I posted some of my experience with Kagi’s LLM features a few months ago here: https://literature.cafe/comment/6674957 . TL;DR: the summarizer and document discussion is fantastic, because it does not hallucinate. The search integration is as good as anyone else’s, but still nothing to write home about.

The Kagi assistant isn’t new, by the way; I’ve been using it for almost a year now. It’s now out of beta and has an improved UI, but the core functionality seems mostly the same.

As far as actual search goes, I don’t find it especially useful. It’s better than Bing Chat or whatever they call it now because it hallucinates less, but the core concept still needs work. It basically takes a few search results and feeds them into the LLM for a summary. That’s not useless, but it’s certainly not a game-changer. I typically want to check its references anyway, so it doesn’t really save me time in practice.

Kagi’s search is primarily not LLM-based and I still find the results and features to be worth the price, after being increasingly frustrated with Google’s decay in recent years. I subscribed to the “Ultimate” Kagi plan specifically because I wanted access to all the premium language models, since subscribing to either ChatGPT or Claude would cost about the same as Kagi, while Kagi gives me access to both (plus Mistral and Gemini). So if you’re interested in playing around with the latest premium models, I still think Kagi’s Ultimate plan is a good deal.

That said, I’ve been disappointed with the development of LLMs this year across the board, and I’m not convinced any of them are worth the money at this point. This isn’t so much a problem with Kagi as it is with all the LLM vendors. The models have gotten significantly worse for my use cases compared to last year, and I don’t quite understand why; I guess they are optimizing for benchmarks that simply don’t align with my needs. I had great success getting zsh or Python one-liners last year, for example, whereas now it always seems to give me wrong or incomplete answers.

My biggest piece of advice when dealing with any LLM-based tools, including Kagi’s, is: don’t use it for anything you’re not able to validate and correct on your own. It’s just a time-saver, not a substitute for your own skills and knowledge.

I recently upgraded to a 7900 XTX on Debian stable, as well. I’m running the newest kernel from Debian’s backports repo (6.6, I think), and I didn’t have that same problem.

I did have other problems with OpenCL, though. I made a thread about this and solved it with some trouble. Check my post history if you’re interested. I hope it helps. I can take a closer look at my now-working system for comparison if you have further issues.

IT WORKS NOW! I will need time to run additional tests, but the gist of my solution was:

1. Backport llvm-18 from sid following the guide you linked at https://wiki.debian.org/SimpleBackportCreation

2. After compiling and installing all those deb files, I then installed the “jammy” version of amdgpu-install_6.0.60002-1.deb from https://www.amd.com/en/support/linux-drivers

3. Downloaded the latest kernel sources from https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git, and simply copied all the files from its lib/firmware/amdgpu folder into my system’s /lib/firmware/amdgpu. Got that idea from https://discussion.fedoraproject.org/t/amdgpu-doesnt-seem-to-function-with-navi-31-rx-7900-xtx/72647

4. sudo update-initramfs -u && sudo reboot

I’m not totally sure it step 3 was sane or necessary. Perhaps the missing piece before that was that I needed to manually update my initramfs? I’ve tried like a million things at this point and my system is dirty, so I will probably roll back to my snapshot from before all of this and attempt to re-do it with the minimal steps, when I have time.

Anyway, I was able to run a real-world OpenCL benchmark, and it’s crazy-fast compared to my old GTX 1080. Actually a bigger difference than I expected. Like 6x.

THANKS FOR THE HELP!

Thanks for the links! I’ve never attempted making my own backport before. I’ll give it a shot. I might also try re-upgrading to sid to see if I can wrangle it a little differently. Maybe I don’t actually need mesa-opencl-ics if I’m installing AMD’s installer afterwards anyway. At least, I found something to that effect in a different but similar discussion.

Update: I upgraded to Sid. Unfortunately, mesa-opencl-icd depends on libclc-17, which uninstalls -18. So I can’t get OpenCL working while the correct libclc is installed.

No idea where to go from here. I’ll probably restore my Bookworm snapshot, since I don’t want to be on Sid if it doesn’t solve this problem.

Update: Running amdgpu-install did not provide those files. There were a few errors regarding vulkan packages when I attempted, I guess because it’s assuming Ubuntu repos. Trying with just opencl and not vulkan succeded, but still clinfo reported the missing files.

