallucinations.
@@korbpw 1. Using it wrong? I'm sorry but there is no definitive right way to prompt an AI, yes there are things that can make the prompt better, be there is no certainity that a prompt will give you a 100% correct answer. But fixing something using the correct logic yourself has better certainity, if both take the same amount of time becuase you are stuck trying to figure out what to ask the AI then why not just fix it yourself and Use AI for things that are just mundane and tedious.
2. Accuracy, right now a really good programmer can match a LLM model's accuracy in fixing bugs if the task is hard to explain to the AI. So its better to just do it yourself
Manual code reviews as a team are highly valuable. We used to do this when I was a junior developer and they were fantastic. We would get together as our small team and review the changes made throughout the week. We literally reviewed every change. Now, our commits were not always the smallest (before the days of DevOps and microservices, though we were mostly doing DevOps and single it web development... it was already microservices... because web pages were their own code). It could still be done today, but perhaps picking a subset of changes to review would work, as well. Being able to discuss the changes, the patterns, etc. that were implemented made for highly quality code in the future.
Sublime text has the remarkable new feature of staying the same forever. It’s like a 90s pickup truck. It does what you want it to do and not much else
I'm starting to get sick of all "THIS IS THE NEXT IDE" videos. First, zed, then cursor, now back to vscode. Man, just use nvim!
Literally me!
avante.nvim looks interesting though
Zed is Actually something new
@@abdullahzafar4401 Tried it, just isnt as appealing to me.
Does nvim have cursor like autocomplete features?
I only use chatgpt and I don't even use it in my editor. I just use it to plan my projects or to learn something new quickly.
I've used both and I feel like the method you're describing just always works better and more reliable, even though you have to go to your browser
Vim users: look at what they need to mimic a fraction of our power
Could you enlighten me what's the equivalent in Vim? Thanks in advance.
But imagine how efficient an LLM would be at managing everything from a CLI ...and it's gone.
@@ichigo_huskyWell if you have time, you can configure anything you want, and it's easy thanks to the neovim ecosystem, you literally create your own editing environment, for free.
@@ichigo_husky Vim can let you do whatever you want with your editor. Same with Emacs.
@@ichigo_husky do it on your own, you can mimic everything and much more in vim
4:40 vitest allows for in source testing, but like you mentioned shortly before the timestamp… Tests sometimes have way more code than the functions they test. The size of files can really blow up. But you can have vite prune it in production and avoid exporting everything.
I often find I only want to test exported code anyway, though. Anything that isn’t exported will be used in the exported function, and their behaviour is tested in a more integrated way which I prefer all around. I’m not a 100% coverage kind of person.
This video perfectly summarized what I've been dealing with writing C# using Copilot or Codeium, partly the plugins don't work, and partly their answers are just complete bullshit or don't write in the style they should write in - and then you post your code into ChatGPT instead and that works way better for some reason
1000% agree on your comment about testing. Thank you! 😍This is why human testers / QA Engineers / SDETs are still going to be needed for quite some time.
Theo Ad section is now a aesthetic PM and a skater boy dev
I'm ready for this to be a whole sitcom
“But it doesn’t seem to work for useful things” seems like a pretty good summary of a lot of LLM stuff involving code 😅
It is becoming increasingly hard to pick the right AI assistant. There are already dozens of them and you cannot fully try them all to decide which one suits you the best.
they're all equally useless. I tried them for like 6 months and they didn't make me any more productive.
Cursor has some awesome UX, and that makes difference for me. Copilot team even though very active they seem to be focusing more on expanding features rather than polishing base experience.
Cursor's product is VSC (MS) and chatGPT (MS). Microsoft can kill it whenever they see it fit.
microsoft does not own chatgpt they just have a big part in it
@@thismightbetwoMS does not need to own the thing to exert control over that thing
Nah. Claude is the secret sauce of cursor and a lot of impressive seeming ai products that are just Claude wrappers. It’s significantly better than chat gpt and therefore anything Msft has. Chat gpt has a serious model drift problem
@@melkorbane Claude is definitely not as good as Chatgpt o1
Better to control than to own (tax reasons). Why all the super wealthy put all their money in Trusts and Foundations, such as the Bill and Melinda Gates foundation. Has nothing to do with charity… same with Warren Buffet 😅
I fear that Junior devs will use these AI features and flood code bases with generated code that is not even needed...
