Mark Woodward

Data Scientist at SIL International | Machine Learning | NLP for low-resource languages.

[Jeremy Howard @jeremyphoward]

For those that hope (or worry) that LLMs will do breakthrough scientific research, I've got good (or bad) news:

LLMs are particularly, exceedingly, marvellously ill-suited to this task. (if you're a researcher, you'll have noticed this already)

Here's why🧵

[Towards Data Science @TDataScience]

In a new deep dive, Rahul Nayak walks us through the process of building an AI research agent capable of answering questions about the intricate details within the Mahabharata, the longest epic poem ever composed. https://t.co/abq3FP9XEu

[Mark Woodward @DataScienceMW]

Realizing (again) how much more useful this app is if I switch to the Following tab. Posts from people I've chosen to follow, which only have a handful of likes and comments, are generally way more informative than posts with 10k likes which have been optimized for the algorithm.

[Mark Woodward @DataScienceMW]

@s_gruppetta_ct I hope you returned home with a KeyError

[Adam Azzam @AAAzzam]

You can do constrained sampling in ChatGPT and micromanage its output. The catch? You only get one token. But you can do a lot with that one token.

You can make it a classifier, a logic gate, or have it choose tools deductively. Here's how 🧵

From my talk this week @trychroma

[Mark Woodward @DataScienceMW]

@mr_nemecek @ylecun @dwhitena Looks like the arxiv page is missing a comma after @dwhitena's name, which then concatenates him with the next author... https://t.co/4IUtnk4IMG

[Yann LeCun @ylecun]

MMS: Massively Multilingual Speech.
- Can do speech2text and text speech in 1100 languages.
- Can recognize 4000 spoken languages.
- Code and models available under the CC-BY-NC 4.0 license.
- half the word error rate of Whisper.

Code+Models: https://t.co/NIGfUZ8KZg
Paper: https://t.co/W15aEWHGIR
Blog: https://t.co/TFKXFtlPwc

[Mark Woodward @DataScienceMW]

Your documents may only be in one language (e.g. English), but you now have a chatbot where users can interact with it in any of the (50+?) languages that are supported by Google translate, and that ChatGPT knows.

[Mark Woodward @DataScienceMW]

Everyone is talking about creating content-specific chatbots, by doing semantic search over documents and feeding the matches as context into ChatGPT (or similar).

But I haven't heard much talk about how adding one more step can make the chatbot multilingual in 50+ languages.👇

[Mark Woodward @DataScienceMW]

Google translate can also tell you the language of the user query.

You then feed this original-language context into ChatGPT, with an additional prompt telling it to respond in <query_language>.

[Cristiano Giardina @CrisGiardina]

Tired of waiting for GPT-4-32k?

Bing Chat Creative Mode uses a GPT-4 model with a context window of ~13,500 tokens!

Here's how you can use it & how I found this out!

[Mark Woodward @DataScienceMW]

Now I've received an email saying my order has been canceled. When I chat with @Xfinity online it tells me to call, when I call it tells me to chat, so I chat and it tells me to call, and I call and get in a loop of endless options, none of which are what I want! 🤦‍♂️

[Mark Woodward @DataScienceMW]

We're moving house next week, so I've just signed up for @Xfinity internet. Not at all impressed by their website, which was super-slow, and the chat person wasn't helpful.

I think this final message sums it up.

"Thanks for your purchase, null."

[Sahil Patel @saaaahiiiil]

I'm currently pursuing an MA in Linguistics from the University of Mumbai. We have a mandatory year long field linguistics course where every year, we try to document either an endangered or a lesser studied language. A few things I learnt from this course, a 🧵

[Mark Woodward @DataScienceMW]

I have to say that a month of working on a project with multiple colleagues and many simultaneous branches, issues, PRs, etc, has helped me to feel a whole lot more comfortable than I was before. But I'm still never quite sure when it's going to surprise me...

[Mark Woodward @DataScienceMW]

Every time I start thinking I've gotten the hang of git, it surprises me and I end up in a tangle of merge conflicts.

Is it just me? Does anyone ever get to the stage where they feel totally comfortable with git?

[Mark Woodward @DataScienceMW]

My Saturday afternoon project was going to be attempting to run Stable Diffusion via @modal_labs.

It turns out it only took me 10 minutes, because they already have the code, and it was so straightforward...!

https://t.co/JdS8T1wrni

[Mark Woodward @DataScienceMW]

I enjoy doing the @0xbnomial machine learning question each day.

I think this latest one must still have been in my subconscious last night, because I dreamed about it and had an insight in my sleep into a machine learning project I'm advising someone on! https://t.co/lFziGVV4ko

[Mark Woodward @DataScienceMW]

Today I learned that in Python the string '\n' has length 1, not 2...!