HapticAI

In recommendation systems, vector DBs shine bright. By encoding user preferences and item features as vectors, we can quickly find the most relevant recommendations. Next time you get a banger Netflix or Spotify recommendation thank the vector DB.

From the outset, these vectors look like random numbers. But in reality they are a dense representations of data, capturing the essence of text, images, or any other type of information. It really is one of these magical things in AI!

What really is RAG and RL?

Today, we thought we would help explain key concepts and designs that are critical for under the hood functioning of our network. We will be starting with RLHF. Specifically, comparing two techniques: Direct Preference Learning (DPL) and Recursive Reward Modeling (RRM).

The initial feedback flow is simple. Users submit a prompt to the LLM, which will then give out two responses to the user. The user can choose which response they preferred and their feedback is submitted to the network.

Once on the platform, you can select a model operator that you want to use, and then choose the specific model you want to use and provide feedback for, from the dropdown section on the top of the dApp.

As a user, once you open the dApp, you will be prompted to sign a transaction with your Ethereum wallet to access the feedback platform. You can view your account status on the bottom left of the application.

It’s been a while since we announced our website and dApp revamp, and initial feedback on UI from users has been positive. Before we dive into some exciting updates, a quick run through on how you can use Haptic today.

2/6  You will soon be able to use your preferred LLM through the Haptic app! For each query, you'll receive two responses. Choose the one that best fits your prompt, and your feedback will be stored and used to refine the models.

Researchers realised is that in the process of recursive synthetic data generation we lose the the perplexity of real world data and shrink the improbable events occurrence over generations further adding bias to datasets