Worse, we don't know what the license of the original code used was.
thats what code reviews by seniors are for.
You don't need to fear it, just accept that it will be the new reality 😂😂😂
Also restricting it to not use certain things or saying "use the packages I already have in the project". But you can be more specific bout how you expect it to make things type safe. Give it examples of good and bad, etc. Also ask it to explain the answers so you can tell it what to not do.
Cursor also has the "include all files as context" which is interesting (CTRL+Enter when you send a chat in Cursor). The inline spinners imply that it's doing some kind of chain of thought pre-processing of all the documents before it reads your instructions.
About AI written unit tests :
Totally agree their point is to ensure the code does what a human would expect.
But I also think there is a use case for ai in it. The same way a dev should understand the code written by an AI, if an AI writes tests well enough for an human to understand and review, the human can assert that the test is testing properly.
What makes unit test writing such a pain is the repetitivenes of the syntax that is needed to make them feel comprehensible.
Using AI would be a huge plus to make it easier.
One may fear that in the long run devs might forget the proper use of unit tests entirely, but if we can get AI to be explicit enough it could also make a positive feedback loop to better understand unit tests as we write them.
Telling an open source project to use a closed source project to review their PR's is kinda off to me..
Are you allowed to work on open source projects on windows or does that being closed source compromise everything?
@@MerkieAE Windows isn't a SAS system where all the data is stored in the cloud. Further Windows could be used by a developer but wouldn't be a dependency directly, you can just use something else without loosing any data.
Honestly, I’m okay with AI/LLMs sucking at writing or rewriting chunks of code.
I can do those things myself and most of the time those kind of things aren’t really a time sink.
Where I want AI/LLM tools to be good at is doing the mundane stuff: autocompleting things I was already planning to write (like comments or rewriting a bunch of same-y code), filling in mock data and checking if I could write something more efficiently.
And that’s already been quite good and reliable with most of these tools (if you keep their pitfalls in mind) in my opinion.
When it tries to do more advanced stuff and I have no idea whether or not the answer is reliable or efficient, I might as well not use it. It takes as much time to google if my LLM tool screwed up as googling how to solve it on my own.
Someday I'll be less greedy and pay for an AI editor assistant.
nah
Or be even less greedy and hire a dev (who will use these to choose anyway 😂)
@@HT79almost all of my coworkers have tried using these and just gone back to the normal methods.
This seems like a normal coding session, when working with AI. It’s equal parts masochistic and alluring to spend an hour to trying to teach the AI to write a 3 line method. 50 shades of AI
AI is gonna take our jerrrbbbs
This is the vid for me where Theo started to implements ads. I like the roll and you did warn us.
Is it wrong that watching Theo struggle with testing was cathartic? It was like, "Welcome to our world." XD
New ad breaks are great!!
I'm more lenient toward the new Copilot features since they are still in the preview stage. Once it's out of preview, I will try it and then decide whether to switch from Cursor. The half-price for Copilot is not a minor benefit for me.
1 quick question, does it require u to pay copilot monthly subscription? I think its 10$ right?
I feel like the 'generate tests' will be really useful for meeting BigOrg code coverage expectations, but as you hint, not actually helpful in testing user expectations.
8:40 "And it's the exact same code"
That has been my experience when trying to make it do something I 'really' needed to get done quickly and didn't know what to do.
AI is, very seriously, dumb as hell. And the people who think it's just about to take all our jerbs are on a similar level of regurgitating information.
And I will admit, I kinda was one of those people... back in Dec. 2022 when a huge increase in tech tricked my monkey brain.
See my comment too.. Things like copilot has gotten SO much worse too. Now its not even usable as a oh rewrite this to take a object instead of a array of object, or oh i missed checking errors after these calls, add it
So ai is "pain in my ass". Now I can focus on my career.😅
As a well experienced dev I don't need a cursor. I write a few lines of code that do many great things. Just learn clean code. Then you just need to learn new things like firebase or any API service integration while working with tasks. ❤
Iam wondering when will typescript and vscode will get annoying enougth that we switch back to sublime and js
I dont know I feel biased but since Cursor came I am really like it, previously for so many years I was a VScode user.
timing is off between the intro to the ad, you say "watch this ad" after the ad
Good catch, just trimmed in YT editor it will persist in a few hours
@@t3dotgg thanks for the great content, love the new ad format.
