Some time ago, I wrote an article showing that AI tools for coding in Julia weren’t that good. Reading a post yesterday, someone mentioned Ask AI from JuliaHub, which specializes in Julia coding. I noticed that it works really well, after I asked a couple of questions (at least better than the other tools).
I was surprised that it wasn’t mentioned anywhere (or at least I only read that in a comment from a post here). In fact, if you search “Ask AI” here on Discurse, there are no results about it. So hopefully, this will make it more noticeable.
One questions for the maintainers of JuliaHub. Is it completely free once you’re registered? I read that there’s a cap of 20 hours per month, but don’t know if it was for computations or it was also applying to Ask AI.
There is a blog post from JuliaHub about it, but I only know about that because someone (outside of JuliaHub the company) posted about this feature on Slack, and this blog post came up in the replies. (It was posted on Zulip too in the same way, by someone outside JuliaHub.)
I believe the 20 hour limit is only for the compute, the AI has no limit afaik. The blog post says:
Login to JuliaHub to ask your Julia questions on JuliaHub now by going to: https://juliahub.com/ui/AskAI. You can sign up with your GitHub, GitLab, Google, or Linkedin account - no credit card is required!
The “no credit card is required” heavily implies there’s no charge for this.
The feature is still marked as “beta”, so maybe they plan to do the announcement once it’s considered ready for release. In my experience, it performs about on par with ChatGPT-4 purely in terms of model output, but the fact that the Phind interface gives source links on the side to follow through and confirm the information makes it much better for me. That ability to verify sources is especially important because of how often these models hallucinate and give false information.
So if the idea is to improve the model and the interface before announcing it publicly, I’m fully behind that, and glad they’re doing it that way. Right now, these tools are prone to missing relevant good packages and instead suggesting outdated or non-existent packages, hallucinating functions that don’t exist, etc. Some degree of that is to be expected with LLMs, but it would be nice if that was a 10% chance of misinformation rather than the current 40-50% when it comes to Julia questions.