Right now, we are providing a free alternative to Copilot and expanding the set of features we support. We are actively working on improving the quality of suggestions and would appreciate feedback.
We guarantee that all users that join now will be able to use the product free forever. We explicitly allow users to opt out of sending data to us post inference. We believe we're in the early innings of code assistants and would like to democratize the technology so we don't have plans to monetize this tier of usage.
The model is hosted internally and trained on public code. For more details on the specifics, please take a look at our FAQ page on the website.
Thanks! We've made a lot of optimizations internally for latency. We've gotten a lot of feedback that it is faster than Copilot but obviously they operate at much larger scale.
I'm definitely for any alternative to an MS solution (especially one where I have to give them actual money to use it), but I had kind of the same experience. Not extensive testing, and I'm using it in PyCharm, so I don't know if that changes the behavior.
The completion was weird. I tried to type "#Function to return the value passed". And it started to autocomplete with what I typed and what it was suggesting all jumbled up and completely unintelligible. Once I hit tab, it completed correctly, but it didn't make any sense until I hit tab, so I couldn't really see what I was completing until it was done.
The more I played with it, I noticed it started to not actually suggest anything. But the codeium plugin is just showing a wait spinner in my status bar, so maybe it's getting pounded with HN people testing it right now.
It looks like after 3 more minutes of testing, it's a me problem. There is one particular file that goes bonkers, which just so happened to be the one I had up. If I start a new file in that project, or work on an existing file in a different project it's fine and the completion works as expected.
I can't make it work with the one file, but yeah...not a codeium problem at all. I'll be sure to test at least 5 minutes next time ;)
I am dubious to use this at work since I can’t be sure it won’t send code from the codebase I am working on to the cloud. When something is free you’re usually the product.
We understand the concern! We describe our security policies in detail on our website, but we offer users the option to opt out from code snippet telemetry. This means after the inference, we do not store any code-based information on our end. We also know that you might not take our word for it, which is why we are working on getting SOC2 attestation for these policies.
Also, to answer your pricing point, we are actively working on additional features beyond code completion. We want to try to make the code completion itself to be free, and definitely free forever for early users as a thank you for using Codeium from early on.
I'm really really curious as to the technical details of getting the inferences cheap and fast enough to enable a copilot like experience when you're not GitHub with inifinity resources.
Sam Altman has said that each chat in chatgpt is single digit cents per chat. GPT3 davinci is 0.02 per 1k tokens, not to mention the high latency. it'd be impossible to build "copilot for law" based on gpt3 because the UX wouldnt work with that latency or cost (you'd be running 100s of inferences if you count every debounced keystroke).
What cost per user/per inference/per day (however you think about it) do you need to get this down to? and what order of magnitude speedups did you need to accomplish to get this done?
I have a ML infra that does between 1 million and 2 million inferences per day (granted it is not a chat bot or code generator, but it is legit ML inference in a similar mode) and it only cost $120 per month for the server (also runs multiple containers and is a web server), which mostly sits idle.
Hey swyx :) Great question, we've got a blog post coming soon with some of these technical and other details... we've employed a lot of tricks here, debouncing included.
Been using this for the past couple of weeks and can attest it's been game changer for my productivity. Was skeptical at first, but I find myself less on Google search and more in flow. Thanks for giving it away for free
Yeah I used Copilot for a few weeks before since it was better known. Ended up cancelling my Copilot because I didn't notice a huge difference in quality/features and Codeium is free (primary motivation tbh)
Cool. Well I guess until there's some community-driven open-source alternative I'll be rooting for y'all against Copilot and CodeWhisperer. Maybe you can pick up some of the Kite talent to join the effort
Has been a huge boost over using Copilot. I accidentally was using Copilot instead of Codeium and was confused why the generations took so long until I realized! Great product
I briefly tried it and it was indeed very fast, but I couldn't get simple things for python. It would just keep adding import statements or try to continue writing my comment (even though I wanted the code that matches the comment).
Also I couldn't find any way to toggle alternative suggestions.
If you start the function you're trying to write (just by writing `def` on a new line), you should usually be able to get it to start implementing it.
You can see alternate suggestions on VSCode by hitting Option + [. We don't currently support seeing alternate suggestions on JetBrains IDEs, but we're working on it!
We have gotten requests for more IDEs, and are actively working on creating support for them so that as many developers as possible can leverage this technology!
In my exactly 3 minutes test, it did worse, even if only slightly.
Right now it is free, but that won't last, right? Do i have to expect any shady data collection practices?
More interestingly, how is this trained? Have you guys done your own finetuning of gpt3 or it's something completely different?
We guarantee that all users that join now will be able to use the product free forever. We explicitly allow users to opt out of sending data to us post inference. We believe we're in the early innings of code assistants and would like to democratize the technology so we don't have plans to monetize this tier of usage.
The model is hosted internally and trained on public code. For more details on the specifics, please take a look at our FAQ page on the website.
I have to say, in my quick tests completions do arrive slightly but noticeably faster than Copilot's.. Which is enough to make this interesting!
Unless you are using a much smaller model, in which case it's too easy ahah.
Thanks and good luck for your further plans, I'll keep trying this
The completion was weird. I tried to type "#Function to return the value passed". And it started to autocomplete with what I typed and what it was suggesting all jumbled up and completely unintelligible. Once I hit tab, it completed correctly, but it didn't make any sense until I hit tab, so I couldn't really see what I was completing until it was done.
The more I played with it, I noticed it started to not actually suggest anything. But the codeium plugin is just showing a wait spinner in my status bar, so maybe it's getting pounded with HN people testing it right now.
I can't make it work with the one file, but yeah...not a codeium problem at all. I'll be sure to test at least 5 minutes next time ;)
What exactly does this mean?
Also, to answer your pricing point, we are actively working on additional features beyond code completion. We want to try to make the code completion itself to be free, and definitely free forever for early users as a thank you for using Codeium from early on.
I'm really really curious as to the technical details of getting the inferences cheap and fast enough to enable a copilot like experience when you're not GitHub with inifinity resources.
Sam Altman has said that each chat in chatgpt is single digit cents per chat. GPT3 davinci is 0.02 per 1k tokens, not to mention the high latency. it'd be impossible to build "copilot for law" based on gpt3 because the UX wouldnt work with that latency or cost (you'd be running 100s of inferences if you count every debounced keystroke).
What cost per user/per inference/per day (however you think about it) do you need to get this down to? and what order of magnitude speedups did you need to accomplish to get this done?
(congrats btw, happy user here thanks to Prem)
EDIT: clicking on the "return to vs code" link from your site was needed. now auth'd
Also I couldn't find any way to toggle alternative suggestions.
You can see alternate suggestions on VSCode by hitting Option + [. We don't currently support seeing alternate suggestions on JetBrains IDEs, but we're working on it!
We will have updates on this shortly!