Are AI Hallucinations a Bug or a Feature? (Vol. 5)

Marc Andreessen suggests that while programmers view hallucinations as bugs, they might also be helpful as features. That is, hallucinations—generative AI output that’s “manufactured” or not based on facts—can be valuable when AI is used as a cocreator, a suggester, and a guesser . When treated as a brainstorming aid, its made-up guesses are rocket fuel for human creativity.

For example, Andreessen explains how lawyers use AI’s “made-up” suggestions during case preparation to imagine novel legal strategies. Conor Twomey says that traders on Wall Street are beginning to use generative AI and vector databases to find trading opportunities — to zig when the masses zag.

Isn’t this creativity?

Might hallucinations be a feature, not a bug?

Listen to Marc Andreessen: Future of the Internet, Technology, and AI with Lex Fridman about why, in some cases, he thinks hallucinations are a feature, not a bug.

Using Similarity Search and Hallucinations for Trading Strategy on Wall Street

“It is said that the military is usually well prepared to fight the previous war—an intriguing old saying that reminds us of how susceptible we are to past experiences of our own when projecting the future.” — Hsin-Yu Chiu, Humboldt University of Berlin

The application of generative AI for similarity-based search and trading strategy development is an emerging area for innovation and exploration on Wall Street. For trading strategy development powered by vector databases and similarity search, hallucinations are a helpful feature, not a bug.

Read the Predicative Ability of Similarity-based Futures Trading Strategies by Hsin-Yu Chiu and researchers at Humboldt University of Berlin for a discussion of the use of similarity search for trading strategy development.

On Vector Databases and Similarity Search

“In trading, we assume that technical traders are “similarity-based,” meaning that their judgments of present market conditions and projections of probable future returns are derived from recognizing vivid, concrete patterns of similarities portrayed by a multitude of technical indicators between market scenarios of the present and past market scenarios.” — Neil Kanungo and Nathan Crone

Similarity search allows for exploration of complex, unstructured data sets, but how is this done exactly? Similarity methods quantify how closely two vectors resemble one another, where vectors are numerical representations of data objects such as documents, images, words, audio files, or temporal data known as vector embeddings.

Read How Vector Similarity Drives Contextual Search by Neil Kanungo and Nathan Crone for more about similarity search and vector databases.

This week, try using AI hallucinations for brainstorming!

See you next week.