VS Code Fork 'Cursor' - the ChatGPT of Coding? (tomsguide.com) 69
"Sometimes an artificial intelligence tool comes out of nowhere and dominates the conversation on social media," writes Tom's Guide.
"This week that app is Cursor, an AI coding tool that uses models like Claude 3.5 Sonnet and GPT-4o to make it easier than ever to build your own apps," with the ability to "write, predict and manipulate code using nothing but a text prompt." Cursor is part development environment, part AI chatbot and unlike tools like GitHub Copilot it can more or less do all of the work for you, transforming a simple idea into functional code in minutes... Built on the same system as the popular Microsoft Visual Studio Code, Cursor has already found a fanbase among novice coders and experienced engineers...
Cursor's simplicity, working from a chat window, means even someone completely new to code could get a functional app running in minutes and keep building on it to add new features... The startup has raised over $400 million since it was founded in 2022 and works with various models including those from Anthropic and OpenAI... In my view, its true power is in the democratization of coding. It would also allow someone without much coding experience to build the tools they need by typing a few lines of text.
More from ReadWrite: Cursor, an AI firm that is attempting to build a "magical tool that will one day write all the world's code," has announced it has raised $60 million in its Series A funding round... As of August 22, the company had a valuation of $400 million, according to sources cited by TechCrunch...
Anysphere is the two-year-old startup that developed the app. Its co-founders are Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger, who started the company while they were students at MIT... Using advanced AI capabilities, it is said to be able to finish, correct, and change AI code through natural language commands. It currently works with JavaScript, Python, and TypeScript, and is free for most uses. The pro plan will set you back $20 per month.
But how well does it work? Tom's Guide notes that after requesting a test app, "It generated the necessary code in the sidebar chat window and all I had to do was click Apply and then Accept. This added the code to a new Python file including all the necessary imports. It also gave me instructions on how to add modules to my machine to make the code work.
"As the chat is powered by Claude 3.5 Sonnet, you can just have it explain in more detail any element of the code or any task required to make it run..."
Andreessen Horowitz explains why they invested in the company: It's very clear that LLMs are a powerful tool for programmers, and that their coding abilities will improve over time. But it's also clear that for most coding tasks, the problem to solve is not how to make LLMs perform well in isolation, but how to make them perform well alongside a human developer. We believe, therefore, the interface between programmers and AI models will soon become one of the most important pieces of the dev stack. And we're thrilled to announce our series A investment...
Cursor is a fork of VS Code that's heavily customized for AI-assisted programming. It works with all the latest LLMs and supports the full VS Code plugin ecosystem. What makes Cursor special are the features designed to integrate AI into developer workflows — including next action prediction, natural language edits, chatting with your codebase, and a bunch of new ones to come... Our belief is that Cursor, distinctly among AI coding tools, has simply gotten it right. That's why, in a little over a year, thousands of users have signed up for Cursor, including at companies like OpenAI, Midjourney, Perplexity, Replicate, Shopify, Instacart, and many others. Users give glowing reviews of the product, many of them have started to pay for it, and they rarely switch back to other IDEs. Most of the a16z Infra team have also become avid Cursor users!
One site even argues that Cursor's coding and AI capabilities "should be a wake up call for Microsoft to make VS Code integration with GitHub Copilot a lot easier."
Thanks to Slashdot reader joshuark for sharing the article.
"This week that app is Cursor, an AI coding tool that uses models like Claude 3.5 Sonnet and GPT-4o to make it easier than ever to build your own apps," with the ability to "write, predict and manipulate code using nothing but a text prompt." Cursor is part development environment, part AI chatbot and unlike tools like GitHub Copilot it can more or less do all of the work for you, transforming a simple idea into functional code in minutes... Built on the same system as the popular Microsoft Visual Studio Code, Cursor has already found a fanbase among novice coders and experienced engineers...
