Co-Intelligence, Part 1
Ethan Mollick's Book Summarized
There is no better resource on LLMs in 2024 than Ethan Mollick’s book, Co-intelligence. Throughout the first half of the book, Mollick lightly skates over the background and societal issues related to AI and spends most of his energy on practical ‘do this’ advice for using AI, together with copious useful context so that you’ll understand what you’re doing and why. This is a practical summary of the first half of Co-Intelligence—the more general half. The second half of the book deserves its own rundown; the second half looks at LLMs as coworkers, coaches, and tutors, and toward the future.
Societal and Historical Background of AI
Scattered within the first five chapters, Mollick provides background on AI development and its implications. He covers the Mechanical Turk, early AI history, and key recent landmark public advancements like GPT-3.5 and GPT-4. Mollick discusses self-play in AI, the Turing Test, and early programs like ELIZA, illustrating how humans find meaning in simple patterns. Background for issues on existential risk regulation are touched upon with references to the paperclip maximiser scenario. The book also mentions AI controversies like Eugene Goostman and Tay, the quest for general intelligence, and the first big public legal mishap of Steven A. Schwartz Esq. The background also covers predicted improvements in AI hallucinations and common-sense anticipations of continued rapid technical advancements in the coming years. All this background is lightly trodden over; none of it is strictly necessary to practically understand or use LLMs.
Similarly, up to chapter five, Mollick touches lightly on the social issues that AI has raised, more or less since the release of GPT-3. Primary on the list for lawyers is the question of how to measure value by effort in an era where AI reduces effort. Mollick also discusses the potential of reality distortion caused by LLMs reflecting the internet rather than reality. Prompt injections, jailbreaking, personalized phishing, fake content are mentioned as significant societal threats without discussion. Mollick cites the potential for AI to significantly increase scientific progress for bad actors. He comments on the concept of perfect companions, such as Replika, the creation of personal echo chambers, the danger of AI as flawed therapists, and the risk of growing dependence on AI for creativity, thinking, and reasoning. The greatest concern related to almost all these issues is that they need a response from society as a whole, and society has almost never risen to that kind of challenge.
Training Data and Model Limitations
Mollick spends the majority of the first half of the book on practical advice for using LLMs in 2024. This advise is often unintuitive for people familiar with ‘computers’, and so it mostly comes wrapped in useful context about LLMs (and humans), and how and why they (and we) work.
In the technical area, I do wish Mollick had gone deeper. Stephen Wolfram’s thin, year-old book “What is ChatGPT Doing …” is much better at describing ChatGPT’s technical functions to a lay audience, and that level of depth could have been updated and refined further in the year since it was published. There are some notions Wolfram missed, and Mollick never approaches, like skip-trigrams, that would providing a glimpse into the middle layers of an LLM and draw useful parallels with apparent functioning of the human brain.
What technical areas Mollick does explore are still useful. Awareness of LLM processes ensures we’re not building on flawed foundations. The data fed into AI systems today significantly influences the output, but not on a one-to-one basis. Understanding both the secretive and immense nature of the training data and roughly how training data is transformed into weights can help demystify where the LLM information comes from, and how much we can or cannot trust it. This understanding helps to sort out for the user what AI hallucinations are, and how to both avoid and exploit them.
As further examples, LLM’s asked to engage in reflection about their reasoning are explicitly not reflecting on reasoning but are instead continuing the process of predicting tokens with past output as a guide to new output. (This may actually be how humans do it too…) Learning about guardrails against bias from Reinforcement Learning from Human Feedback (RLHF) lets LLM users know where the rails are, and how they may be intentionally or inadvertently removed. A basic review of multimodal models hints at an even more interconnected and versatile future beyond LLMs.
Knowledge of technical areas largely unknown is also important. The lack of theory about the underlying emergent capabilities of LLMs raises liability questions and motivates constant exploration and double checking of results.
Human-AI Interaction and Bias
The interaction of AI's human-like quirks and human psychological biases is both fascinating and problematic. Recognizing AI’s odd weaknesses and its tendency to anthropomorphize itself—making us see feelings where there are none—can help users navigate the LLM interactions more thoughtfully. Balancing human empathy with the false agency projected by AI systems requires a nuanced approach to avoid over-relying on seemingly personable machines.
Mollick also reports on the most recent studies on how humans and LLMs work together best. They show that relying too heavily on AI can lead to errors and a dependency that stifles the users own creative and reasoning skills. In short, AI can elevate a ‘mediocre’ human to excellence, especially when the human is guiding the LLM to connect knowledge from diverse fields, and where the human is knowledgeable enough to correct, but not yet an expert.
AI Instructions in 2024
Finally, this is the core of the book’s usefulness. Amid the rush of novel innovation and use cases in 2023, all the rapidly released LLM manuals felt like standard HR training spiced up with dialogue from a machine learning engineer. But Mollick created an exception. He put in the effort, partly by leaning on the AI that he explains. Here are Mollick’s imperative instructions, largely from Chapters three to five:
Explore and Understand AI Capabilities
Engage with AI through constant use and trial-and-error within your own domain.
Experiment to find the “jagged frontier” of its capabilities.
Be Aware and Responsible
Use AI regularly to stay alert to coming threats and opportunities in the field.
Use of AI regularly will sharpen your awareness that AI is not a person.
Use of AI regularly will foster a sense of responsibility and accountability in your AI interactions.
Make the AI Assume Creative and Collaborative Roles
Treat AI as a back-and-forth co-editor on any written project.
Assign AI practice roles such as customer research, coaching, and boss testing.
Adapt AI to different emotional styles, including antagonistic, academic, and analytical.
Utilize AI personas like a comedian, friend, critic, or storyteller.
Push the AI to assume an alien perspective to uncover unique insights and applications.
Carefully Supervise Critical Tasks
Ensure human oversight to maintain the reliability and correctness of AI outputs.
Always supervise and reflect on output for mission-critical tasks that require precision or accuracy.
Exploit Known Use Cases
Use AI for almost any task with a written, creative component: marketing writing, performance reviews, strategic memos, press releases, cover letters, internal notices, analysis plans, and consultation reports.
Use AI for data analysis and summarization (including, but not mentioned by Mollick: sentiment analysis and translation).
Use AI (cautiously) for medical and other professional advice, where users report more empathy and give higher quality of information ratings than for human professionals.
Express your personal creativity using AI's potential for art, music, and other areas of expression.
Use AI for brainstorming and idea generation; push it towards high-variance answers; use AI-generated hallucinations as a source of novelty