jsnns

Hey! We have that and it's integrated with all of our tools and data!

I’ve found this personally very useful — it can already access tons of SEC documents, news articles, company reports, etc. What else would you like to see us add?

Thanks for your interest! We decided to split the AI system into Thinkers, Actors, and Communicators for several reasons.

Firstly, it makes each part more reliable and easier to maintain. This division allows us to craft specific prompts for each component based on their function and the information they deal with. The segregation also allows for better data management and focused improvement in case of performance issues or potential upgrades.

Secondly, it provides a smooth data flow. The Thinker receives a prompt tied to the user's question and starts gathering relevant data. This data then flows to the Actor, which uses it to respond to the user's query or to execute tasks. For example, if a user says "build an automation to buy the dips for the top 10 stocks by market cap in the US," the Thinker first gathers the top 10 stocks from our screener, then the Actor creates the strategy.

Finally, the Communicator receives the outcomes of the actions from the Actor, and it's responsible for relaying the information back to the user in an easy-to-understand format. It ensures that the actions taken by the AI and the results are transparent and understood by the user.

Regarding the prompts for each agent, they outline the commands each agent can access, as provided by DionysusDSL. We use the inputs to these commands to form a "tree" of possible types the AI might need to use. Then, we build a prompt demonstrating these types, allowing for accurate and context-specific responses. Each agent's prompts differ based on their role and the tasks they need to perform.

Hello, Hacker News community!

I've been deeply involved in the AI sector for quite some time now as a software engineer, founder, and previously with roles at Bridgewater Associates and NVIDIA. Recently, I've been pondering the limitations and potential improvements of Large Language Models (LLMs) such as GPT-4.

While LLMs have demonstrated impressive capabilities in understanding and generating human-like text, they are inherently limited by their lack of access to real-time data. This restriction effectively freezes their knowledge at the point of their last training update, leaving them unable to provide information about events or knowledge that have emerged post-training.

Moreover, a significant portion of their training and computational resources is spent on memorizing facts, functioning in essence as massive, intricately structured fact databases. This focus on fact storage often limits their potential for reasoning and creativity.

Given these constraints, I've proposed an approach that aims to shift the focus of LLMs from fact storage to enhanced reasoning. By allowing LLMs to access and integrate real-time data from their training corpus during inference, I believe we can transform these models into dynamic reasoning engines that can provide more accurate, current, and contextually relevant responses, while making more efficient use of their training and computational resources.

I've detailed this proposal, along with its technical implementation and associated challenges, in a recent blog post: https://www.workbyjacob.com/thoughts/from-llm-to-rqm-real-ti...

I am sharing this here not just to present my idea, but to spark a broader discussion on how we can push the boundaries of what's possible with LLMs. What are your thoughts on this approach? Do you see other potential solutions to the limitations of LLMs? How can we address the challenges that such a transformation might pose?

I am eager to hear your insights and engage in a productive discussion on this topic. Thank you for your time and consideration.