Cursor's simplicity, working from a chat window, means even someone completely new to code could get a functional app running in minutes and keep building on it to add new features... The startup has raised over $400 million since it was founded in 2022 and works with various models including those from Anthropic and OpenAI... In my view, its true power is in the democratization of coding. It would also allow someone without much coding experience to build the tools they need by typing a few lines of text.
More from ReadWrite: Cursor, an AI firm that is attempting to build a "magical tool that will one day write all the world's code," has announced it has raised $60 million in its Series A funding round... As of August 22, the company had a valuation of $400 million, according to sources cited by TechCrunch...
Anysphere is the two-year-old startup that developed the app. Its co-founders are Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger, who started the company while they were students at MIT... Using advanced AI capabilities, it is said to be able to finish, correct, and change AI code through natural language commands. It currently works with JavaScript, Python, and TypeScript, and is free for most uses. The pro plan will set you back $20 per month.
But how well does it work? Tom's Guide notes that after requesting a test app, "It generated the necessary code in the sidebar chat window and all I had to do was click Apply and then Accept. This added the code to a new Python file including all the necessary imports. It also gave me instructions on how to add modules to my machine to make the code work.
"As the chat is powered by Claude 3.5 Sonnet, you can just have it explain in more detail any element of the code or any task required to make it run..."
Andreessen Horowitz explains why they invested in the company: It's very clear that LLMs are a powerful tool for programmers, and that their coding abilities will improve over time. But it's also clear that for most coding tasks, the problem to solve is not how to make LLMs perform well in isolation, but how to make them perform well alongside a human developer. We believe, therefore, the interface between programmers and AI models will soon become one of the most important pieces of the dev stack. And we're thrilled to announce our series A investment...
Cursor is a fork of VS Code that's heavily customized for AI-assisted programming. It works with all the latest LLMs and supports the full VS Code plugin ecosystem. What makes Cursor special are the features designed to integrate AI into developer workflows — including next action prediction, natural language edits, chatting with your codebase, and a bunch of new ones to come... Our belief is that Cursor, distinctly among AI coding tools, has simply gotten it right. That's why, in a little over a year, thousands of users have signed up for Cursor, including at companies like OpenAI, Midjourney, Perplexity, Replicate, Shopify, Instacart, and many others. Users give glowing reviews of the product, many of them have started to pay for it, and they rarely switch back to other IDEs. Most of the a16z Infra team have also become avid Cursor users!
One site even argues that Cursor's coding and AI capabilities "should be a wake up call for Microsoft to make VS Code integration with GitHub Copilot a lot easier."
Thanks to Slashdot reader joshuark for sharing the article.
Never going to work for many use cases... (Score:4, Interesting)
AI coding seems great but good luck using it to write, say, a conversion tool for a weird undocumented proprietary format or a high performance direct3d11 graphics engine (both things I have had a hand in writing)
A tine who thinks AI can replace all coding either has a vested interest in AI coding or doesn't understand programming.
Re: (Score:3)
How was your direct3d11 graphics engine coded? I mean, you had a specification right? You don't think an AI could have been asked to write some of the functions if you described it? Go ahead and describe one, let's see if AI can do it.
Re: (Score:3)
No, I didn't have any kind of spec and I doubt you could come up with a description of what some of the giant mess of this code in our engine does that would allow any AI to produce it.
Re: Never going to work for many use cases... (Score:4, Insightful)
Re: (Score:3)
Considering that the engine I am referring to started out life as a bunch of enhancement patches to game running on Direct3D8 (before eventually becoming its own standalone thing and being ported to Direct3D11) and has been hacked together by a bunch of people in their spare time (quite a few of whom are no longer around), its not surprising the codebase has so little documentation.
Re:Never going to work for many use cases... (Score:5, Interesting)
I've been using Claude a lot recently for coding a React.js app, and it's really good. There was only one issue that I even had to "help" it with (e.g. it didn't fix after three iterations), and the only "help" I needed to give it was some basic debugging info (copy-paste from the DOM). With everything else, it's been enough to just describe the symptoms. All the more impressive, to think about it, since it's having to "work blind", just from text descriptions of what I see and what I want (though I guess I could provide screenshots or illustrations!).