Was sad to see you had to change the way you did the ad stuff but this is fun
Supermaven blows all this crap away. It has been a real revelation for me, and I was a Copilot early adopter.
I've been using Zed and that experience has been pretty great so far
The functionality added seems more in response to Sourcegraph Cody.
AI coding tech will be as commoditized as calculators in the next 5 years. Every product remotely related to apps will have it as a feature. Every product that could benefit from such "AI extensions" but does not risk becoming not competitive. It's kind of like mouse input. Imagine having a Windows or Mac app that can only be used with the keyboard?
Last year one term under which that was coined was "Polimorphic apps" which rewrite themselves based on user input.
I do belive that the end destination of this kind of thing is integration into OS, OS provides GUI, and it will provide AI friendly app layer on the operating system level removing the need for the majority of small apps and possibly allowing customization of large apps.
That is what I suspect will be slowly happening, will take a while, a lot of "tech dept" to untangle both on software and human habits level.
I got told to use cursor… but from my view point I just wait until there more stable tool come out many years before I start using it. But I don’t know why people get mad t me. I just want stable not new feature
My thoughts exactly! I see everyone talking about cursor and it was tempting to use it, but honestly Im not gonna waste time to learn it if in a couple of months vscode(or some of the other main ones) come with and incredible update that'd make cursor obsolete... which Im pretty sure it's gonna happen to cursor next year.
Who actually thought Cursor was going to fork a Microsoft project and out compete them when it comes to DevDiv and AI?
I've been using cursor for a little bit. And while I like a lot of it. Sometimes I find that I spend more time fighting with it's autocomplete than working with it. When it gets it wrong, it so much harder to ignore it's suggestions. And sometimes if feels like it's making changes without me confirming them.
The insane feature of cursor is "Cursor lets you breeze through changes by predicting your next edit.". All other feature is not important, I will not go back to Copilot until they have the feature
Does it generate code on the fly based on your previous code like cursor?
Writting and endpoint the way your team does it and have it auto complete or generate the rest of the endpoints for you is amazing. (Based on your models or services as well).
Having the AI adapt to your code
"none of that makes any sense to me, let's see what they really added." LOL
As of now, I'm often disabling the AI helpers. By default, they keep distracting me so much that I forget what I wanted to do, by behaving like a chatty person looking over my shoulder and talking bullshit. It's mostly noise obfuscating useful code completions. Mostly because it's missing context I guess. When I'm about to write a comment, I do this because there's some non-obvious reason. I don't want an AI to guess obvious things and talk over me, but that's what happens right now.
But I use AI from time to time to help writing tests. That's so much boilerplate code to write, and I keep forgetting the nuances required for writing tests. However, this also only works well if there's not too much context needed.
I can use continue dev extension to send directories in context and I can use any model, including local Ollama...
Is all of this for Copilot only ?
ai assistant driven development is impressive but still not hacked. you switch from a designer into an assembler. I wouldn’t recommend new people just dive into it as you do need some experience to stitch things together in a way that is functional and maintainable
Behold, the result of multi-billion dollars worth of research and funding - broken on first few prompts. Here's to wishing the trend continues 🙏
lol solution to when every possible key binding is already taken. "F*ck that, delete"
The no unchecked index typescript error thing is bullshit, the array length was checked to be 1 so reading 0 should not return undefined.
I can totally understand the copilot messing it since I would've too.
Same. I was looking at that and I couldn't find the issue. The array is known to have exactly one item and the type prevents it from being null/undefined, so what's the problem?
I get serious Map Men vibes from these sponsor segments 😆
Like a wise man once said:
"If it's not free, it's not for me"
"If you're gonna ask me to pay, get the f away"
The title of this should be CoPilot (in VS Code) is fighting back. I thought the VS Code codebase was being updated to work more like Cursor 😢
what vscode theme are you using? I like that
Ive no need of cursor when supermaven already satisfies my needs
Supermaven is also starting to work on an editor.
@@Ruzgfpegk You mean their own editor? Is this gonna be a VScode fork?