I don’t think I can get this working without a whole newer llvm.

Ah, somehow I didn’t see 18 there and only looked at 17. Thanks!

I tried pulling just the one package from the sid repo, but that created a cascade of dependencies, including all of llvm. I was able to get those files installed but not able to get clinfo to succeed. I also tried installing llvm-19 from the repo at https://apt.llvm.org/, with similar results. clinfo didn’t throw the fatal errors anymore, but it didn’t work, either. It still reported Number of devices 0 and OpenCL-based tools crashed anyway. Not with the same error, but with something generic about not finding a device or possibly having corrupt drivers.

Should I bite the bullet and do a full ugprade to sid, or is there some way to this more precisely that won’t muck up Bookworm?

Can you explain more about your workflow? Do the Nix packages have their own isolated dependency resolution? How does it work when Debian packages depend on a library you get from Nix, or vice-versa?

Thanks, that’s good advice. There are lower-numbered gfx* files in there. 900, 902, 904, 906. No 1030 or 1100. Same after reinstalling.

Looks like these files are actually provided by the libclc-15 package. libclc-16 has the same set of files. Even libclc-17 from sid has the same files. So I guess upgrading to testing/unstable wouldn’t help.

apt-file search gfx1100-amdgcn-mesa-mesa3d.bc yields no results, so I guess I need to go outside of the Debian repos. I’ll try the AMD package tonight.

Thanks! I didn’t see that. Relevant bit for convenience:

we call model providers on your behalf so your personal information (for example, IP address) is not exposed to them. In addition, we have agreements in place with all model providers that further limit how they can use data from these anonymous requests that includes not using Prompts and Outputs to develop or improve their models as well as deleting all information received within 30 days.

Pretty standard stuff for such services in my experience.

If you click the Chat button on a DDG search page, it says:

DuckDuckGo AI Chat is a private AI-powered chat service that currently supports OpenAI’s GPT-3.5 and Anthropic’s Claude chat models.

So at minimum they are sharing data with one additional third party, either OpenAI or Anthropic depending on which model you choose.

OpenAI and Anthropic have similar terms and conditions for enterprise customers. They are not completely transparent and any given enterprise could have their own custom license terms, but my understanding is that they generally will not store queries or use them for training purposes. You’d better seek clarification from DDG. I was not able to find information on this in DDG’s privacy policy.

Obviously, this is not legal advice, and I do not speak for any of these companies. This is just my understanding based on the last time I looked over the OpenAI and Anthropic privacy policies, which was a few months ago.

Yeah, I wouldn’t be too confident in Facebook’s implementation, and I certainly don’t believe that their interests are aligned with their users’.

That said, it seems like we’re reaching a turning point for big tech, where having access to private user data becomes more of a liability than an asset. Having access to the data means that they will be required by law to provide that data to governments in various circumstances. They might have other legal obligations in how they handle, store, and process that data. All of this comes with costs in terms of person-hours and infrastructure. Google specifically cited this is a reason they are moving Android location history on-device; they don’t want to deal with law enforcement constantly asking them to spy on people. It’s not because they give a shit about user privacy; it’s because they’re tired of providing law enforcement with free labor.

I suspect it also helps them comply with some of the recent privacy protection laws in the EU, though I’m not 100% sure on that. Again, this is a liability issue for them, not a user-privacy issue.

Also, how much valuable information were they getting from private messages in the first place? Considering how much people willingly put out in the open, and how much can be inferred simply by the metadata they still have access to (e.g. the social graph), it seems likely that the actual message data was largely redundant or superfluous. Facebook is certainly in position to measure this objectively.

The social graph is powerful, and if you really care about privacy, you need to worry about it. If you’re a journalist, whistleblower, or political dissident, you absolutely do not want Facebook (and by extension governments) to know who you talk you or when. It doesn’t matter if they don’t know what you’re saying; the association alone is enough to blow your cover.

The metadata problem is common to a lot of platforms. Even Signal cannot use E2EE for metadata; they need to know who you’re communicating with in order to deliver your messages to them. Signal doesn’t retain that metadata, but ultimately you need to take their word on that.