My main limitations have been:
* Maximum conversation length (a growing problem as my code grows)
* Number of queries per day
* Its tendency to only write the changed parts (something I could fix by telling it not to, but then the first problem gets worse), so I can't just copy-paste it all at once.
* Time I waste doing copy-paste, edit errors, etc.
I could get past the first two problems - and probably the third, and part of the fourth- by paying for the premium tier, but I've always balked at adding another $20/mo bill (I'm a cheapskate). But if it comes with an IDE that simplifies the workflow, so I can just type to Claude and have the project automatically updated... yeah, that might pull me off the fence; I could iterate SO much faster. I think I'll try it out this evening.
Re: (Score:2)
(The one thing it had trouble with was a Leaflet map covering up dropdown menus; it was trying to fix it by e.g. adjusting the Z on various components, but Leaflet was doing something weird with the DOM (I don't remember exactly what), which made it so that no Z change would solve the issue)
Re: (Score:2)
I tried it with a simple request ("Create a function that renames multiple files in a directory, using PowerShell") and it obliged.
function Rename-MultipleFiles {
param (
[Parameter(Mandatory=$true)]
[string]$DirectoryPath,
[Parameter(Mandatory=$true)]
[string
Re:Never going to work for many use cases... (Score:4, Informative)
It does.
Claude doesn't do images.
To be fair, there are practical reasons - LLM token-related memory needs are O(N^2) with respect to context size. Also, for a given amount of memory, you can increase the batch size if the context size is smaller, and high batch sizes improve performance. Claude's free tier max context size is variable based on load.
I don't even pay for TV service. I hate ongoing costs.
30000 users. Also, the Slashdot reporting is wrong - they raised $60M at $400M valuation. Total funding thusfar is $71M.
That said, if you're investing in startups based on their current revenue status vs. their current valuation, you shouldn't be in the business of investing in startups. Startup valuation is deeply related to forecasts. That doesn't mean that $400M is "the right number" - it could be way off; startup valuations are extremely difficult. Nor does that mean you can calculate the estimated number of users so simply. E.g. an investor might have the view that there's only a 1% chance that they hit it big, but if they hit it big, they'll be worth $40B.
Re: (Score:2)
Some AI companies like Midjourney are already plus, so that's a demonstrably false statement.
Re: (Score:2, Funny)
Re: (Score:1)
I think you are giving 99% of the coding community too much credit. Giving an individual an even slightly more detailed objective is unlikely to return a timely workable solution. But saying you need a simple function to do x in struct b..... to give a general scenario? AI chat appears to do that extremely well. A project lead takes those elements and decideds if they play nice with each other. The conceptualization you are crediting (most)coders with goes away when the specs are handed out
Re: Never going to work for many use cases... (Score:3)
If it works for some use cases, isnâ(TM)t that enough? We have spoons and they donâ(TM)t work for most use cases.
Re: (Score:2)
The only thing that matters is that AI is getting better, and you and I aren't.
Re:Never going to work for many use cases... (Score:5, Insightful)
AI coding seems great but good luck using it to write, say, a conversion tool for a weird undocumented proprietary format or a high performance direct3d11 graphics engine . . .
AI makes the easy part easier. It doesn't solve any hard problems. For that you need devs.
Who knew that the easy part needed to be easier? AI vendors.
Re: (Score:1)
...it will not and cannot work well for any task that requires actual understanding. Coding is such a task.
"Cannot" I can see, but "will not" is a bold prediction.
Re: (Score:2, Interesting)
Having used 4o a few times to produce C code for some simple functionality I'm pondering of using, it's indeed not for anyone who has barely any knowledge of programming because they'll end up with a giant clump of code in one file for sure.
It might technically work, but it's far from easy to read, understand, and modify.
If I use those code snippets, I'll have to rewrite them just to fit my writing style and source files structure.
