Please note: Updates from YC's Startup School have now been relocated to our blog. If you're receiving this message for the first time, it's John here sending it to you.

What we've done and learned

Nexus

On my way to work on Monday, March 11th, the concept of employing the Monte Carlo method to evaluate digital marketing performance emerged. This idea, originally suggested by Bobby, revolves around simulating uncertain environments with the Monte Carlo method. Further research led me to understand that regardless of the domain, the technology we develop at Nexus can conclude that LLMs are adept at producing realistic random samples of human behavior. This insight clarified our next steps.

Firstly, we aim to demonstrate that individual simulations accurately mirror human behavior, urging us to examine if each profile yields logical outcomes. Secondly, we seek to confirm that the probability distribution from numerous simulations presents promising results, compelling us to test this by running a campaign for a hypothetical product to see if it significantly impacts our success metrics.

In the first phase, we completed the initial version (well it's actually the second version) of the manual simulation. During testing, we discovered that the model occasionally does not respond to certain provocative images, such as those found in Tom Ford advertisements, due to moderation constraints.

TalkToDART

We've achieved significant advancements in the development of TalkToDART, particularly with the RAG engine. This innovation, developed entirely in-house, promises to enhance the development of other modules. Our approach leverages LLMs for task routing based on complexity, assigning tasks that demand a more deterministic format to API calls, regardless of whether they are developed internally or by external partners.

In the development of the RAG Engine, we've implemented a data fetching mechanism inspired by computer architecture principles, specifically caching. Data requests are initially searched in our database, serving as a cache; if found, results are instantly returned. Conversely, in the event of a cache miss, the module queries the data source (DART) to fetch the required information.

Dylan has significantly improved the performance of the Reranker and the Refiner. The Refiner's role is to refactor the initial response from the module, making it easier to process for subsequent computational steps.

ASQ

We've polished the architecture for ASQ v0.1 alpha and have decided to make our design public. We believe the key differentiator for products using multi-agent systems lies not in the architecture itself, but in the design of user interfaces and Modules. Modules serve as essential tools for ASQ, underscoring the importance of being equipped with highly functional capabilities.

TalkToDART serves as our pilot Module, demonstrating our capability to develop state-of-the-art performance products. We believe we're nearing our goal and are excited to showcase how our product will exceed expectations.

Tough Call

Many of you know that we have been working on three products: TalkToDART, ASQ, and Nexus. We know that putting resource into a single product is a norm when it comes to achieving somethings exceptional, not only startups. So, this week I decided to make a tough call: to shift Nexus down to stealth mode.

It was really hard because I love both concepts. But I had to choose one as we needed to compress our timeframes. Just like Phil Knight states in his autobiography Shoe Dog, grow or die.

This decision does not signify abandoning Nexus. We will continue to allocate time to it, albeit sparingly. To quantify, the distribution of our focus would approximate a 19:1 ratio.