Announcement

Today: New stuff in the repo!

Coming weeks: New competitions!

We are updating our repo today. First we are updating the base script for the first time since launch. We are also introducing the first in a series of security features, to be dropped over the next few weeks.

1 Today's Code Drop

1. New Base Script: Since launch, miners have massively improved prediction accuracy. So we want to make sure that new miners are building from a better base.
2. Opt-In Copy Protection: This new feature allows miners choose their own compute service, as an alternative to Chutes. Note that if you have already used Chutes on a UID, you will need to register a new one. Remember! (Aas well as being a bit naughty) copying makes it impossible for validators to achieve consensus.

2 Forthcoming Competition Changes

We are also working hard on the next two competition changes:

1. Update Scoring to Semantic Bias: While there is some value in lexical prediction scoring, ultimately it is the semantic score which matters most for translation. We will make this change shortly.
2. Introduction of a Confidence Quotient: Having the best predictions in the world will not improve latency, unless the system knows when it is making a good prediction. And... some utterances are more predictable than others:-

3 Confidence Examples

These show the extremes of predictability. LLMs can predict group A very early, but can't predict group B at all. This has led us to product decisions. For B2 names of attendees could be in a RAG data-set or prompt. If the order is wrong it won't affect the semantic score.

A: Predictable at 2 words:

1. Our Father who art in heaven, hallowed be thy name
2. I solemnly swear that I am up to no good

B: Unpredictable to the end:

1. My fave songs are: Downtown, For No One, Visions of Johanna, 96 Tears & Roxette
2. Attending today's event are: Liesl, Ximeng, Edgar, Retaj, Marielle & Arev