C-LARA

OpenAI have released the new o1 model earlier than expected! This has upended our schedule in a very good way, and I’ve spend most of the week experimenting with it. We have made a bit of progress on some other things as well.

O1 model

The new o1-preview model, previously known as “Strawberry”, is now available. It uses Chain of Thought (CoT) reasoning throughout, trained using a reinforcement learning method: this means it’s slower but much smarter. I have integrated it into C-LARA, and you can select it from the “Edit configuration information” screen. Note however that you’ll only be able to use it on your own API key if you’re in Tier 5, which means you’ve paid at least $1000 to OpenAI. So, at least for now, we’re back where we were a few months ago: most people will only be able to access o1 by getting C-LARA credit put on their account.

We need to test o1 more, but initial impressions are that it lives up to the hype. I have in particular created some English texts glossed in Ukrainian, a language pair we’ve previously found challenging. Vladyslav says they are much better than the gpt-4 versions.
Amazing Aisha Loves Comparative and Superlative Forms – Ukrainian version (o1)
Journalist Jamila Loves Subject-Auxiliary Inversion – Ukrainian version (o1)
Supermodel Lisa Loves Phrasal Verbs – Ukrainian version (o1)
The Unicorn and the Killer Robot

It’s obviously important to evaluate C-LAR A with o1 properly, and we should discuss how to do that. One straightforward option is to do another iteration using the texts from our ALTA 2023 paper.

Priority list

Reinforcement learning and Chain of Thought for MWEs. I used gpt-4 to MWE-annotate the dev portion of Francis’s Sherlock Holmes data. We need to discuss the results, and also see what happens when we repeat using o1.

Support for non-AI languages. Sophie and I are making progress on this functionality. After I fixed an issue last week, Sophie says that, at least so far, it’s functioning correctly.

Encyclopaedia article

We will submit a version of the Encyclopaedia article to ReCALL Journal. Branislav is discussing with the editors; it looks like we will be able to publish it in January, when it should still be interesting.

New Kanak languages project

We got a bit of French funding for a new project with Kanak languages; it will focus on creating picture books and picture dictionaries, with the content reviewed by Kanak community members. I’m very curious to see how this works out.

Next Zoom call

The next call will be at:

Thu Sep 19 2024, 18:00 Adelaide (= 08.30 Iceland = 09.30 Ireland/Faroe Islands = 10.30 Europe = 11.30 Israel = 12.00 Iran = 16.30 China = 18:30 Melbourne = 19.30 New Caledonia)