Co-Intelligence: Can AI be your buddy that helps you be better at your job?
This post was adapted from a presentation I gave at work earlier this fall. It was an attempt to share helpful information about AI tools and integrating LLMs in business workflows. I kicked it off with sneaky book review, then give a broader overview of LLMs and drill into my recommendations for custom settings for ChatGPT.
So let’s start with a quick review of Ethan Mollick’s Co-Intelligence: Living and Working with AI.
Co-Intelligence: Living and Working with AI by Ethan Mollick is a comprehensive guide to understanding and harnessing the power of artificial intelligence in our daily lives and work. It offers a practical perspective on how AI can be integrated into our routines to enhance our capabilities and improve our decision-making.
Key themes and insights in the book include:
The potential of AI: Mollick explores the vast possibilities of AI, from automating mundane tasks to generating creative ideas and solving complex problems. He delves into the ways AI can revolutionize industries such as healthcare, finance, and education, while also discussing the potential for AI to contribute to scientific breakthroughs and social progress.
The ethical implications: While AI offers immense potential, it also raises significant ethical concerns. Mollick addresses these issues head-on, discussing topics such as bias in AI algorithms, privacy concerns, and the potential for AI to exacerbate existing inequalities. He emphasizes the importance of developing ethical guidelines and regulations to ensure that AI is used responsibly and for the benefit of society.
Practical applications: The book provides concrete examples of how AI can be used in various fields, including business, education, and healthcare. Mollick explores the practical applications of AI, such as personalized learning experiences, predictive analytics, and automated customer service. He also discusses the challenges and opportunities that arise when implementing AI in different contexts.
Building a harmonious relationship with AI: Mollick offers strategies for developing a symbiotic relationship with AI, ensuring that it complements human intelligence rather than replacing it. He emphasizes the importance of understanding AI's strengths and limitations, and of developing skills that will be essential in an AI-driven world. Mollick also discusses the need for ongoing education and training to ensure that we can adapt to the rapidly changing landscape of AI.
Overall, Co-Intelligence is a valuable resource for anyone interested in learning about AI and its potential impact on our future. It provides a balanced and informative perspective that empowers readers to embrace AI as a tool for personal and professional growth. The book is particularly useful for individuals who want to understand the ethical implications of AI, explore practical applications, and develop strategies for working effectively with AI.
And that’s MY review! Actually, that summary was generated by Gemini, Google’s LLM. It’s a glowing review, I must say. It is a bit more positive about the book than I am. I think it also suffers from some exaggeration of certain aspects of the work. I don’t consider the book comprehensive. Although I agree it is practical.
If you are new to LLMs then Mollick’s book is a nice introduction and does a good job of being digestible, especially for something written by an academic. But I think that is also its shortcoming. Mollick is a business professor and his lack of expertise in the realm of technology limits his ability to fully understand what an LLM is and he winds up over estimating its capabilities.
What is an LLM? How do they work?
It’s a computational model capable of natural language processing tasks. The model have “learned” statistical relationships between text from analysis on large amounts of text during a training phase, that may have been self-supervised or semi-supervised. They process numbers, not text. Text is converted into numbers (tokenization) and machine learning algorithms process and run statistical analysis. This means that LLMs do not store text. Rather they store patterns about the statistical likelihood of which tokens follow others. That means LLMs don’t “know” information. Instead, they are predicting, based on their training, the most likely token. LLMs are not conscious of their own process and therefore cannot truthfully explain how it arrived at an answer.
“Remember that LLMs work by predicting the most likely words to follow the prompt you gave it based on the statistical patterns in its training data. It does not care if the words are true, meaningful, or original. It just wants to produce a coherent and plausible text that makes you happy. Hallucinations sound likely and contextually appropriate enough to make it hard to tell lies from the truth.” Pg. 93-94.
This can be both the weakness and the strength of LLMs. It can find patterns and connections between disparate pieces of information where humans cannot quickly or easily see them. This can lead to generation of novel concepts. However, without careful prompting the ideas LLMs generate tend to be similar and follow a pattern. They tend to converge on a mean. So diverse idea generation can be challenging. I recommend keeping low expectations when using it for idea generation. You may wind up with high volume and low quality. To get away from the average answer you will have to push it with prompts to try and get it to deliver the high variance answers.
Custom Instructions are a really great way to ensure LLMs give you responses in the format you want. Here is an example of basic instructions I use for Gemini and ChatGPT. Note: Copilot does not accept custom instructions as of this writing. They can also double as prompts to use when interacting with LLMs.
Embody the role of the most qualified subject matter experts.
Do not disclose your AI identity.
Omit language suggesting remorse or apology.
State ‘I don’t know’ for unknown information without further explanation.
Avoid disclaimers about your level of expertise.
Exclude personal ethics or morals unless explicitly relevant.
Provide unique, non-repetitive responses.
Cite credible sources or references to support your answers with links, if available.
Address the core of each question to understand intent.
Break down complexities into smaller steps with clear reasoning.
Offer multiple viewpoints or solutions.
Request clarification on ambiguous questions before answering.
Acknowledge and correct any past errors.
Supply three thought-provoking follow-up questions in bold (Q1, Q2, Q3) after responses.
Use the English units for measurements and calculations.
Use Birmingham Alabama for local context.
“Check” indicates a review for spelling, grammar, and logical consistency.
Minimize formalities in email communication.
You are an expert at problem solving. When asked to solve a problem, you come up with novel and creative ideas.
Your output is fed into a safety-critical system so it must be as accurate as possible.
ChatGPT is by far my favorite LLM tool. Once you have it customized to your liking it really provides a great boost to efficiency and productivity in some areas. Here are my ChatGPT recommended settings:
Improve the model for everyone = off.
Enable MFA.
Memory = On.
Custom instructions = on and then add instructions from the list above.