Ad
More Related Content
What's hot (20)
Similar to Leveraging Generative AI & Best practices (20)
Ad
More from DianaGray10 (20)
Ad
Recently uploaded (20)
Leveraging Generative AI & Best practices
- 1. Leveraging Generative AI Ashling Partners | Solutions Engineering | Alp Uguray, 4x UiPath MVP
- 2. 2 Senior Solutions Engineer at Ashling Partners 4x UiPath MVP Award Host & Creator at Masters of Automation Podcast (https://themasters.ai) Alp Uguray Introductions
- 3. Innovation Ambition Matrix HOW TO WIN WHERE TO PLAY TRANSFORMATIONAL Large market opportunity identified but very different from what we are doing today. ADJACENT Not doing today, but plugs right into what we are doing today CORE Already doing it today Develop New Products & Assets Add incremental Products & Assets Use Existing Products & Assets Serve existing Markets & Customers Enter Adjacent Markets Serve Adjacent Customers Create New Market Target New Customers
- 4. 4
- 5. 5 Good ones (Utopic Use) • Leverages AI versus. Manual execution productivity gains • Augmentation in task execution as HITL suggestions and recommendations Not so good ones (Most likely) • Job Displacement / Re-write • Digital Misuse • Digital Divide • Vulnerability increase with cyberattacks Worst ones (Cautious view) • Data Privacy • Fake Content and IP Law • Failure of Regulations • LLMs dominate the communication lines - Don’t know who you speak and widespread adoption of personalized Face, Voice and Text Importance of Scenario Planning Driven by productivity gains and improved Customer and Employee Experiences, Conversational AI dominance depends on a few different outcomes based on its adoption
- 6. 6 Focus on realistic applications that can complement existing business capabilities. • Prioritize applications based on ease of implementation and risk level, gradually moving towards more complex and valuable ones. An example of a key application is using generative AI for knowledge management, which can provide immediate value across various business functions Do not have a perfectionist attitude towards the development of AI applications, which could trap you in the proof-of-concept phase without ever delivering value. • An iterative product development approach where applications are developed to solve specific customer or employee problems and are then continuously adjusted based on feedback until they're ready to be scaled. This ensures that the efforts have purpose and contribute towards transforming the industry standards The importance of ensuring that AI adoption doesn't compromise the organization's data and intellectual property security, customer data security, brand credibility, and legal protections. • Collaboration between leaders from operations, technology and data teams, and the legal department to create guardrails that empower the organization without hindering it. Some Guiding Principles in Adoption
- 7. What’s prompt engineering? Prompt engineering is the ‘art’ of optimizing natural language for a LLM. Effective prompts provide the relevant context and detail to a LLM, therefore improving the accuracy and relevance of the response. The quality of prompts directly affects the output of the model. Effective prompts help the model understand your request and generate appropriate responses, in complex or ambiguous scenarios. Tips / Tricks – • Zero-shot Learning: never seen your data, but makes inferences based on understanding • CoT (chain-of-thought) reasoning, ‘break it down, step-by-step’ • Providing relevant context, ‘I am’ or ‘you are’ • First, do ‘xyz’, then do ‘xyz’, finally…
- 8. 8 Zero-shot learning This is a problem set up in machine learning where the model is asked to classify data accurately it has never seen before during training. In other words, the model is expected to infer classes that were not part of its training data. The model typically leverages high- level abstractions and understandings learned from the training data to make accurate predictions on the unseen classes. Zero-shot learning is especially important in settings where it is costly or time-consuming to collect large labeled datasets for every possible class. Few-shot learning Few-shot learning refers to the concept where a machine learning model is able to generalize well from a small number of examples – often just one or two, hence the term "one-shot" or "two-shot" learning. In a traditional machine learning context, models are often trained on large amounts of data, but in few-shot learning, the idea is to design models that can extract useful information from a small number of examples and make accurate predictions. This is similar to how humans can often learn concepts from just a few examples. Shot Learnings
- 9. Some considerations Data privacy and security: • Avoid using real customer data or any personally identifiable information (PII). • Use anonymized or synthetic data sets whenever possible. • Ensure data storage and transfer follow best practices and comply with relevant regulations, such as GDPR or HIPAA. “Hallucinations” - ChatGPT can make stuff up. • Be aware of potential biases in data sets and algorithms, which could lead to unfair or discriminatory outcomes. • Use techniques such as data pre-processing or algorithmic adjustments to minimize the impact of biases. Responsible use of AI: • Ensure that your solution aligns with ethical principles and responsible AI guidelines. • Avoid applications that could be harmful, discriminatory, or promote misinformation.
- 10. 10 Reinforcement Learning Prompt Engineering Chain of Thought How to get the best out of AGIs
- 11. 11 RLHF Reinforcement learning from human feedback further aligns models. (Diagram from OpenAI ChatGPT announcement.)
- 12. 12 Prompting with the “format trick” “Use this format:” is all you need. © 2 0 2 3 S c a l e I n c .
- 13. 13 Specifying tasks using code prompts Prompting through partial code. © 2 0 2 3 S c a l e I n c .
- 14. 14 Specifying tasks using code prompts Prompting with imaginary variables. © 2 0 2 3 S c a l e I n c .
- 15. 15 Using an external interpreter to overcome model limitations in conversational Q&A. “You are GPT-3” © 2 0 2 3 S c a l e I n c .
- 16. 16 Chain-of-thought prompting Figure 1 from Jason Wei et al. (2022). © 2 0 2 3 S c a l e I n c .
- 17. 17 Zero-shot chain-of-thought Figure 1 from Takeshi Kojima et al. (2022). © 2 0 2 3 S c a l e I n c .
- 18. 18 Zero-shot chain-of-thought Figure 2 from Takeshi Kojima et al. (2022). © 2 0 2 3 S c a l e I n c .
- 19. 19 Zero-shot chain-of-thought Figure 2 from Takeshi Kojima et al. (2022). © 2 0 2 3 S c a l e I n c .
- 20. 20 Self-consistency and consensus Figure 1 from Xuezhi Wang et al. (2022). © 2 0 2 3 S c a l e I n c .
- 21. 21 Q&A