NLP/backend engineer, sometimes JA/EN interpreter
@Oracle←@mantra_ja←@UTokyo_IIS←@UW←@eBay
I like languages , combat sports and the outdoors
日本語も一応話せます
Recto — a truly 2D language
https://t.co/ipiOM3ngmm
What if your code lived in two dimensions instead of a flat stream of text? Recto is a 2D programming language that uses nested rectangles to encode structure and recursion directly in space.
Happy to report that the paper I've been working on with some friends at @NAIST_MAIN_EN was accepted to ACL main, proving that at least a few people still care about MWE research.
Consumer facing open source has gotten really good... needed screensharing software, file deduping software, etc. to help an aging family member organize their computer and the OSS solutions (@rustdesk, czkawka) were as good or better than the commercial offerings.
I just tried this after hearing about it on @latentspacepod, and it's actually pretty incredible.
Feels an awful lot like the future of search, or at least certain types of search.
Looked for NLP folks focused on language learning and found myself, which was flattering 😂
@bclavie@bclavie Thanks for the sanity check - I was assuming the model was optimized for longer sequences and newer GPUs, and I have neither, so this makes sense.
You made me realize I'd skipped over unpadding in the blog post though, which I went back and read properly. Cool stuff.
ModernBERT is too modern for me apparently, I'm getting worse GPU utilization than normal BERT (with identical hyperparameters otherwise).
Just realized I'm missing a flash attention install though, so that might be part of the issue...
In testament to my ability to procrastinate doing more important work, I've produced another blog post about weird NLP problems. This time it's compound word generation for artificial languages.
https://t.co/ic73CatmJO
(日本語版もあります)
@adamnohejl Yeah, something to talk through in discord, but I do think we should have thought more about presentation, which we could do while still being honest. "IAA is only 37%" by itself doesn't look great, but readers won't notice we computed it differently unless we tell them.
What is the bar for calling something an agent? If simply having multiple prompts used in an application means that each of these is an agent, then I don't think I've ever seen an LLM-based system that wasn't "agentic"
Someone found the input method I wrote for myself a while back on Github, and has started contributing PRs to it. At first it was mostly minor stuff, but they just fixed a pretty substantial bug... open source is cool sometimes.
https://t.co/wq0JY3ymqE
Another week, another speech model from Kotoba! We're open-sourcing kotoba-whisper-bilingual, a super-fast, end-to-end Japanese↔English speech translation model! The project is led by @asahiushio. Feedback is appreciated, and up next is our speech GenAI API!
Suddenly regretting not publishing anything to #ACL2024 ... probably not going to get another chance to see my favorite field of research and my favorite sport in the same place.
Happy to announce that I'm starting a new role as a Principal Applied Scientist at @Oracle Health Data Intelligence as of today.
Looking forward to exploring the world of medical NLP!