C-LARA

An AI collaborates with humans to build a language learning app.


Weekly summary, Sep 19-25 2024

I have continued to test the o1-preview model: it’s extremely interesting! The new functionality for non-AI-supported languages is nearly there, and the Melbourne Uni students are making good progress.

O1-preview model

Here are a couple of nontrivial things I did with the new o1-preview model over the last week:

  • A 3100 word science fiction story. o1-preview can write much longer texts than gpt-4 and keep them coherent. The quality of the story was IMHO quite good, though I haven’t received much feedback yet.
  • A complex image creation task. We need to test o1-preview’s abilities more carefully, but this initial example was encouraging.

After correctly adjusting the function that calculates costs of OpenAI calls, I however discovered that using o1-preview is extremely expensive. Doing the MWE-annotation step for “Finley the kitten converts to Islam” costs about $0.25 with GPT-4, but $8 with o1-preview! So we need to evaluate more, and plan carefully how we will use it.

Support for non-AI languages

We seem to be nearly there. Sophie found another issue, a messy problem in alignment, but I have hopefully fixed it. We don’t currently know of any more problems, though more testing is definitely needed.

Melbourne Uni student projects

Two of the Melbourne Uni groups report good progress:

  1. Image annotation. This group is working on integrating the Segment Anything Model into C-LARA. If they succeed and they sound optimistic, we may be able to do two things:
    1. Create C-LARA images where you can select a part of the image by touching/clicking and e.g. show a popup or play audio;
    2. Say to the AI that you only want to regenerate a part of an image which is mostly correct.
  2. Recording audio. This group is developing functionality to do direct recording of audio for C-LARA, similar to LiteDevTools in LARA. It will make it much simpler to add audio for individual words in languages where no TTS is available. They also sound optimistic about success.

These projects should be very useful for the work we’re doing with the New Caledonians.

Next Zoom call

The next call will be at:

Thu Sep 25 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)



Leave a comment