LLM’s Can Only Achieve Superintelligence with Help
I give LLMs a hard time because I know what they are capable of, I know they utterly lack common sense, too often get caught in faulty “reasoning” loops, among other cons. There exists a wide chasm between reality, investor perception and therefore valuations. I do not agree with Elon Musk when he says that more compute and time will eliminate these problems, because LLMs fundamentally are not reasoning - they are recognizing patterns. Pattern recognition does not suggest consciousness the way “reasoning” does, and therefore does not present the opportunity for outlandish valuations. However…
LLMs can become extremely powerful when knitted together in a purpose-built intelligence bundle to drive full process automation.
Imagine a frontier LLM that sits on top in a parent-child relationship, with specialized models (language or other model type), databases and applications that automate the tasks that would otherwise be performed by one hundred, or one thousand humans, or more.
In the case of Investment Management, perhaps Anthropic’s Claude would sit on top and engage with the user in a conversational manner. Claude would have hooks into Bloomberg for market fundamental data, would be integrated into OneDrive to access Analyst models, would link to a robust calculation program for running numbers, would deploy agentic coding capability for executing scripts on an as needed basis (generating a visualization of a sensitivity analysis as an example), would integrate with FinBERT for sentiment analysis, with CEORater for Management scores, and would synthesize these disparate information sources (and others), to help the user generate alpha. A mother model that orchestrates a super-sophisticated, automated process that ingests and analyzes untold amounts of data. Think of it as an automated team approach to Superintelligence. This is possible. All it takes is a bit of imagination and institutional will to make this happen now.