I even noticed it keeps using i++ throughout instead of ++i, which to me show
Re: It really does not matter ... (Score:3)
Re: (Score:2)
Yep, that is about what I expect. It may be able to help the incompetent with very simple code, but that is it and even there it may fail. Fro the competent, I suspect that it takes the same time or longer to write code with its "help" than without.
Re: (Score:3)
i++ is standard and makes no performance difference in most situations (excepting custom objects that override the operator). As a general rule, ++i is used only when you care about the return value. You technically could use it in most places, but again, that's not "the norm".
You're coding weird and thinking you're outsmarting the compiler. You're not.
Also, you can very much ask a LLM to refactor your code into multiple files. Claude also has artifacts to keep track of them. This app is taking it furth
Re: (Score:2)
The name of the language for object-oriented C is literally "C++". The postfix increment operator is the "normal", most common one that people use, so complaining that the AI used it is.... certainly a choice? As a general rule, the distinction only matters for objects that override the function.
I mean, there's lots of things, stylisically, that say ChatGPT does that I disagree with (for example, four-space indents in Python instead of two), but I grudgingly accept that it's doing things the way that most
Re: (Score:3)
I even noticed it keeps using i++ throughout instead of ++i, which to me shows the AI does not understand C in depth.
OK boomer! That's been irrelevant for probably 20 years. Maybe more.
For example:
https://godbolt.org/z/6KnaqhT3... [godbolt.org]
identical code is produced.
Re: (Score:2)
I see those who say post or pre incremental doesn't matter, don't understand the implication of using it beyond the single i++; statement and rely on the compiler to figure out what they meant.
Use ++ or -- in a combined statement (for example calculating the position in an array based on sizeof), and things turn out different because the compiler might not figure out whether to use i as is or the incremented version.
Writing ++i is a good, safe habit which causes no mistake (and bugs) on what value of i is s
Re: It really does not matter ... (Score:3)
Iâ(TM)ve been working with Cursor for the last six months now and paying for it. Itâ(TM)s certainly not going to replace a coder anytime soon, generating code from the chat prompt is marvelously boilerplate. Has it been a time saver? Absolutely yes!
What I love most about Cursor is writing a comment on what the code should do next and the AI provides a suggestion. I find that these suggestions are really helpful for articulating the solution. Sometimes it just spits out a complete solution, obvious
Re: (Score:2)
So I just started using it yesterday. A lot of frustrations early on - the IDE needs refinement, and even crashed on me once - but once I got used to it, and got past some of the bugs, the utility is obvious.
For anyone just starting out, protip: if you do ctrl-k (ask for code changes) within a file, it will only work on a selection of code, not print any text output, and often bug out (esp. when working with Claude, which is by far the best LLM for code offered). But if you do ctrl-k after clicking on a f
wake me up when reviewer builds a real app with it (Score:5, Insightful)
wow, the author built a demo app with it. Something pretty much every ide could do from a template since like.... 1990 or so.
yawn.
slashvertisment much?
what I would like to see is an AI that can do quality refactorings of my rust code, eg move struct into a new file, rename it, and change all its imports and references to it. now that would be an actual help.
I don't need regurgitated code snippets from stackexchange or github, thx.
Re:wake me up when reviewer builds a real app with (Score:5, Insightful)
It's always the same marketing bullshit that we've been hearing for the last 30 years. "Anyone can build an app in minutes!". Conveniently ignoring that you need a lot of knowledge even before coding a single line - architecture, deployment, ux, qa, testing, etc. A novice wouldn't get anywhere, especially "in a few minutes". instead of making the job easier, these tools increase productivity.
Re: (Score:2)
"Anyone can build an app in minutes!". Conveniently ignoring that you need a lot of knowledge even before coding a single line - architecture, deployment, ux, qa, testing, etc.
No, not really.
Yes, anyone could now build a very simple app withing minutes, and they don't even need to know anything about architecture, deployment and all that crap.
This application is not going to build the next-gen triple-A game or the next-gen OS for you, but it can help people create simple apps and/or scripts which just work, without having to worry about optimization, architecture and so on, because it won't matter for personal use or infrequent usage.
