🐙 Promptpack: How to build a second-brain (featuring AI)

Hi, it's Azeem.

This week my colleague Chantal shared how she uses generative AI to build a second brain and augment her research skills. This was originally published in Exponential View. Enjoy!

Hi,  Chantal Smith here.

In this Promptpack, we’ll take a look at how to set up a knowledge base, aka your second brain. 

We’re a little more than half a year into the generative AI revolution. Along with millions across the world, we’ve tested and implemented new tools and practices. We’ve shown you how to get started with quantitative analysis using OpenAI’s Code Interpreter. We've shared our favourite prompts to turn ChatGPT into a thought partner. This time, we’ll go into how we use AI-powered tools to research, process, and store knowledge.

As Microsoft CEO Satya Nadella said last week, LLMs make natural language the primary interface with computers, and add a layer of reasoning over data. We increasingly see this in our work, which for a lot of the time is about finding clues, building knowledge and making sense of it.

The process of research, whether in an academic, professional, or personal setting, is a constant interaction between resources, tools, and the data landscape. In this Promptpack, I will share:

• How I use Notion and its AI capabilities as a smart knowledge repository;
• How I use Perplexity and Elicit to augment search;
• Why ChatGPT remains my favourite AI tool.  

AI-powered (re)search

For most people, the way to find information is to google it. Google is the internet’s most popular website, and was visited 84.6bn times in June 2023 alone. You search it using keywords. But the new way of searching is to use AI, both its natural language capacity and its increased ability to filter through knowledge. 

I tend to use two different but complementary tools, one designed for academic research, the other a more general search engine. 

For academic research

Elicit is a research assistant using language models to automate parts of researchers’ workflows. It’s driven by research questions to which it responds with a list of relevant sources, and a short literature review created from the four most relevant articles. Think of it as Google Scholar that understands natural language and summarises the conclusions of the top articles. The advantages of it understanding natural language is that it will show you papers within the theme you’re researching, even if it doesn’t include the exact keywords you searched for. 

Where Elicit really excels is in biomedical research. It makes meta-analysis and other cross-research surveys really easy to undertake. So worth trying those fields if that is your bag.

But for now, let’s ask “What makes an innovation ecosystem successful?”. 

From this result, I might iterate in multiple ways. 

• The top result is a little old, therefore I will filter out articles published before 2020 thanks to the “filter” button.
• I could focus my research question more on the aspects I’m curious about. For example, “What kind of social relationships are necessary in a successful innovation ecosystem?”
• If I am more interested in the types of approaches I could use to answer my question, I might ask “What theories help explain the success of innovation ecosystems?”

The results may not always be relevant, but I find Elicit useful to get a fast overview of the state of the academic literature for a question. 

For more general search

Perplexity.ai is closer to a general search engine, as it will prioritise more mainstream sources of information, such as major media outlets. It has a chat interface, which enables more interaction and iteration than Elicit. Of the browsing chatbots that I’ve tried, Perplexity has proven most sensible and nicest to use, mainly because it offers the sources upfront, at the top of the page.

Below, I asked Perplexity the same question as Elicit: What makes an innovation ecosystem successful?

I was intrigued by the social aspects in Perplexity’s answer, so I followed up on them with “Tell me more about the people and culture parts, and focus on relationships.”

To go further, I might:

• Ask Perplexity to “Give me a list of experts and sources on the question of relationships in innovation ecosystems”.
• Ask “Is this a relevant subject, or should I reframe it?” for it to explain to me why my question matters. It’s worth noting that it tends to agree and can lack criticality.

Thanks to these two tools, you often get a thorough overview of a subject and the relevant data sources to further explore it, both from a strictly academic and a more mainstream point of view. Now, let’s explore the AI-enhanced way of storing this information. 

AI-enhanced knowledge storage

Summarise in an easy-to-read way 

Notion is a productivity and note-taking app for organisations that is well suited for amassing knowledge. It lets you build databases in which each page has a source of information, an insight, or even a chart. That’s not unusual, but what distinguishes it as a brain-builder is Notion AI. It does all the same things as other AI writing assistants, like brainstorming ideas and creating social media posts, press releases, poems, or even pros and cons lists. But on top of that it has an autofill feature that lets you apply an AI command across the different pages of your database. 

Let’s start at the beginning. You can duplicate the Exponential View template here, and adapt it to your own needs. Let’s go through how it works.

It is a database, with two kinds of sources as examples: one is an article from The Guardian, another is a table of global EV penetration. These sources are pages in the database, with content (what the source says, that you copy-paste into the page) and properties (the link to the article and our AI magic).

When I go through my sources, I often want to know at a glance (1) what the source contains, and (2) why it matters to me (in this case, I care about the relevance to technology). To implement this, I have created two “AI auto-fill properties”, that each apply the same prompt to all pages that I will put in this database. Here are the prompts:

1. Summarise in 5 bullet points.
2. Explain in 1-2 sentences why this is relevant for technology.

In a video you can access here, I show how I created the page and applied these prompts.

As I show in the screenshot below, these prompts also work on quantitative data. Simply insert it as a table. However, I would be extra careful on the AI’s interpretations of numbers, especially since Notion AI wasn’t specifically trained for quantitative analysis. 

As a result, you have a database in which you can store knowledge either as words or numbers, and a view of what it is about and why it is relevant at a glance. I’m impressed by the quality of the output, too. Although the AI always risks hallucinating facts and misunderstanding data, it is correct often enough for it to be a great time saver.

Go further: Transform data into knowledge

There are already many ways to make sense of data and transform it into knowledge that is useful to you. New AI tools pop up every day. However, I always return to ChatGPT for two reasons. First, it is the tool most widely experimented with, so there’s a lot of collective intelligence and learning available to make the most of it. Second, it remains one of the highest performing and useful LLMs out there, even as the quality of its outputs has varied since its inception. However, chatbots are really a personal choice. While I (Chantal) prefer ChatGPT, 

Azeem Azhar  swears by Claude. For qualitative analysis, I would recommend checking out our first Promptpack, which suggests ways of using ChatGPT to help you think. It includes prompts to (1) find connections, (2) apply frameworks, (3) explore scenarios, and (4) question assumptions. In a nutshell, AI functions here as a brainstorming partner.

For quantitative analysis, you can use OpenAI’s recently introduced Code Interpreter. It does a brilliant job to help with quantitative analysis, as Nathan Warren shows in this second Promptpack. Here, AI acts as a slightly unpredictable junior data analyst with a great memory. It is less reliable than more straightforward statistical software or Excel functions, but it can do many tasks and then critique its own analysis. 

A special mention to EV member Gianni Giacomelli and his team at MIT who have opened the beta to Supermind Ideator, an AI-assisted tool to help with idea generation. What makes it particularly innovative is that it proposes a series of ideas based on different logics, for example how a market works, how a community works, or how democratic decision-making works. 

As you experiment with generative AI, it’s important to remember that these tools are far from perfect: they frequently hallucinate, so never take their output as gospel.

I would love to hear and learn from you, dear members. 

• What are your favourite generative AI tools?
• Is there any advice you’d like to share with the rest of the community?