In this article we explore how we can further reduce randomness in terms of AI outputs even in other areas of SIKE such as the Dreams area. Dreams is where users can add various ideas (in a post-it notes type format) and make queries from them. This is great for project management, meeting tracking, managing clients/users, ideation and more. What’s great is that users can then write queries and get answers from what’s in these notes and answers will only stem from there. This is an easier way of managing information but can lead to issues of stochasticity as we detail below.
In terms of the stochasticity issues, we detail this and how we solve it in other areas of SIKE in our previous article “Reducing Stochasticity in Corporate Applications of Language Models through Semantic Programming”.
Here’s how randomness plays out in the Dreams area. Imagine an initial scenario where we are looking at former WeWork founder Adam Neumann, and how he might comment on a particular news article. In this example, we have a note on his background, a note on the style in which he speaks and another on the news article that we want him to comment on.
Here’s what’s in the STYLE note
and the NEWS ITEM note
Given this setup, a user can ask a query like the following on these notes “using the about note on Adam Neumann and the Style note, respond as he might to the news item” and you see the response start to form in this animation with the full response below.
Another user could ask a different but what they think is a similar query, in this case it’s “how would Adam respond to the news item“. You can see the animation here and the full response below.
Even if we asked the same question again, without steering and providing the aspect of the steps we want the model to take, we get a different kind of answer – see the animation for it here and the full response below.
We can see from the above that we get an answer that is generally in the same kind of style but can differ in terms of structure and even language.
What we’ve found that works quite well is to actually provide a set of instructions as a note which you can see below.
We could add more to these instructions in terms of specificity but here we show how we can do something like have the response and then have a 3-dot point summary and a historical quote.
Running a query on this the first time is as simple as saying “Run Instructions” which you can see in action here.
Each time the user needs to run this query they would just type Run Instructions and to make changes they specify that in the INSTRUCTIONS note.
The full result is below.
Running that query again (animation here) produces a similar style of result.
Outcomes
What’s important to know here is that through the specific context (in this case the notes we have in this Dream thread) and the specific steps (e.g. the instructions), we can get reliable outputs each time.
This can make usage of the Dreams area less stochastic if that’s what you’re after though it can certainly be useful for ideation and creative purposes too.
As we mentioned in our earlier blog on stochasticity, users can also use the Teams function to manage how this works for steps that are quite procedural (see example below), but it’s also good to know how you can do this in the Dreams area of SIKE as well.
For more information on SIKE, please check out the details here (or email us at we@notcentralised.com).
