our team has been focused on extensively testing and refining the capabilities of Kruel.ai. The model has demonstrated impressive stability, even though its performance is slightly slower due to the correct implementation of its mathematical framework. A standout feature of this iteration is Kruel.ai’s ability to “converse” about its own codebase. With a comprehensive view across all its files, the AI has shown exceptional accuracy in identifying relevant elements and recognizing necessary updates to functions across different Python files. This level of contextual understanding surpasses previous iterations, where Kruel.ai was more prone to confusion when processing large quantities of code.
We have streamlined the codebase by reducing redundant relationships and optimizing the structure, making significant enhancements to the embedding logic. Ownership tracking has played a pivotal role in disambiguating overlapping entities, resulting in more accurate information retrieval. Furthermore, we overhauled the AI architecture to bolster its comprehension of code structures and to track associations between functions. The system now varies its response depending on the complexity and nature of the task, offering more contextually appropriate outputs. By incorporating books, extensive datasets, and similar knowledge sources, we have enabled the model to extract and understand key concepts effectively.
This current build is the most advanced among all six versions of our memory system. While there are still improvements to be made, Kruel.ai’s comprehensive understanding allows us to confidently shift focus to refining and expanding other aspects of the system.
Looking ahead, there are plans to integrate features such as a “chain of reasoning,” similar to what OpenAI has implemented in their latest models. A recent discussion with Lynda during a four-hour drive provided valuable insights into how we might achieve this level of cognitive processing. Even without an explicit chain of reasoning, Kruel.ai has already generated numerous impactful ideas, demonstrating robust independent thought.
In parallel, we’re delving deeper into advanced algorithms to build more resilient mathematical frameworks for detecting unseen patterns. Additionally, we revisited components from V3-V5 of our memory system, particularly the category-based memory logic used for health data, and reintegrated it into V6. This transition from context-based to math-based processing has been transformative, allowing us to eliminate the need for “Memsum” a previously expensive data processing method we used to enhance speed and understanding. This move has not only improved system performance but also significantly reduced our daily operational costs, now averaging around $1.50. Moreover, by refining dynamic intent processing, Kruel.ai can now generate new intents alongside static ones, enhancing its ability to track key information across various domains. This upgrade provides additional data points that bolster its reasoning capabilities, leading to more nuanced decision-making and intelligent responses.
Our current embeddings are accurate but quite large and complex. There’s significant potential to reduce their size while retaining their quality. Here’s how we’re considering tackling this challenge:
Dimensionality Reduction: Methods like PCA (Principal Component Analysis) and t-SNE (t-Distributed Stochastic Neighbor Embedding) can help compress embeddings into a lower-dimensional space. This not only decreases the memory footprint but also helps zero in on the most informative aspects of the data.
Quantization & Pruning: By lowering the precision of numbers in the embeddings (for example, converting 32-bit to 8-bit), we can substantially reduce their size. Pruning less significant parts further shrinks memory and speeds up processing without sacrificing accuracy.
We’ll be testing these methods along with some other concepts in mind. The goal is to preserve accuracy while maintaining a response time that competes with the best AI systems out there.
Here’s a peek at the beginnings of our model – can you guess which part is its understanding of its own code? Keep in mind, this model is just a few days old. One thing that stands out is how the system learns over time. When it processes new data like PDFs, manuals, or even full Python files, it builds its network, which can make initial responses a bit slower. But ask a follow-up question on the same topic, and you’ll notice how quickly it recalls and uses its newfound knowledge. That’s because it’s not just storing data – it’s optimizing pathways for faster, more insightful predictions.
This real-time learning ability is a personal favorite. Most models learn from inputs, but seeing a system refine its pathways for prediction and understanding over time – now that’s exciting.
In the future, we’ll need to explore enhancing the Whisper system’s voice input to improve vocal recognition. Imagine a robot that can both hear and see: we want to make sure it only responds to authorized users. We wouldn’t want just anyone nearby to control conversations or issue commands, so ensuring secure voice recognition is crucial. It’s on our roadmap, though not an immediate focus.
