AI Governance

Transcript

1. Ian Mulvany - CTO BMJ Implementing AI Governance BMJ Case Study
2. Goal of this talk: What actions can each of us take now? How might we best share what we are learning?
3. Things are strange
4. How might we make them less so?
5. 1. 12 Months since ChatGPT became available 2. LLMs predictably get more capable with increasing investment, even without targeted innovation. 3. Many important LLM behaviors emerge unpredictably as a byproduct of increasing investment. 4. LLMs often appear to learn and use representations of the outside world. 5. There are no reliable techniques for steering the behavior of LLMs. 6. Experts are not yet able to interpret the inner workings of LLMs. 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider! https://cims.nyu.edu/~sbowman/eightthings.pdf
6. 1. The pace of iteration is insane 2. LLMs predictably get more capable with increasing investment, even without targeted innovation. 3. Many important LLM behaviors emerge unpredictably as a byproduct of increasing investment. 4. LLMs often appear to learn and use representations of the outside world. 5. There are no reliable techniques for steering the behavior of LLMs. 6. Experts are not yet able to interpret the inner workings of LLMs. 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
7. 1. The pace of iteration is insane 2. They are going to get more powerful 3. Many important LLM behaviors emerge unpredictably as a byproduct of increasing investment. 4. LLMs often appear to learn and use representations of the outside world. 5. There are no reliable techniques for steering the behavior of LLMs. 6. Experts are not yet able to interpret the inner workings of LLMs. 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
8. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. LLMs often appear to learn and use representations of the outside world. 5. There are no reliable techniques for steering the behavior of LLMs. 6. Experts are not yet able to interpret the inner workings of LLMs. 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
9. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. They look like they are intelligent 5. There are no reliable techniques for steering the behavior of LLMs. 6. Experts are not yet able to interpret the inner workings of LLMs. 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
10. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. They look like they are intelligent 5. We don’t know how to control them 6. Experts are not yet able to interpret the inner workings of LLMs. 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
11. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. They look like they are intelligent 5. We don’t know how to control them 6. We don’t know how they work 7. Human performance on a task isn’t an upper bound on LLM performance. 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
12. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. They look like they are intelligent 5. We don’t know how to control them 6. We don’t know how they work 7. They will be better than us at things 8. LLMs need not express the values of their creators nor the values encoded in web text. 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
13. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. They look like they are intelligent 5. We don’t know how to control them 6. We don’t know how they work 7. They will be better than us at things 8. They are a tool, and will be used by those who use them 9. Brief interactions with LLMs are often misleading. LLMs - 9 Things to consider!
14. 1. The pace of iteration is insane 2. They are going to get more powerful 3. We don’t how, they just are 4. They look like they are intelligent 5. We don’t know how to control them 6. We don’t know how they work 7. They will be better than us at things 8. They are a tool, and will be used by those who use them 9. You need to spend time with them LLMs - 9 Things to consider!
15. 1. Stay the same 2. Get incrementally better 3. Get dramatically better 4. Become sentient
27. Customer Context Certainty Disruptiveness Language Generation How we interact with information Workflow / Scaling Value of high quality data
29. Let’s Get Boring
31. Kind of task Risk Level Autonomous decision, high stakes domains - self driving, clinical diagnosis, Peer review Expert Augmentation, coding, legal drafting, scholarly text analysis Mechanical tasks: document analysis, copywriting, code cleanup Danger Danger Will Robinson!, stay very far away Understand, explore, exploit Just do this already!
32. Policy GenAI
33. Governance Data Workflow Output The risk is our reputation
34. Editorial Users Governance Reputation Legal Tech Governance Expertise
35. Example use cases inside BMJ • Software development • User research • SAP management • Report analysis • Copy editing
37. Conclusions