Hacker News new | past | comments | ask | show | jobs | submit login
AI-Powered JavaScript Completions (kite.com)
141 points by csmajorfive 13 days ago | hide | past | web | favorite | 94 comments





FYI: Kite as a company had a backlash from the community after they acquired popular Python packages and injected their spyware and ads. So far they've learnt from that lesson, but who knows if they'll stay that way.

We hear you. We agree this was a mistake, and over the ~three years since then we've listened to the community, for example by releasing local processing for Kite.

Why is "send usage metrics" still on by default in the Kite.app preferences?

I was prepared to give Kite another shot but seeing that checkbox just made me lose confidence. Caddy almost got destroyed by this! Please please please reconsider and make telemetry opt-in, you don't have enough goodwill given Kite's history with exfiltration, and the product has access to too much vital information to be playing around with users trust like this.


Were the minds that decided this was a good idea in the first place fired from the company? I guess it was greenlighted from the very top. No idea how anybody could ever trust you ever again.

That’s actually worse than the last time I remember them being in the limelight for questionable privacy regarding sending your code to their servers.

Yes, code extraction is deeply worrying. I installed Kite once, but never enabled it for this reason.

The ads were specifically added into two community supported Atom extensions after Kite employees got contributor access to both of them. For those curious, here are the relevant issues:

https://github.com/atom-minimap/minimap/issues/588

https://github.com/autocomplete-python/autocomplete-python/i...


if you're concerned with Kite, you may want to try TabNine (https://tabnine.com), which works for JavaScript, Python, and all popular languages

Suggesting alternatives is fine in principle, but astroturfing a competitor's thread is not. You guys have crossed the line. Please stop and do not do this on HN again.

https://news.ycombinator.com/item?id=23160881

I suppose I should add that we have no interest in the competition or which product is better, but we have a strong interest in preserving the integrity of HN threads, and the community demands that of us.


Adam from Kite here. The team and I are thrilled to launch our new deep learning engine for JavaScript completions.

After years of work building our Python product — now to over 250,000 monthly active users — we decided we were ready to expand a new language. JavaScript was a clear choice given how well it pairs with Python, and it was the our most-requested language.

We took what we learned from building our Python product, adjusted it to JavaScript’s nuances, and trained the model on 22 million open-source JavaScript files. We then spent months dogfooding, user testing, and applying advanced model filtering techniques to ensure the UX was ready for everyone to enjoy.

We think the end-result will help you have more fun writing JavaScript, while also helping you ship software faster. We can’t wait to get feedback from you to help us improve our JavaScript completions from here! We'll be online to answer your questions here today as well.

Since a lot of folks ask, Kite continues to work 100% locally and Kite for JavaScript is free. We plan to release support for more languages later this year, and are taking requests at kite.com/letmeknow. Thanks!


Adam, why did you remove your response [1] to the hijacking of Atom packages from the Kite blog?

Kite was collecting usage data without consent through an unrelated and popular Sublime Text package [2], and they've sent the data to an IP address instead of a domain name, presumably to conceal their involvement. Kite has also transferred private package names to that IP address.

When they were discovered, Adam Smith has said that they've forgot that the data collection was taking place [3], after having paid the maintainer of the package to send data to Kite.

This is on top of Kite covertly buying up popular Atom packages to push their service on developers. [4]

I'm surprised their domain wasn't banned from HN, it has been repeatedly demonstrated that Kite can act maliciously whenever it's in their interest.

[1] https://news.ycombinator.com/item?id=14857944

[2] https://news.ycombinator.com/item?id=14902630

[3] https://forum.sublimetext.com/t/rfc-default-package-control-...

[4] https://news.ycombinator.com/item?id=14836653


Adam,

Intelligent autocomplete is great. Big fan of it. However, Microsoft has already built this out in VS and VS Code as IntelliCode, and it supports Python, Typescript, JS, and Java.

VS Code is the most used JS editor by far and has first party support for this that comes out of the box. What is the reason for Kite for the VS Code users out there?


Hi there, thanks for the question.

IntelliCode has a similar vision but uses much simpler models. IntelliCode can essentially give you to the top 5 attributes for a global type.

Kite uses statistical models which are orders of magnitude larger and richer, both in terms of the context they understand and the sophistication of the completions provided.

Here is a good example from the blog post:

  var server = express(); server.
IntelliCode returns:

  use, set, get, listen
Kite's first completion is

  use(bodyParser.json
(bodyParser is defined earlier in the file.)

