Suggestion to Damien: Migrating to an NN (Neural Net)

Anki Suggestions
61

Update: I will no longer be working on Clanki and RWKV-mini. My Claude Code account randomly got banned and when I requested further investigation, I got this:

image-4
image-4535×251 13.2 KB

I could maaaaaybe work on RWKV, granted it would be very slow and more limited, but there is absolutely no way to maintain an entire Anki fork alone, on my own.

Ngl, I don’t feel great right now.
And I’m not rushing to try Codex or other alternatives because the country where I live is blocked by every LLM provider. For Claude Code I was using a burner phone number + a card with a billing address that has nothing to do with where I live (different country), which is probably why I got banned. Probably. They never specified what exactly I violated.

I’ll keep using my Clanki build, at least until FSRS-7 is integrated into Anki natively. Then I’ll probably switch to official Anki. Maybe someone will look at the history of my Clanki fork (aside from UI, it’s mostly performance optimizations) and use it as a starting point for Anki speedups, but the chances of that are very slim.

EDIT: some of my Clanki speedups are being ported to Anki:

62

I haven’t tried it yet but OpenCode seems to be a popular workaround. Since it’s open source, it can be used from any country, users can use the API of their preferred AI model, and local LLMs can also be used if they have a GPU.

63

I’m now using OpenHands (open source harness, kinda like Claude Code) and GPT-6 Luna (best in terms of price-performance) in Flex API mode (higher time-to-first-token aka latency, lower throughput tokenss, and worse server uptime, but 2x lower cost) at “medium” reasoning effort.
So I’m using a poverty LLM in poverty mode at poverty reasoning effort :face_exhaling:
I’ll probably end up spending 100-200$ per month, comparable to my Claude subscription, except with much lower level of intelligence (8-12 months behind the frontier). Well, better than nothing, I guess.
I’ll keep working on RWKV-mini and Clanki.

EDIT: OpenHands is…not great, to put it politely. I’ll look for other open-source alternatives. But yeah, the overall plan is financially workable, the question at this point is whether I can find an open-source harness that isn’t a buggy mess, and, of course, quality of code written by behind-the-frontier models.

EDIT 2: OpenCode straight up doesn’t work for me. Yeah, finding a good open-source harness is going to be a problem. My last couple of updates have been oscillating between “We’re so back” and “It’s so over”.

image
image1192×900 107 KB

EDIT 3: guess I’m forking OpenHands

64

Re: RWKV development (not so sure about Clanki though):

I don’t know the nature of your work, but if you can make good progress via async “batch” style development, I can donate Astra / Astra Pro use to your efforts. You’d need to give me what needs to be run and I can provide the outputs. So obviously this doesn’t work for tight work loops but perhaps there’s something I can contribute that augments what you do with the “poverty” models meanwhile…

An easy way to do this for me would be via a GitHub pull request. You can comment with what you want done next, comment on lines, and upload any artifacts (as long as it’s not like arbitrary binaries), and I can update the PR with the result.

65

I have decided to stop working on RWKV and Clanki. OpenHands is a buggy mess and GPT-6 Luna is not intelligent enough. I’ll try again in 6-12 months, perhaps by then OpenHands will be less buggy and cheap models will catch up to frontier models of today.

In the meantime JSchoreels (Jonathan Schoreels) · GitHub will continue my work on RWKV and will integrate it into his own fork of Anki.

EDIT:
There’s a lot of perf improvement PRs based on my Clanki:

Though they are relatively minor. The big ones were

and

EDIT2: ok, I found a decent open-source harness (Goose), but GPT-6 Luna isn’t great, so progress will be slow