Here's a random example: I needed a PowerShell
Re: (Score:2)
Re: (Score:2)
Sounds like someone who hasn't worked in an office.
Who said anything about an office?
This is not for office productivity or coding. This is for home use where you can automate some simple things for your own stuff, without having to rely on random strangers on whichever forum, or even for adding some functionality to existing code you got from Github or wherever.
Maybe it's a niche case, but tools such ChatGPT and now this help me tremendously, because my (too many) side projects involve having to write small stuff in a gazillion different programming and sc
Re:wake me up when reviewer builds a real app with (Score:5, Informative)
It's always the same marketing bullshit that we've been hearing for the last 30 years. "Anyone can build an app in minutes!".
More like 50 years. [wikipedia.org]
Re: (Score:2)
It's always the same marketing bullshit that we've been hearing for the last 30 years. "Anyone can build an app in minutes!". Conveniently ignoring that you need a lot of knowledge even before coding a single line - architecture, deployment, ux, qa, testing, etc. A novice wouldn't get anywhere, especially "in a few minutes". instead of making the job easier, these tools increase productivity.
There really is a sea change this time, but yes - the LLM tools are making project management skills way more important than simply grunt coding skills. Because it's the grunt coding skills that the LLMs are good, fast, and tireless at.
They hid it fairly well... (Score:2)
... and did a better job at it then most. But, in the end, this was a PR release (aka "Slashvertisement"):
wake me up when there is a local tool (Score:1)
Here's the source code for the demo app (Score:2)
$ cat helloworld.cursor
Hey ChatGPT, write me a program that says Hello World!
But being an Electron app, this VSCode fork won't run it without half a gig of RAM and disk.
Re: (Score:3)
So I pasted it into Claude [pasteboard.co] and got this [pastebin.com] and this test case [pastebin.com], from which I get this output [pastebin.com].
You're the one tasking this problem, so you tell us whether it works and whether you want improvements anywhere. This took seconds, and even if you're not happy with it, it's bare minimum a framework (and requests for changes also take seconds).
Re: "More or less" is doing a lot of heavy lifting (Score:2)
For larger instances I'd recommend using something like Concorde TSP solver [uwaterloo.ca] (free) along with pyconcorde [github.com], or Gurobi [gurobi.com] (commercial) instead of solving the TSP solely using Python... but otherwise, pretty cool that Claude pulled together a solution like that.
Boilerplate (Score:3)
Most of the time I can't express an algorithm in English any more succinctly or clearly than in code.
And half the time as I write the code I realize my initial idea was incomplete.
And sometimes it was flawed, requiring debugging.
I get how AI can help with huge frameworks which are tedious and full of boilerplate code (that is, they suck) but how can it code an algorithm for me that I don't yet know how to describe?
Re: (Score:1)
It's rarely "writing algorithms in English". You describe the big picture and let it implement the big picture, instead of dealing with the low-level algos. If there's something too complicated and you don't want to describe it, no worries, just ask for a skeleton function for that part of the work for you to fill out on your own.
The other thing is lets you iterate quickly. So if you implement an algo, and then realize, "oh, I'm stupid, that physically cannot work", it's not some massive work to start ove
Re: (Score:2)
https://www.cursor.com/ [cursor.com]
Linky? (Score:2)
Would it have killed someone to actually link to it?
Seriously, the Tom's guy goes so far as to put his own playground code on github and link to that, but he can't link to the tool itself, in a news story about the tool?
(Yes, yes, google, but I thought the whole point of the web was to hyperlink to stuff when you mention it. Also, get off my lawn!)
Re: (Score:2)
(Yes, yes, google, but I thought the whole point of the web was to hyperlink to stuff when you mention it. Also, get off my lawn!)
Nah the whole point of the web is to let people argue endlessly over stuff most people don't care about, post stuff on Instagram and think they are an influencer, and allow idiots a much broader audience to show off their stupidity.
Now, if you'll excuse me, I'll go back to yelling at clouds...