@@besizzo My other comment got blocked because I linked to their website but yeah, they raised $12M for it and the wording seems to point to a new editor.
@@besizzo I think they meant Visual Studio or Rider, IDE in general
@@solivagrey1707 My comments went to the mod queue because I put an URL to their website, but they raised $12M "to build a text editor" so it looks like it's a new one.
lol I respect that "I'm gonna leave this here so that it fucks me the next time I open VS Code", a man who values comedy over productivity even at his own expense
Using comments to drive these ai completion tools has been my preferred method for working with ai. Gotta admit, Cursor is pretty slick though.
the ad sections are getting better lol
Legendary Battle!
Hmm... New Ads...
Maybe add a more obvious transition ("And now, a word from our sponsors") - the flow needs to be smoothed out a little somehow.
Otherwise, cool idea... May have been good to go slightly deeper on the Review App suggestion, and how it works.
Every video… “btw I’m an investor” when you probably invest spare change through acorns lol
Have you tried void and how do you think about it when compared to cursor?
Cursor is way better for someone who doesn't code everyday and needs some heavy help. Glad they added o1 to paid plans
Has anyone noticed VSCode became real slow lately? Everyone used to say it was slow but I had not noticed that until very recently. I wonder why.
Would be great if they could use the whole folder as context, not just selected files. That way you could make updates to the whole codebase. But what do i know. Im not even a dev.
I think using AI to write unit tests is fine as long as you are actively reviewing the tests that it spews out in considerable detail, otherwise there's no point writing tests. Unfortunately I suspect the reality will be that too many devs will implicitly trust the AI generated tests without any consideration.
This is why competition is great.
some VSCode settings have to close the VSCode and reopen to work. Maybe this is the case. Just entirely close VSCode and reopen it and test again.
Do you have a .cursorrules ? You should try to finetune your cursor to your project. You can add examples and instructions focused on your style, or have always in mind the TYPESAFETY of everything
Theo switches tools like they are clothes.
The pig-latin usage is the most useful feature 🤣
I believe the inline chat/commants and chat are separate settings, and the instructions you gave it were set in the Chat setting, not in the inline command setting.
Honestly I don't need any of that... I just want the Cursur Tab and cursur predict feature and nothing else
Not respecting instructions is a staple of generative ai
It's funny how programmers see the writing on the wall yet keep looking forward to better AI assistants but Figma doesn't have the spine to release Make Designs. Really hoping there's a Cursor-like Figma killer in the works. v0 doesn't really solve the rapid ideation and prototyping plain designs do
I love the comparision. Can you do the same for ZED editor with Zed AI?
Yay!
Hi
@@d5hcode Hey!
I'm too old fashioned, does not have any ai in my editor and did'nt try one. Only watching you tried a lot of them :D
With Cursor I can chose which AI API is used.
In this video here you use co-pilot which uses the despicable OpenAI API.
Tons of bad feeling from me for everything-OpenAi.
vscode is Microsoft-owned, and they of course want to push co-pilot hard since they've invested tons in OpenAI.
What's the VS code theme in-use?
Zed with Supermaven and Claude AI is just AWESOME
My laptop have 16gb ram and as a react-native dev,
I need atleast a browser with some tabs and an emulator always opened
Vscode is good but it's SUCH A RAM HOG
I wish a better font and easier prettier integration in zed but still better overall than vscode
ok now when is the LSP going to catch up
Commemting at 8 mints....Well in 8 mins I would have written the test myself and get it working including settting it up and installing dependencies
These tests have so many mocks, that the tests get practically pointless
Could you share the name of the dark theme and the font being used?
I auto like your videos before watching btw 🙂
I highly recommend you check someone really good with Cursor to see how you could use it to it's full potential.
You're using the "Composer" feature a lot, where it's not the best to use.
Also, you biased the Cursor solution with the VScode solution. That's why it was bad, i tried the same thing and it was way better.
In your last video about Cursor, I saw that you'were not using the context and the CMD+L at all, which is the most powerful at current moment.
Cursor makes same mystake theo: cool
Later vscode does the same. Theo: this is not good as cursor😂😂😂😂
They need to add somthing like cursor compose to be still in the game. To have an ai with multi file edits