Finally, IntelliCode only returns completions for attributes (foo.) whereas Kite returns deep learning completions in all the places they are useful.


Appreciate the reply and the additional explanation, best of luck!

The link has examples of how this is different than VS Code.

Intellicode is an extension add on from Microsoft. It's not default in VS Code. It doesn't look like they're making a distinction here.

It says it right in the blog post. VSCode only does fuzzy matching for javascript. This is an "intelligent" autocomplete, so the suggestions should be better.

No, IntelliCode is also powered with AI and does not do fuzzy matching. Fuzzy matching happens if you disable intellisense. https://visualstudio.microsoft.com/services/intellicode/

Fuzzy matching is done almost never, even ordinary intellisense is type aware

Hello adam how much memory it will consume? I just use tabnine before, and after 12 hours I uninstall it. Because it is use lot of memory and CPU Usage.

Hi there, Kite at most about half of what TabNine uses, often less than that.
whadar 12 days ago [flagged]

Congrats, but it seems that now the free plan has very few features, almost all paid? :(

It's abusive to try to manipulate a competitor's thread this way. Please stop now and don't do this on HN again.

https://news.ycombinator.com/item?id=23160881


Semi-related thought~ I find myself often dreaming about the AI-assisted development processes of the future. I'll frequently be writing code and start thinking "this task I'm doing is pretty formulaic and constrained-- it should be automatable someday, even if we haven't figured out how yet."

Examples:

1. I memoized the arguments for a util function. Repeat this on the very similar functions defined alongside it.

2. I need to put a wrapper component around six similar components. Just do it for me after seeing me do it once.

3. I imported a new logging mechanism and called a specific method from it in one of my class's methods. Repeat for four similar classes.

Basically I want a "repeat after me" functionality that can intelligently infer which part of thing I just did is generic and can thus be replaced with the next item in the series. In a dream future it would also infer what the next items in the series are.

I haven't a clue how such a thing could be created though, or I would already be on it myself. It would have to understand the code not just syntactically but conceptually.

God it would make development less tedious though. The difference between today and that future is the same as the difference between today and the days of punchcards.


When is it better to solve those problems through the organization of your code rather than an autocomplete feature?

The things you mentioned all seem to yield to the decorator pattern.

Creating the right abstraction (instead of using autocomplete) would benefit future collaborators who may not have access to the same autocomplete functionality.


The GP wanted to repeat some code modification 2-6 times. At that scale, adding abstractions often just adds unnecessary indirection and makes the code harder to read.

Depends on how often they intend to make changes to that code. But in general I think you are right.

Even if that was true, which I'm not arguing one way or the other, it wouldn't take much effort of imagination to come up with other use cases that aren't solved by decorators.

This has a taste of the old "they would have asked for a faster horse" argument.


I actually agree with you. I wasn't implying that the OP should consider refactoring their code. Just trying to understand for myself when I would prefer autocomplete over organization of code.

> I imported a new logging mechanism and called a specific method from it in one of my class's methods. Repeat for four similar classes.

How would an IDE know the local variable you've declared, that is a suitable match to the logging method's parameter?


How indeed. I think basically a rudimentary version might guess and then ask you if its mapping is right.

i.e. a modal asking:

"As done to MethodA, do to MethodB?" "As done to VariableNameA, do to VariableNameB?"

And the B values would just be dropdowns that would let you correct it. Idk.

I'm just imagining what you could do with some decent speech-to-text, NLP, and an engine that understands code conceptually on a basic level.

"Create a new enum on line 14 called SuchAndSuchCategories, with the values A, B, and C. Replace the string literal on line 27 with SuchAndSuchCategories.B."

"Move method A above method B."

"Add a new method to class X called doSomething() that accepts three parameters: a string called A, a generic called B, and a SomethingClass called C. Inside the method, start a for-loop...."

Like, idk. These aren't repeat-after-me tasks but they're complimentary to them. It would just be great to be able to just talk code into existence.


>"As done to MethodA, do to MethodB?"

Don't know about you, I would have no idea of a method name's exact spelling from a declaration multiple folders and files away.


Well, I mean you wouldn't really need to know the spelling. It would be picking the methods from the code, having parsed it. There would be no spelling involved. It's just a matter of it checking that its target is correct-- "Is this the one you wanted to change?" And if the dev doesn't know the answer to that, then they couldn't have made the change without the tool, either.

TabNine already does this "repeat after me" autocomplete really well, and Kite probably does too, you should give them a try!

These kinds of repetitive tasks are really where AI-based autocomplete shines!