Re: (Score:2)
https://www.cursor.com/ [cursor.com]
Re: (Score:2)
Are you not longer on their site? Looking for a link, I mean.
Possibly even enraged enough to leave a nasty comment about the lack of links?
Or even so enraged you sign up for an account to be able to comment?
I'm sure longer read-times and more "engagement" are numbers that increase to value of advertisements on that site. We lost Anandtech already, losing Tom's Hardware would likely be too much if that would happen this year.
But yes, I completely agree that it would have been better, if there was a link to t
Cursor is neat (Score:3)
Tried it, made it write a snake game. The game had some fundamental flaws like you could click a 180 degree direction and kill yourself, but still, a snake game within minutes.
However, that's a trivial game. LLMs are known to solve trivial things, and I'm not too sure of the value. It's also a fact that the LLM vendors are artificially boosting trivial code via finetuning to make their product look better - so yes, of course it will be amazing for trivial code (which is still useful, don't get me wrong).
I've recently been working with NestJS where ChatGPT is often wrong when I want to do something, because there is not a whole lot of data to work with. No doubt Cursor can help with some boilerplate but I'd wonder just how useful it is in a corporate environment (lots of proprietary garbage) versus GitHub Copilot. I'm a GitHub Copilot subscriber and it seems to meet the bar of AI assistance. That is, constant code suggestions, ability to query LLM, and highlight and replace with LLM suggestions. I haven't tried Cursor in a while but I don't agree that its implementation is a "wake up call" if you have real work.
Re: (Score:3)
However, that's a trivial game.
It's not just a trivial game, it's a trivial game for which examples abound.
Re: (Score:2)
Then ask it for a different game, or throw a twist into the rules.
I regularly use Claude for coding, and if you're trying to claim it can't handle anything remotely complex that's not common on the internet, you're outright gaslighting. Some complex things might take more than one iteration, but as if that doesn't apply to humans too? Its "first draft" is much more likely to work straight out of the box than my "first draft", and it produces it in seconds.
Re: (Score:2)
I regularly use Claude for coding, and if you're trying to claim it can't handle anything remotely complex that's not common on the internet, you're outright gaslighting.
You need to get out more. Disagreeing isn't the same as gaslighting. Me having a different experience from you isn't the same as gaslighting. I don't find the AI tools very useful.
They're good for bashing away at big well known APIs. They're good for doing kind of repetitive autocomplete on steroids when you have a lot of cases. They're OK
Re: (Score:2)
The ChatGPT of ___ (Score:3)
This is going to become the marketer's catchphrase. Any time a marketing department has to say their product is "the ChatGPT of (whatever)" you can be sure that this company _hopes_ you will think their product is as "amazing" as ChatGPT. If you've got to drop the name, there's probably a reason for that.
So what's new here, exactly? (Score:2)
Is it that you can swap out LLMs? Or is it that it's somehow superior to GitHub Copilot and other IDEs?
Re: (Score:2)
So, I started using it yesterday. And yes, you can swap out different LLMs, though the two provided Claude versions are definitely better than the other options. But that's not the real issue.
So it of course has the "mundane" things like code completion.
It also has the ability to chat with your whole codebase, asking any questions about it, adding in additional information, etc, like, "Hey, this variable is null in this function when I expected it to be set - how could it possibly have been null?" and get
Re: (Score:2)
Thanks, that summary is helpful.
My first reaction is that that's pretty far short of "it can more or less do all of the work for you, transforming a simple idea into functional code in minutes." And it's not that much different from what GitHub Copilot can do.
Now, one thing it does seem to be better at, from your description, is the ability to make changes in multiple files at once. That is certainly an incremental step beyond GitHub Copilot.
Weasel words (Score:3)
it can more or less do all of the work for you
So it can't actually do all the work for you.
Re: (Score:2)
Re: (Score:2)
Good point! I've certainly had Github Copilot do more than I asked, but the result was not helpful!
Re: (Score:2)
If AI worked, it would be in the VM (Score:2, Insightful)
Ominous name (Score:2)