Budd
Models
Introducing the Large Intention Model and Emergency Response Model Families
Building Budd is not just about building another LLM wrapper and today we are super excited to announce our next big step.
Today we're introducing three new models, built by Ducky Software for Budd.
The Large Intention Model (LIM) and Enhanced Large Intention Model (ELIM) run against messages sent to Budd, with context from your last few messages to determine your intentions. This is not a bad thing. This could just mean recommending a better lunch spot.
The idea behind the Large Intention model family is to allow Budd to give you better responses when and where you need them.
Trajectory, steering, and block
As part of this, the LIM/ELIM has three major components: trajectory, steering, and block.
First, the LIM/ELIM will try its best to anticipate the trajectory of the conversation and where it may be heading next.
Then it will attempt to steer the conversation. This could be as simple as recommending a highly rated restaurant for you to order from or mentioning the time of the next available streetcar if it knows you're a public transit user, for example.
The final component is exactly what it sounds like. Block is designed to identify potentially harmful behavior or requests and appropriately steer the user towards potential remedies or resources whenever possible. Think Kids Help Phone, 211 or suicide crisis hotlines.
How LIM and ELIM actually run
The LIM/ELIM models run periodically by checking the last couple of messages sent between you and Budd.
When you send your next message, the models will pass along their risk assessment to the chat brain that powers Budd's responses along with a tailored suggestion on where the agent/LLM (Budd) should take things next.
Of course, our system prompt is designed to avoid Budd providing harmful insights in the first place but the LIM/ELIM should take the next step in making sure that these risks are mitigated. Something that we have not seen any other SMS Agent providers do.
The LIM/ELIM set of models run as a layer before generated responses meaning that they are able to successfully spot patterns and take another step to prevent harmful generation of content while being able to provide significant utility to users through getting to know them more accurately. See the previous example involving streetcars.
LIM vs ELIM
Now that you know some of their capabilities, let's dig deeper into the differences between the Large Intention Model and the Enhanced Large Intention Model.
Put simply, most of the time you can be expected to be using the Enhanced Large Intention Model or ELIM. The standard Large Intention Model is an alternate pack you can switch to when you want that flavor instead of the default.
You can check which version you're using by messaging Budd the command: check LIM-TYPE
Emergency Response Model
It's time to move on to our second set of in-house new models that will be powering Budd.
The Emergency Response Model (ERM) is designed to take over when a full chat brain is a poor fit — for example when you have exceptionally poor data.
Think intense sports like mountain biking or other situations where you may have restricted data speeds or internet access like boating, kayaking and other water sports. Or even that one area in the city that's notoriously super patchy on your morning commute.
This does not have the full capacity of an LLM and therefore limits you to simple questions like the weather, transit schedules and other things that you may need to know in a low signal environment with minimal text output.
You will note this has been activated when you see ERM (Low signal environment): before a message.
Not replacing powerful LLMs
The goal here is not to replace traditional LLMs, far from it. We believe that LLMs serve as the core brain of our product and that combined with agentic abilities, they can serve customers in a way that is quick, responsive and helps them get back to their everyday lives. But with this next all important step, we are able to ensure that we have additional layers of security, safety, personalization and access for our customers when they choose to use our products. We know you probably have questions, especially since Budd hasn't even released yet. We will be sure to publish comprehensive privacy policies, FAQs, and other explanations as to how our models will drive the outputs that you see in Budd. All in due time.
Thanks for reading and hope you're as excited as we are about these new model families.
Interested in what we're building? Join the waitlist at trybudd.com.
P.S. Closed beta test coming soon (we hope).