I’m glad this works locally. I remember Kite for Python, one of the reasons we couldn’t use it was because it sent the code to a remote server. Glad to see this.

It's difficult to build a solution that works for all the use cases but I hope the team persists and so far, looking good! I love the examples of comparisons in the links posted, everyone should go check them out to see the difference if they're wondering if they should pick this up over the other implementations!

Why is Kite trying to access all of the folders on my machine?

When Kite starts, it indexes the .py files on your system so that it can have a ready-made code index to serve responses from. This is why on macOS, Kite will ask for access to certain folders on first startup.

There are some folders that Kite ignores, namely folders that usually contain private data. More info here: https://help.kite.com/article/97-adding-libraries-from-pytho...


Reminder that Kite hijacked data in the past [0]. Why should it be trusted now?

[0] https://news.ycombinator.com/item?id=14902630


Congrats Kite, Adam & the entire team for launching JS & Pro. Been a fan for a while. Fingers crossed!

How is this better than any IDE anyways? The examples look trivial and don’t seem even better than ultisnips for vim, forget IntelliJ. Are there better examples anywhere?

I can only speak about TabNine which appears to be very similar. It doesn’t just complete on APIs, it actually understands the code you’re writing. It will suggest variable names based on patterns, it will suggest what you initialize your variables to based on what they’re called. It will suggest comments based on the context. Sometimes you can just write the name of a function, and then tab through the suggestions line by line and it will write the whole procedure for you. It’s witchcraft.

1. IntelliJ also suggest variable names based on patterns.

2. Is there any good example of it suggesting good comments? AFAIK, comments are hard even for humans to write, so if they are able to generate good comments, it should be a major advance in NLP.

3. I don't believe this. Unless these are very trivial functions which are copy pasted across the internet, this is very hard. Is there any showcase or example where I can see this for myself?


I don’t know IntelliJ specifically. Maybe they do something similar that works as well.

It obviously doesn’t write a complete perfect comment for you, but I mentioned it because it’s not really specific to code, it’s just a really good and smart autocomplete that works on anything, even prose, and that can take unexpected patterns into account.

I probably wouldn’t believe it either, unless I had experienced it. Yes, it’s happened for simpler procedures, and you might need to tab through to pick the right completion - but still. It is unbelievably good at figuring out what type of code goes with which identifiers.


It's hard to believe until you try it, but it can absolutely autofill almost complete functions based on context and patterns learned from you've been typing in the project so far.

Think about how good GPT-2 is at writing essays given only a one-line prompt, now imagine that kind of predictive power applied to your code and you'll get something close to TabNine.


I haven't tried this, but I'd think that having auto completion in this way, is highly distracting. I usually start out with the basic idea of what I want to write and while writing I go back, alter, jump around, etc. If some AI starts suggestion whole chunks of code for me, I'd just waste time reading what the completions are and not thinking about the actual problem I am writing.

Sure if you are just re-implementing the same thing over and over, then this might come in handy, but so would using OOP.

Again, haven't tried, just some thoughts.


So, when this runs completely local, WHY do I have to pay monthly? Why can't it be bought?

And it costs more than my whole IDE (IntelliJ complete license for _everything_) as well...


I'm guessing if they gave you an all-in price you wouldn't want to pay what they ask anyway.

I am paying ~200 dollars per year for my IDE. What price would be justified for a auto-complete plugin for bespoken IDE?

It will be not for sale because they would not want to sell it.

Subscriptions are a tool that can both extract more money from customers and also increase the customer base by reducing the upfront cost.


This looks more like a problem that could have easily been solved with traditional algorithms, ranking keywords, etc.

I guarantee your opinion of AI-based autocomplete will change if you try it for a week. It's completely different than anything heuristics-based. In fact, it often suggests complete English phrases when I'm writing prose that are contextually relevant and grammatically valid (which makes sense, it uses GPT-2 under the hood I think). It's crazy how good it's gotten over the last few years.

I recommend TabNine but maybe kite is just as good these days, idk.


Your comment convinced me. I just downloaded the TabNine extension for VS Code and started using it for the work I'm doing (Python).

I tried it for about 10 minutes, but had to disable the extension. Two reasons:

1. My editor no longer brought up the useful function signature hints for library methods I wanted to call, which is really annoying.

2. There was a noticeable increase in lag from when I hit a key on my keyboard to when the corresponding character appears on screen. It was hard to get used to.

Have you noticed these issues? Would appreciate any tips you have for how to get over this - maybe just stick with it a while?

Thanks.


