noe casas

ML engineer at Apple. PhD in Neural Machine Translation. Background in ML, data science and SW engineering. Chinese lang enthusiast. Life-long learner. 学无止境

[MetaData.Cat @MetaDataCat]

La plataforma Langtern empra la IA per ajudar els professors a crear exercicis a partir de textos i vídeos en línia, que classifica per nivell de dificultat

https://t.co/HEj7lWfqiZ

[noe casas @noecasas]

You can find the slides and jupyter notebooks of my talk at PyDayBcn "Intro to neural nets with PyTorch" at https://t.co/Rc1uFtmyMP @PyBCN

[Marina Moro López @marinamorolopez]

Hoy es el #PyDayBCN 👩🏻‍💻🐍 De momento aquí estoy aprendiendo un poquito de PyTorch gracias a @noecasas 🤭

[noe casas @noecasas]

@erogol I was planning to get familiar with ggml at some point, but for now I have been just lurking and tracking its progress. There is already a TTS for ggml: https://t.co/J8n5Az2FMk . Maybe parts of it can be useful for a future "xtts.cpp"

[noe casas @noecasas]

@Azure will your prices reflect the same reduction on gpt-3.5-turbo? Will the new models be available via Azure Cognitive Services?

[OpenAI @OpenAI]

GPT-4 and GPT-3.5 Turbo models in the API now support calling your custom functions, allowing the model to use tools you design for it. Also — reduced pricing & new model versions (including 16k context for 3.5 Turbo): https://t.co/dalfgEQ9k2

[Python Barcelona @PyBCN]

@nataliasirera The PyDataBCN 2023 has been incredible!

We are grateful for all the organizers, volunteers, speakers, sponsors and assistants.

See you at our next event!

[Marta R. Costa-jussa @costajussamarta]

We are pleased to share a multilingual extension to holistic bias dataset which allows to unveil demographic biases for languages at scale, see our first findings: https://t.co/Az8rYZdAaQ

[noe casas @noecasas]

@Baidu_Inc @BaiduResearch 连普通的ERNIE接口还不可以正确使用。一位百度代表给我说明了我现在不能购买机器资源是因为：由于他们现在机器资源有限，为了保证用户体验，ERNIE 3.0暂时不售卖了。请增添机器资源来正确提供ERNIE的服务。

[noe casas @noecasas]

@gdb • Longer context
• Random seed parameter
• Tiktoken in JS/wasm to count tokens correctly (https://t.co/yhAObbmKZQ)
• Models tailored to specific languages, to use fewer tokens to represent the same text, e.g. for Chinese chars

[Langtern @langternapp]

New version of Langtern with improved flashcards!
• Auto-advance mode
• Can go back to the previous cards
• In Chinese, flashcards can play the word pronunciation
• Pronunciation played automatically in auto-advance mode

Great to review vocabulary while doing something else!

[noe casas @noecasas]

@ramsri_goutham @OpenAI To count tokens, I was previously relying on JS package gpt-3-encoder, recommended by OpenAI, which matches the tokenizer web tool. After switching to tiktoken, my estimation and the actually consumed tokens matched perfectly.

[noe casas @noecasas]

@ramsri_goutham @OpenAI I found out at https://t.co/wxoD3UHw6R after seeing large discrepancies in the actually consumed tokens and my estimation

[noe casas @noecasas]

The situation has improved a bit with the new cl100k_base tokenizer used in the new gpt-3.5-turbo models

[noe casas @noecasas]

@gdb: is in @OpenAI 's plans to mitigate this issue? (e.g. offering models that are trained on vocabulary that represents other scripts more efficiently). My case is with simplified Chinese, which has approx num tokens = 2 x num chars

[Ramsri Goutham Golla @ramsri_goutham]

Most people don't understand @OpenAI GPT-3's tokenization and how expensive/inefficient it is to build a GPT-3 app in non-English!

English: John - 1 token
Telugu: స్తి - 12 tokens 🤯

Byte pair encoding divides a character like స్తి further!

స్తి (12 tokens)= స్త + ి (8+4)

[Langtern @langternapp]

Do not hesitate to drop us a line at [email protected] if you have any doubts or feedback to share! 6/6

[Langtern @langternapp]

You can also extract the whole vocabulary (by HSK level) of a video or an essay to preview it, and save it as flashcards to review later. 5/6

[Langtern @langternapp]

And the new main feature: the online content search engine, where you can search videos and essays by keywords/content but also by HSK level and duration/length! Check it out! 4/6

[Langtern @langternapp]

In the Online Content, we have also added social media accounts in Chinese from Twitter and Weibo, web novels, short stories and podcasts with transcription. 3/6

[Langtern @langternapp]

The Online Content section now has more videos in Chinese from Youtube and now also from Bilibili, organized by category and channel. With an improved popup dictionary for captions! 2/6