Key Takeaways from Today’s AI Class – Cargo Cults, Prompting, and the Rise of Intelligent Agents, and what about Privacy
In today’s Part II of Harvard AI Diaries, I reflect on a series of thought-provoking sessions exploring how AI intersects with human intelligence, decision-making, and future technologies.
We covered so much today and my critical thinking along with ideas of industry application was on over-drive.
I’m just over two days in and already I am hugely motivated to bring the learning to a broader audience. Being here benchmarks my own knowledge and I was delighted to see my own experience of building Custom Chatbots being shared with the class today.
We started the day with a new theory I hadn’t heard of before, presented Dr Edlyn Levine – the Cargo Cult AI analogy. It’s a warning of sorts: many of us may be imitating intelligence without truly understanding it.
AI systems can mimic smart behaviour, but they don’t understand context or causality. The danger lies in humans jumping to conclusions based on surface-level patterns, mistaking correlation for causation. This class reminded us that critical thinking must stay in the loop.
We explored six fundamental limitations of AI: its lack of universality, transparency, and causality; its dependence on hardware; its vulnerability to manipulation; and its purely data-driven nature.
The big takeaway? AI is powerful, but not infallible. The human in the equation is still the single most important driver of AI value.
Then we moved into prompting techniques, guided by Prof Teddy Svoronos.
His TIC framework, Task, Instructions, Context, is a foundational tool for good prompting. For complex, unfamiliar, or repeatable tasks, prompts should be fleshed out, step-by-step, with examples and structure.
Chain of Thought (CoT) prompting, in particular, helps reasoning models break problems into logical steps, essential for data analysis, policy planning, or anything involving logic.
We looked at the distinction between reasoning and non-reasoning LLMs. Older models, like GPT-2 or base GPT-3, operate like predictive parrots, fluent but shallow. In contrast, reasoning models such as GPT-4, Claude 3, and DeepSeek can simulate logical thought, plan steps, and use tools. These models are suited for tasks involving complex reasoning, not just content generation.
Distilled models like DeepSeek were another highlight. These are compressed versions of larger models, faster, cheaper, but still capable of impressive performance if trained and prompted properly. They’re ideal for real-world deployment where cost, latency, and efficiency matter.
We also explored AI Assistants vs AI Agents.
- AI Assistants (like a well-tuned CustomGPT or chatbot) are tailored to specific workflows, designed to respond to user queries, support tasks, or streamline repetitive actions, often integrated into tools like Google Workspace, Canva, or Grammarly.
- AI Agents go further. They are autonomous, goal-oriented systems that can plan and execute multiple steps without needing continuous prompting. Think of an agent that produces a detailed policy brief from scratch, or coordinates a logistical task like ordering supplies. They exhibit traits like planning, decision-making, and interaction with external environments. Examples include research agents, scheduling assistants, and operational bots.
When and how to use GenAI was a great discussion that ended the day. It included a section on privacy which we unanimously care about. This was very helpful.
These agents signal a future where AI is not just reactive, but proactive and embedded in systems, helping organisations think, act, and decide faster, with human oversight.
The final reminders from today: Learning to prompt well isn’t just a trick, it’s a core professional skill.
And as AI becomes more embedded into our daily tools and systems, knowing when to trust it, when to guide it, and when to question it will define how effectively we use it. But use it we must!
About the Podcast Host
Joanne Sweeney is a specialist in public-sector marketing and digital communications. She’s the author of Public Sector Marketing Pro and the creator of the Professional Diploma in AI for Public Sector Marketing. With years of experience in training, consulting, and advocating for digital transformation, Joanne provides insight and guidance for government agencies navigating the digital landscape. She is also the host of the award-winning Public Sector Marketing Show podcast.