1. the priority order of popup hints is customizable in VSCode I believe. Unfortunately, I use sublime so I'm not sure of the exact fix in VS, maybe someone else can chime in with their experience here.

2. I haven't encountered this in SublimeText other than when I first open a large new project for about ~30sec, maybe it's trying to index a large project folder? I know you can set up restrictions on the folders it tries to index to make sure it's not going wild on your whole home folder by accident.

Check out https://www.tabnine.com/faq/#magic_strings and maybe try playing around with some of those config options?


Thanks. Will give it another go.

Think Kite is down, trying to download but getting service unavailable when I open up this page: https://alpha.kite.com/release/dls/mac/current

Thanks, we believe we fixed this. Could you try again?

Thanks for your patience!


Fixed now, much thanks

I can't put my finger on it, but this seems less like a helpful incremental improvement and more like sowing the seeds of our own destruction. Like it's obviously bad to only write code that you've seen other people write in circumstances you can recognize as similar, and it's bad to only write code nobody has ever written in history, and this just barely crosses the line.

very stoked on this! Thanks! Been using the python version for months and absolutely love it!

Wow! Excited to hear more about Kite in the news

Code extraction alert! Would not trust this company. Google them

Incredible stuff!

Any reason to use this over TabNine[1]? I’ve been using that for months, and only have good things to say about it. It’s like it’s reading your mind. Strong recommendation.

1: https://www.tabnine.com


Daniel from Kite here. TabNine is great! And at the end of the day, Kite's completions and TabNine serve the same purpose. However there are a couple of differences in how we designed Kite vs TabNine.

First, the most important consideration for us was if our completions model could provide a good user experience 1) locally and 2) with minimal CPU and memory usage footprint. While we're constantly tuning the product's performance, Kite already uses less memory and CPU than TabNine. We know how important it is to save resources (and battery life!) when you're developing, so this was a huge sticking point for us.

Second, we've also put a ton of thought into dealing with the visual noise that completions of this nature introduce. The completions that are predicted by models such as Kite's and TabNine's can sometimes be irrelevant or redundant, so it's important to design filters that reduce the amount of what we call "low value" completions shown so that developers are less distracted. In Python, we have a best-in-class static analysis engine that we use to filter out low value completions. We don't have our own proprietary JS static analysis engine yet, but we've done some UX "hacks" to make sure Kite's JS completions aren't redundant and don't crowd out other completions the user may find useful e.g. native completions from the IDE.

Third, (and this is a point mainly for Python), is that Kite is much more than just completions. We also provide help for calling functions, instant documentation and we have some more features in the works. Completions certainly help developers physically write code, but we know that "writing code" also includes a lot of reading, debugging, etc. So we're trying to build a product that helps out developers in all phases.


I just downloaded this yesterday and it’s already been insanely good. Really stoked to see more players and improvement in this space.

I agree that it's wonderful, it impresses me every day. When I saw kite I immediately thought about TabNine.

Does Kite's VSCode extension work with Github Codespaces?

You can now install vscode-server on a remote machine and connect to it over ssh from your local vscode or even from vscode in the browser (using Visual Studio CodeSpaces or the upcoming Github Codespaces)

Will Kite work with this??

Github CodeSpaces annoucement a few days ago: https://news.ycombinator.com/item?id=23092904


Kite currently doesn't work with remote coding in VS Code. That said, we're always open to integrating with more platforms, especially if there's strong enough demand from our users.

Off-topic maybe, but do you feel the any kind of lag with TabNine. I've dropped using it after a month as its slowed down my editor.

I've come back to TabNine after a hiatus and do see the occasional lag or the process thrashing at 100% of a core. Depending on mood I'll kill it or let it recover. I find myself forgiving any sporadic lags thanks to the productivity gains.

Dror from TabNine here. We've introduced some big improvements in TabNine's resource consumption in the last few weeks, and there's more to come. In addition, TabNine's Professional version (free trial available) lets you use our GPU cloud servers which means dramatically reduced load on your machine.

Are there any plans to allow using local GPU?

Could you share the OS and editor you're using? we're working to make TabNine super performant.

Thanks for the comment, I came back to TabNine after your previous comment, and I am glad that the issues (mostly performance) looks to be resolved, yet I need to play more with it.

Edit: OS: mac os 15.4.2 Editor: Vim and VScode


if you use WebStorm or IntelliJ, Codota (https://www.codota.com) is also very good

And, interestingly enough, Codota just acquired TabNine a few months ago :)

What? That's pretty massive. So Codota+TabNine got you covered on all languages & IDE's.

