The explanations and Ask AI answers I get are pithier and better. Also, the Better AI Model option in settings was something like “…uses the state-of-the-art GPT-4o model…” but now I see only a mention of a more advanced AI.
Yes, well spotted. Since 30 September, half of Readlang users have been switched from GPT-4o mini and GPT-4o to OpenAI’s newer GPT-6 Luna (standard) and GPT-6 Sol (advanced, used by the Better AI Model option). The other half are still on the previous models for now so that I can compare the two groups and check for any potential issues. Assuming that there are no problems I expect that within the next couple of weeks everyone will be switched to the new, better models. When I do that I’ll post an update here on the forum.
Thanks for the feedback. Please let me know if you notice anything that seems worse than before.
5 Likes
Everything seems better now. No adverse effects whatsoever. I love Readlang.
2 Likes
@steve I also noticed the more concise explanations and I like it. Does this mean that I’m already on the new model or were there some general changes in the prompt?
In case you are willing to share in the upcoming post I would be interested in to how the models compare.
1 Like
Yes, you’re in the group that’s on the new GPT-6 models.
The prompts for the other group haven’t changed. For the new models I made a few small prompt tweaks they needed, for example asking them not to use bold or italic formatting in explanations, but nothing asks for shorter answers. The more concise explanations come from how the models themselves are interpreting the prompt.
For the record, here’s roughly how they compared in my tests before the switch:
- Translations: on around 20,000 real Readlang translations across 126 language pairs, graded by another much slower and more expensive AI model, GPT-6 Luna scored 87 out of 100 compared with 77 for GPT-4o mini, and GPT-6 Sol scored about 91 compared with 84 for GPT-4o.
- Explanations: the new models’ explanations are about half as long as GPT-4o’s, and the AI judges preferred them over the old models’ explanations more than 90% of the time.
Thanks for the feedback. Please let me know if you notice anything that seems worse than before.
3 Likes
I strongly agree with the AI judges 
The new model uses half the amount of words but I don’t feel that there’s anything missing in terms of content.
I also feel like the “wrong capitalization for German” issue might be gone (it was a lot better after you turned off the special treatment for German, but I still got a capitalized adjective occasionally). I will try and pay a little more attention to that.
1 Like
Nice! This sounds awesome. I just noticed that it’s different and there’s a noticable increase in accuracy. I am assuming I am in the GPT 6 Sol group? I have the advanced option ticked. I have used Sol before and it is a really good model.
1 Like
Yes, you’re in the group that’s on the new models, so with the Better AI Model option enabled you’re getting GPT-6 Sol.
1 Like
It would be nice if you offered us to sign in with chatgpt (or whatever they called the new feature they released during the last devday) so that we can use our own openai subs to get the the best possible ai, that will also save you from paying openai for our inference.