[DataFramed AI Series #2] How Organizations can Leverage ChatGPT
Tue May 09 2023
New Innovations and Job Displacement
00:00 - 16:51
- New innovations will displace some jobs but also create new ones.
- Enterprises need to frame the use of generative AI as augmenting human ingenuity rather than replacing jobs.
- Leaders should think about the human component when implementing automation to avoid job displacement.
- Leaders should consider the human component when thinking about automation and potential displacement of jobs.
- New innovations will displace some jobs, but also create new ones and evolve skillsets.
Enterprise Use Cases and Prioritization
00:00 - 22:27
- Throughout the episode, they discussed how organizations should be thinking about chat GPT use cases.
- They discussed a prioritization framework for enterprise use cases of large language models.
- GPT models can support document intelligence and eliminate the need for live agents for ingestion analysis extraction all within the conversation.
- Using GPT on the back end can analyze how conversations are happening, do named entity recognition, summarization and augment data scientist capabilities.
- One model can serve vastly different use cases across the enterprise, including legal triage and RCA for help desks.
Technical Stack and Deployment Challenges
00:00 - 22:27
- The technical stack needed to deploy large language model based applications was also discussed.
- Challenges associated with deploying generative AI in production include data selection, creating an inclusive dataset, fine-tuning, accuracy, and hallucinations.
AI Literacy and Skillset
11:09 - 44:24
- What AI literacy looks like in the context of these tools was also discussed.
- A level of skillset is still required even with large language models, as users must know what they want from the model and how to interpret its output.
- The current education system needs more leaders who are proactively technology-focused and technology-leaning.
- Learning by doing is powerful, as building something useful can show the opportunity that lies within AI.
Generative AI Applications and Efficiency Gains
16:23 - 44:24
- GPT technology can enable systems to communicate with each other without hard coding.
- GPT models can reduce friction for customer conversations and help engineers write better code.
- Efficiency gains in content generation are incredible with tools like Power BI.
- Generative AI solutions can create more efficiencies in a company's talent pool and accelerate growth.
Privacy and Bias in Generative AI
16:23 - 28:05
- The privacy angle of generative AI was explored in this episode.
- Data scraped from 499 million websites is biased due to human embellishment and false information.
- Models are not wrong, they are simply untrained.
- Fine tuning and prompt engineering changes can help control how a response is generated.
- A human in the loop is necessary to monitor data for toxicity, injustice, and bias.
Cloud Infrastructure and Responsible AI
22:03 - 33:24
- Generative AI can sit native inside a secure cloud infrastructure to run applications.
- Cloud providers offer secure infrastructure for running models trained on large amounts of data.
- Enterprises need to build their own secure tenant within cloud providers' shared responsibility model.
- Accenture created a center of excellence and guidelines for using models like Chatsupitina.
- The company also provided an enterprise sandbox on Azure tenant for employees to play with the model without risking client data.
- Azure OpenAI service launched with ResponsibleAI Toolkit built-in, which is part of the solution and not just an option.
- Enterprises are calling for more visibility into black box AI models, and some companies are building model cards that explain how decisions are made in neural networks.
Regulation and Explainability
33:09 - 39:01
- Auditing a black box model is difficult without explainability and auditability mechanisms.
- Responsible AI teams are often shed during tough economic times, making regulation necessary.
- Collaboration with consumers and regulated industries can help build inclusive data sets for regulation.
- Regulation should start at the application level, not research level.
- Transparency and explainability are important in AI models, but many black box models exist today.
Future of AI Models
43:58 - 46:28
- Models will get better and more specific to industries.
- Companies serving medical communities will be able to leverage this technology in meaningful ways.
- Reduction of time to value or time to market for these companies is going to be very interesting.
- In 12 years, organizations of all sizes will start leveraging this technology in meaningful ways.
- App Store for models where people can download legal GPT, bio GPT, etc. is exciting.
- Encouragement for everyone to learn by doing using GitHub repos available from Microsoft, AWS, Google, etc.