AGI is a product problem

AI was supposed to change the way we work.

We were promised the most important technology of the century, unprecedented societal change, even the end of life as we know it...

…and so far, all we have is a chatbot.

A chatbot that has only really transformed 1) cheating on homework and 2) copywriting.

A chatbot that hallucinates, regurgitates the internet, and can't process the private information the economy runs on.

A chatbot that doesn’t know how to work.

Chat was just the beginning.

In response to these limitations, Hebbia sought to pioneer a solution.

In 2020, we deployed the first operational RAG (Retrieval-Augmented Generation) system – AI that for the first time could read private data.

And user after user, query after query – RAG failed to overcome the limitations of chat.

For 84% of real-world user queries, RAG, and any chat-based AI failed.

The questions that mattered:

1. Weren’t answered in existing documents, they were net new insights.
2. Were too complex, requiring many steps (not just one search) to get to the right answer.
3. Required too much data – across every file (not a few potentially related quotes).
4. Weren’t answered within text alone– but across charts, tables, images…
5. We're too important to rely upon a black box system.

Foundation models are powerful – and if orchestrated correctly – could already solve almost any task.

This isn’t a problem solved with the next model – with 10 trillion parameters or 10 million token context windows.

This isn’t a technology problem. ‍ This is a product problem.

It’s time for an infinitely capable AI product.

One that can process any amount of information, answer any question, and show its work.

To achieve true AGI, a product must be:

• Autonomous: Empowered to make and update decisions, interfacing with an environment, humans, and AI agents alike.
• Composable: Able to break down complex problems into understandable steps, enabling both AI and users to navigate through solutions with clarity.
• Flexible: Allowing both humans and AI agents to modify sources and steps in ways that update final outcomes.
• Transparent: Providing clear insights into each decision and action taken by the AI.
• Meta: Communicating its own limitations, with awareness of those limitations.
• Specialized: Able to call on expert “tools” – without the need for retraining and fine tuning.
• Self-improving: Designed to leverage prior knowledge for new tasks.
• Multimodal: Able to pull in any inputs and outputs at scale – more than just text.
• Unbiased: Uncovering patterns in the data and across data sets.
• A blank canvas: Allowing any user or any AI to explore in any direction, not constrained to a single thread or line of questioning.

Hebbia’s latest product represents the culmination of three years of dedicated research and development into this platform.

Matrix is the interface to AGI.

Launching March, 2024.