Please Please Me AI
A simple framework for adopting AI
Why Bother With a Framework?
We’re at the frontier. Humans have never seen anything like AI before. It’s new and strange, we don’t really understand how it works (even the people building it), and hundreds of millions of people use AI every day. But where are we going?
Whether you’re adopting AI as an individual, team, or organization, using simple categories can help to avoid getting stuck in endless play.
Categories
Explorations are open questions: Will I like playing piano?
Experiments are assertions with a testable hypothesis: I met a bloke named John Lennon who likes the same music as me. I bet we can write some good songs together.
Implementations are when you take the results of a set of related experiments and turn it into something lasting: John and I have written about 50 songs and decided to start a band. We’re calling it The Beatles. Let’s see how this goes.
What separates these categories is the level of commitment and the degree of curiosity. Generally speaking, broad curiosity and lower commitment characterize explorations, whereas implementations have focused curiosity and high commitment
Exploration
Getting started with AI is more like learning an instrument than learning a traditional piece of software. There’s no manual explaining everything so you just have to get in there and see what’s possible.
Playtime is crucial for developing intuition and intuition leads to more focused questions, i.e. experimenting.
Principles:
Learn by doing. Share with others.
Focus on observations more than goals.
Play with different models (e.g. GPT 4 vs o3) and tools (e.g. ChatGPT, Perplexity) to get a sense of similarities and differences.
Improve how you communicate with AI by reading OpenAI and Anthropic best practices.
Be smart about explorations: Don’t use sensitive data without knowing what happens when you do. Watch out for big bills. Don’t violate terms of use.
If you think you should have started last year, start today. It's fine.
Don’t stop exploring. AI is evolving fast.
Eventually, some of your play will lead to patterns worth testing.
Experimentation
Once you get a feel for what’s possible you'll naturally start making predictions about what will happen, I believe that if I give my AI access to computer logs, then I will be able to troubleshoot problems faster. That's experimenting.
An effective experiment statement should include:
A clear hypothesis – what you believe will happen.
An intervention or change – what you’re testing.
A measurable outcome – what success looks like.
A time frame – when or how long the test will run.
“I believe that [intervention] will result in [measurable outcome], and we’ll test this by [method] over [time frame].”
Example of a small experiment, i.e. short and cheap:
“I believe that creating an AI agent that summarizes the latest Deep Learning articles and converts them into a podcast will save me one hour per day. I’ll test this by building a simple tool, using it during my commute for one month, and tracking time saved compared to my current reading habits.”
Example of a medium experiment, i.e. several weeks with a budget:
“We believe that extending the podcast agent to cover all topics within our research department’s scope will save the team an average of 50 hours per week and surface at least one new insight monthly. We’ll test this over six months by comparing time spent and insights recorded with and without the agent.”
Principles:
Don't wait for others to run experiments for you. There's nothing fancy about an experiment.
Experiment in areas you know and partner with others for unfamiliar areas.
Write down your experiments, the hypothesis, results, and the date you ran it. This helps you track progress.
Rerun experiments when models, tools, or prompts change.
Group related experiments. This helps in deciding what to implement.
Watch out for false fails. If you believe in an idea, don’t give up after one bad result. Maybe the setup was off or you measured the wrong thing. The Beatles got rejected by several record companies before finding the right match.
When experiments prove out consistently and the stakes get higher, it’s time to make a bigger bet.
Implementation
Implementations can be seen as bigger, longer, fuzzier experiments. Democracy is both an experiment and an implementation….but it’s probably more useful to think about it as the latter.
When a commitment becomes substantial enough to need long-term investment, consider it an implementation. Characteristics of an implementation include things like recurring spend, supporting services, progress reviews, etc.
For example, you might run an experiment over a year to compare different investment strategies. You tally the results at the end of the year and decide to commit $10,000 to the best one. This is the implementation.
A bigger example is creating a new company: “We believe that there’s a market for our custom podcast agent so we’re going to invest $500K to hire engineers and create marketing campaigns. Our goal is to test product-market fit by releasing v1 in Q1 and tracking month-over-month user growth, average time used, and churn over the next year in order to guide our next investment decision.”
Principles:
Implementations have risks. Know what they are and work to reduce them.
Resource implementations appropriately.
Set ground rules for how an implementation will be governed.
Measure what matters and take action based on the data.
Treat implementations as living systems. Your first version is a starting point, not the end. Adapt as you learn.
Final Thought
When viewed from a high level, figuring out how to weave AI into our lives isn’t that different from other things we do. This simple approach, while not novel by any stretch, can help by reframing today’s strange AI frontier as something familiar.
Rule of thumb: Explore constantly, experiment often, and implement the few ideas you truly believe in.
You just might find your Beatles.