Welcome to the future - AI with a train of thought 🧠

February 4, 2025

Fifth Dimension Team

In a major technological breakthrough, a Fifth Dimension engineer has fundamentally changed how AI thinks by giving it something it never had before - the ability to maintain a continuous train of thought, leading to significantly better reasoning on complex analytical tasks.

When running Meta's Llama 3.1 8B model through our newly developed AI architecture, without changing the model itself or giving it any extra training, we achieved 89.01% accuracy on mathematical tasks and 91.04% on scientific reasoning - dramatically outperforming the base Llama 3.1 8B model.

The Current Challenge with AI

Current AI systems have a fundamental limitation - they completely reset their thinking process between each word they generate. While they can see that conversation with the user, any documents shared, and their internal knowledge base, they have to rebuild their entire understanding from scratch with each new word they write.

Although current systems perform well, there’s room for improvement in every one because the models don't have this ongoing “train of thought” that allows them to reason better - only their training, context and previous words.

Until now.

The Breakthrough 💡

Several months ago, our Founding AI Engineer, Thea Aviss, had a remarkable insight about improving AI's reasoning capabilities. Today, this vision has become reality with the State Stream Transformer (SST).

This revolutionary advancement allows AI to maintain a continuous "train of thought" while processing information. Instead of analysing each document or data point in isolation, this new architecture enables AI to maintain its own internal reasoning process between each word it writes, leading to significantly better reasoning overall.

When testing against Meta’s base Llama 3.1 8B model against the same model in our AI architecture, we saw incredible results:

In mathematical reasoning: SST achieved 89.01% accuracy with no prompting aids, beating Llama’s 84.50% even when using their most advanced prompting techniques
In scientific reasoning: SST achieved 91.04% accuracy, significantly outperforming Llama’s baseline of 83.40% and the best effort of 86.86% with advanced prompting
Enhanced abilities for state awareness, hypothetical planning, uncertainty, self-questioning, and self-correction

What makes this particularly significant is that these improvements come from better processing - not from bigger models or more data. We've helped AI think more effectively with the knowledge it already has.

Why This Matters for Real Estate

‍While most technology companies focus on processing more documents faster, we've taken a fundamentally different approach. Though speed is important, it isn't just about doing the same things faster - it's about enabling truly intelligent analysis of your real estate data.

Because our AI can now maintain a continuous train of thought, it can step back from problems, recognise when something seems "off", and correct its own thinking - just like a human would. This means even fewer attempts to get to the right answer in complex analysis, as this breakthrough means that AI can now reason even more deeply about the complex real estate scenarios your team members are working on every day.

‍Looking to the Future

‍At 5D we don't follow the crowd - we lead it. Our commitment to innovation continues to redefine what's possible in AI, bringing game-changing advantages to our customers that you won't find anywhere else.

2024 saw us revolutionise document processing, developing technology that outperformed industry giants like OpenAI, Google, and Nvidia to transform how real estate professionals work with their vast document repositories.

Today, we’ve unveiled an AI that actually thinks and has a train of thought internally, leading to significantly better reasoning for complex real estate tasks.

Tomorrow? Well, we're already working on something extraordinary with this new AI technology - and we can’t wait to tell you.