We used Anthropic's flagship model, Claude 3.5 Sonnet, to identify an investment opportunity that could outperform the market over a one-year period.

We used Anthropic's Claude 3.5 Sonnet to generate the latest stock recommendation for The GPT Investor. This essay will outline the technology, methodology, and result of this process.

Background:

Anthropic, one of the world's leading AI companies, was founded in 2021 by a group of former OpenAI researchers, including CEO Dario Amodei. The company's mission is to develop AI systems that are not only highly capable but also safe, reliable, and aligned with human values.

Since its inception, Anthropic has raised an impressive $7.3 billion in funding, including significant investments from Amazon ($4 billion) and Google ($2 billion).

One of Anthropic's notable achievements is the development of the Claude AI model family, which excels in tasks requiring advanced reasoning, math, and coding skills. In March 2024, Anthropic took center stage in LLM (Large Language Model) research after releasing Claude 3 Opus, which outperformed the existing LLM leader, GPT-4, in benchmarks. Claude 3 Opus was so impressive that some AI researchers claimed the model exhibited signs of self-awareness.

On June 20th, 2024, Anthropic released a newer model, Claude 3.5 Sonnet, which they claim to be more intelligent than their previous model, Claude 3 Opus.

Anthropic claims that Claude 3.5 Sonnet outperforms OpenAI's latest flagship model, GPT-4o, in most benchmark tests, including those for graduate-level reasoning and reasoning over text.

Methodology:

Our approach for this round of recommendation focused on two considerations:

1. Claude's lack of internet access: While Claude has an impressive list of features including code generation and document editing, it lacks the straightforward ability to browse the web. It's likely that the Anthropic team believes web browsing might negatively affect the LLM's responses by incorporating information from a limited number of sites it would access before generating a response. This concern is valid, as hundreds of millions of dollars are spent training these models but when these models perform web searches, the system can only afford a tiny number of searches due to computational constraints imposed on the end user. (An end user isn't allowed to dictate how many sources the system should consult during a web search since each additional source costs money.) So while web searches have the clear benefit of access to the most updated information, they likely have the significant disadvantage of potentially skewing responses based on low-quality or limited searches. We designed our methodology with this limitation in mind.
2. The GPT Investor's challenge with investment timelines: At The GPT Investor, we've noticed that the hardest part about using LLMs to generate stock recommendations has always been determining the timeline of the investment. While The GPT Investor has consistently outperformed the S&P 500 (Note: 90% of professional money managers never beat $SPY), many of the stock recommendations made significantly more gains AFTER the scheduled date for liquidation. It was clear that the models picked the right stocks but couldn't accurately account for the timeline we had instructed in the prompts. To address this issue, we've started generating longer-term recommendations (e.g., 1 year). Since Claude 3.5 Sonnet's training cutoff date is April 2024 and it doesn't have access to the internet, we tried to mitigate the issue by providing the model several references related to the current time and date that it would find helpful.