There's evidence that you guys have been astroturfing this thread. We ban accounts and sites that do that. Please stop now and never do this on HN again.

https://news.ycombinator.com/item?id=23159058


I hope my comments in the thread have not been interpreted as astroturfing. I have no affiliation with TabNine, I’m just a happy user.

I was talking about other accounts than yours. You're fine.

Having tried Kite a few years ago, and then TabNine, I ended up sticking with TabNine.

The thing that convinced me in the end was that Kite had a super heavy editor integration UX that needing a constantly-running Electron "copilot" app, even uploading some of my code without clearly asking for consent first. This has maybe changed now, but at the time that was my impression.

Meanwhile, TabNine has almost no UI, it's just the tab button and nothing else, and somehow the autocomplete "just worked" amazingly well with any language I wrote, including English prose. The ratio of "fuss" to "gain" was just so different between the two. One I barely remembered I had even installed anything, my editor just suddenly became a pair programmer overnight, and the other felt like I was running a whole electron app just to get occasional library headers autocompleted and docstrings on hover (in only a few supported languages).

I wish the Kite team all the best though, and I hope to see their solution grow and improve. There is definitely room for more players in this space, and I know they've learned many lessons from their past issues with code uploading UX, so I hope people give them a break over that past PR gaffe.


Thanks for the thoughtful message!

We hope you'll give Kite another shot. We're always improving it and have covered a lot of ground in the last few years.

In particular, a common complaint about TabNine on Twitter is that it's heavyweight. Based on our comparisons of today's Kite versus today's TabNine, we believe Kite uses less CPU/battery, and far less memory. (My teammate dhung posted a longer comment about TabNine.)


To clarify, my complaints about Kite being heavyweight are 90% about the UX flow being heavy, not the CPU/RAM usage.

Kite seems to require a background daemon, Electron copilot app, and editor plugin, with a multi-step setup to customize what things are shown on hover vs ctrl+space vs tab. By comparison, Tab Nine was a single command to install in SublimeText, and I have never had to interact with it since other than pressing "tab". It seems to be completely language-agnostic with no additional setup as well.

FWIW, here are a few pain points from trying Kite today:

- Kite is a 1GB download vs TabNine's 64mb

- Kite Engine is using 300MB of RAM on my machine right now, and I don't have a single editor open, and there's no way to stop the daemon in the UI

- "sending usage metrics" is on by default in the preferences :/ this is really scary to someone who remembers Kite's issues with code exfiltration and telemetry in the past, haven't ya'll learned that lesson? I had hoped all of that stuff was disabled these days but I guess not.

- there is no uninstall option provided, and you cannot delete the .app on macOS because "the application is in use", yet there is no UI to stop or diasble the daemon anywhere, do I really have to manually `pkill kite; rm -RF /Applications/Kite.app` to remove it?

- Your "Start free trial" link on the homepage appears to redirect to forms filled with real user email addresses? Who do those addresses belong to, are you leaking your users emails? [screenshot link redacted]


Response from Kite about the emails leaked:

> Thank you very much for calling this out. A test version of that webpage was temporarily deployed by mistake. The issue was fixed immediately and we’re in the process of notifying the specific users who were affected.

Welp, those were real user emails being leaked.


I've been using https://www.tabnine.com/ for similar AI/deep learning autocompletion, highly recommended.

TabNine has been absolutely life-changing for me, I spend most of my coding time these days just pressing tab.

I can't wait until they release a shell autocomplete solution, I imagine it'll blow the fish shell's (already amazing) autocomplete out of the water. (https://github.com/codota/TabNine/issues/171)


I bought TabNine back in 2018 used for a while and stopped, now I see it's a subscription service. Anyone still using the old version? Is it still working?

(Dror from TabNine here) If you're lucky enough to have bought the perpetual license for TabNine back in 2018, you don't have to purchase a subscription as all perpetual licenses are honored! Just reach for support@tabnine.com if you need help.

Wow, fantastic! Thank you very much!

I was going to mention this. TabNine supports more languages, how does Kite's product differ?

Really love TabNine [1]. Same thing, AI as an editor addon for super smart completions. Would be interested in knowing how Kite compare!

[1] https://www.tabnine.com/


So this is just code completion? What exactly makes it qualify as "artificial intelligence"? Did they throw some PyTorch or TensorFlow into the mix?

> We’ve trained a new deep learning model on 22 million open-source JavaScript files to ensure Kite works with your favorite libraries and frameworks like React, Vue, Angular, and Node.js.



Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: