Masters of Scale - The 'ghosts' that shape AI & our future (with Kevin Slavin) | Masters of Scale Summit 2024
The speaker presents three 'ghost stories' to illustrate how past data and errors continue to influence the present and future, particularly through AI. The first story is about Henrietta Lacks, whose cells, known as HeLa cells, are immortal and have been used extensively in medical research, highlighting how data (in this case, biological) can persist and impact the world. The second story involves a spreadsheet error in a 2010 economic paper by Harvard economists Reinhart and Rogoff, which led to widespread austerity measures based on incorrect data. This error shows how flawed data can replicate and shape policies. The third story discusses the Enron scandal, where emails from the company were used to train AI models. This example illustrates how data from fraudulent activities can be repurposed and continue to influence AI development. The speaker emphasizes the importance of transparency in AI data sources, as demonstrated by Bloomberg's disclosure of using Enron emails in their AI model.
Key Points:
- Historical data can have lasting impacts, as seen with Henrietta Lacks' cells used in research.
- Errors in data, like the Reinhart-Rogoff paper, can lead to significant policy changes.
- Transparency in AI data sources is crucial, as shown by Bloomberg's disclosure of using Enron emails.
- AI models are influenced by the quality of data they are trained on, highlighting the need for accurate data.
- Understanding the origins of AI data can help identify biases and errors in AI outputs.
Details:
1. 🎤 Setting the Stage: Stories of Ghosts and AI
1.1. Introduction to AI and Storytelling
1.2. Case Study: AI-Driven Audience Engagement
2. 👻 The Immortal Legacy of Henrietta Lacks
- Henrietta Lacks was born in 1920 in Virginia and died at 31 due to cancer.
- Her cells, known as HeLa cells, are unique because they continue to replicate indefinitely, unlike normal human cells.
- HeLa cells have an unusual chromosome count of 35 to 40, as opposed to the typical 23 pairs found in human cells.
- These cells have been crucial in medical research, aiding in the development of polio vaccines and AIDS research.
- The use of HeLa cells has also sparked ethical debates about consent and the commercialization of human biological materials.
3. 📉 Data Disasters: The Economic Ghosts of Austerity
- In 2010, Harvard economists Reinhart and Rogoff published a paper claiming that public debt exceeding 90% of GDP leads to economic growth collapse, which influenced austerity measures worldwide.
- Key policymakers like Paul Ryan and David Cameron used the paper's conclusions to justify cuts in healthcare, retirement, education, and transportation.
- Grad student Thomas Herndon discovered that the paper contained a spreadsheet error, omitting data from Denmark, Canada, Belgium, Austria, and Australia, which undermined its conclusions.
- This error led to significant policy decisions based on flawed data, impacting healthcare, culture, and public services globally.
- The replication of this bad data resulted in substantial economic and social policy changes, illustrating how data errors can shape the world.
4. 💼 Enron Emails: Lessons in AI and Financial Fraud
- Bloomberg developed its own GPT model for finance, openly sharing construction and evaluation methods, including the use of the big bench evaluation to assess strengths and weaknesses.
- The Bloomberg GPT model excels at deciphering complex truth dependencies, benefiting from high-quality data, including both private and public sources such as YouTube subtitles and GitHub code.
- Enron emails, made public after a major fraud case, were used as a data source, providing insights into fraudulent communication patterns, highlighting the importance of data transparency.
- The Enron scandal, involving $185 billion in losses, serves as a historical example of financial fraud, with Enron emails becoming a valuable resource for understanding corporate malfeasance.
- Bloomberg's transparency in using Enron emails contrasts with the general lack of disclosure in AI development, emphasizing the need for openness in AI training data sources.
- The Enron email dataset has been instrumental in training AI to recognize fraudulent behaviors, illustrating the potential for historical data to inform future AI decision-making.