tootyskooty

I'm working on Repple (https://repple.sh)! It's a modern spaced repetition x incremental reading/PDF library app with a few (tasteful!) QOL AI features.

I've been using Anki for 10+ years and love it but always wanted something with a cleaner UX and a reader view. The recent Anki ownership change pushed me to finally make something, and it's seeing some traction :)

Right now I'm focusing on getting the reading and note-taking view to be nice. I used to use Polar Bookshelf (RIP) but that went away, trying to make something better.

The flashcard side also has a REST API btw!

I'm building Repple (https://repple.sh), a free flashcard app with spaced repetition scheduling and some (tasteful imo) AI additions. I've been using Anki for 10+ years and love it but always wanted something with a cleaner UI and that's a bit easier to experiment with. The recent Anki ownership change pushed me to finally make something.

Added a REST API (https://repple.sh/developers) a few weeks ago so you can build on top of it. Decks, cards, reviews, etc.

Feature delta over Anki:

- Tab-autocomplete for text fields

- Automatic image-gen for image fields

- Optional rephrasing that changes wording each review to avoid pattern matching

- Basic PDF library & incremental reading support

- "Orphan" card detection; i.e. knowledge that isn't well connected

- ... + a bunch of other qol improvements like semantic search, etc.

I'm wondering about this too. Would be nice to see an ablation here, or at least see some analysis on the reasoning traces.

It definitely doesn't wipe its internal knowledge of Crystal clean (that's not how LLMs work). My guess is that it slightly encourages the model to explore more and second-guess it's likely very-strong Crystal game knowledge but that's about it.

See pretraining section of the prerelease_notes.md:

https://github.com/DGoettlich/history-llms/blob/main/ranke-4...

Since it now includes 4 thinking levels (minimal-high) I'd really appreciate if we got some benchmarks across the whole sweep (and not just what's presumably high).

Flash is meant to be a model for lower cost, latency-sensitive tasks. Long thinking times will both make TTFT >> 10s (often unacceptable) and also won't really be that cheap?

both have had questionable content for a while, it's a wonder people are still paying for them. especially given that LLMs exist (and youtube for that matter).

Since no one has mentioned it yet: note that the benchmarks for large are for the base model, not for the instruct model available in the API.

Most likely reason is that the instruct model underperforms compared to the open competition (even among non-reasoners like Kimi K2).

Shameless plug: if OP is looking to stay on d3, he could also try slotting in my C++/WASM versions[1] of the main d3 many-body forces. Not the best, but I've found >3x speedup using these for periplus.app :)

[^1]: https://www.npmjs.com/package/d3-manybody-wasm

Working on a new interface for learning with LLMs that creates courses on any topic.

https://periplus.app

The goal was to make the learning material very malleable, so all content can be viewed through different "lenses" (e.g. made simpler, more thorough, from first principles, etc.). A bit like Wikipedia it also allows for infinite depth/rabbit holing. Each document links to other documents, which link to other documents (...).

I'm also currently in the middle of adding interactive visualizations which actually work better than expected! Some demos:

https://x.com/mato_gudelj/status/1975547148012777742

I've been getting a lot of vulnerability "spam mail" recently that's clearly AI-generated.

It's a surprise every public bounty program isn't completely buried in automatic reports by now, but it likely won't take long.

would be nice to finally see multi-turn coding benchmarks. everything we have so far is single-turn and that's clearly not a realistic scenario.

Hey, sorry for seeing this late!

You can send me a message on discord (dcbcdefb), or email (support (at) periplus dot app).

Also available on Twitter!

I suspect one can go a lot further by adopting some tweaks from the GPT-2 speedrun effort [0], at minimum Muon, better init and carefully tuning learning rate.

[0] https://github.com/KellerJordan/modded-nanogpt

Thanks a lot! I did do some of these things (namely Reddit) and that worked well, just the number of places that allow posting is limited and I don't want to get too spammy. Will continue there.

Main conceptual issue I've been having with other marketing (e.g. influencer) is that there isn't a well-defined audience to market this to. Usually edtech targets students/schools and Periplus doesn't fit there too well. Need to find what works I guess.

I'll spend more time on it from now on :). Thanks again.

re: mobile playback controls -> on it

I have no idea how to market this honestly. Most of my users came from a single discord server, but the feedback is overall pretty positive.

If you have any tips I'd be super grateful (gave you a follow on X).

Just added a proper playback control component on desktop, allows changing rate, rewinding & persists across pages :)!

Yeah it's still hard to deal with LLM gaps (fwiw Study mode would also be prone to this). I do try to catch the super obvious stuff and put up a disclaimer but it's far from perfect.

I had some prototypes basing the generations in websearch but the APIs are still super expensive on that front + the models tend to overindex on the results.

Glad you like it!!

I've been working on it on-and-off for about a year now. Roughly 2-3 months if I worked on it full-time I'm guessing.

re: playback speed -> noted, will add some controls tomorrow

There are both in-document quizzes and larger exams (at a course level).

I've also been playing around with adapting content based on their results (e.g. proactively nudging complexity up/down) but haven't gotten it to a good place yet.

I gave it a shot with periplus.app :). Not perfect by any means, but it's a different UX than chat so you might find it interesting.

Honestly thought they would take this a bit further, there is only so much you can do with a prompt and chat. It seems fine for surface level bite-sized learning, but I can't see it work that well for covering whole topics end to end.

The main issue is that chats are just bad UX for long form learning. You can't go back to a chat easily, or extend it in arbitrary directions, or easily integrate images, flashcards, etc etc.

I worked on this exact issue for Periplus and instead landed on something akin to a generative personal learning Wikipedia. Structure through courses, exploration through links, embedded quizzes, etc etc. Chat is on the side for interactions that do benefit from it.

Link: periplus.app

Still working on https://periplus.app!

It's an environment for open-ended learning with LLMs. Something like a personalized, generative Wikipedia. Has generated courses, documents, exams, flashcards, maps and more!

Each document links to more documents, which are all stored in a graph you grow over time (very Obsidian-esque).

No, you don't need to try to keep up with new tools. I would recommend you try the models though, even for a short time every few months. Send them questions or things you're working on, and see how they do. Provide sufficient context.

It's a good approximation to say that all tools are thin wrappers on top of the models, and having a good grasp of what the models can/can't do right now gets you 80% of the way there.

Before LLMs (not hard-set order): IDE/interface -> Stack Overflow -> Docs -> Library code -> Github

LLMs now slot in first or second, typically completely eliminating SO. Others still provide value.

One underdiscussed advantage is that an LLM makes knowledge language agnostic.

While less obvious to people that primarily consume en.wiki (as most things are well covered in English), for many other languages even well-understood concepts often have poor pages. But even the English wiki has large gaps that are otherwise covered in other languages (people and places, mostly).

LLMs get you the union of all of this, in turn viewable through arbitrary language "lenses".

Google will also have good results to report for this year's IMO, OpenAI just beat them to the announcement

The conclusion is that research takes time to productize, and this is cutting-edge research. OAI employees stated that there isn't anything math-specific (think AlphaGeometry) about this model. It's a general system.

Honestly might be more indicative of how far behind vision is than anything.

Despite the fact that CV was the first real deep learning breakthrough VLMs have been really disappointing. I'm guessing it's in part due to basic interleaved web text+image next token prediction being a weak signal to develop good image reasoning.

I'm assuming he means the "generate an image and order 500 stickers" one.

Thanks, I didn't know about Photoprism! My phone seems to do similar auto-tagging but I found it to not be flexible enough.

Honestly might do a CLIP powered version myself. I only need image ("similar picture") and language search, doesn't seem that difficult.