How I explained AI to Private Equity

A framework to get started and creating projects

Hey friends,

The ​Masters of the Universe​ have discovered a new tool in their dealmaking.

Artificial Intelligence. Specifically, ChatGPT pro (aka ​the $200 version​).

How do I know this?

Last Friday, an NYC-based Private Equity firm brought me in for a training session so that they could learn how to best use these powerful tools.

The 75-minute session covered frameworks, prompting advice, using the "Projects" feature, the existence of a “hive mind” and my favorite…

You need to learn Model Jiu-Jitsu.

(They gave me permission to reference this conversation anonymously.)

The G.R.A.I.L. Framework

I began the session with a series of simple principles to show the range of possibilities.

(And BTW, ​ChatGPT is a phenomenal brainstorming partner​ for this.)

Here’s the acronym, GRAIL.

Daft Punk is playing in my house...

Generate

Draft emails, proposals, marketing copy and presentations for your specific audience (and in your style and voice).

Use Cases: Generate an investment committee memo, draft a quarterly report with industry insights.

Research

Identify trends, break-down data, or suggest frameworks for evaluating business decisions—ideal for market analysis or due diligence tasks.

Use Cases: Competitor analysis, pricing strategies, evaluating potential acquisition targets.

Automate

Optimize repetitive tasks by creating custom tools, using integrated coding and workflow automation to handle routine processes (that drain your time).

Use Cases: Auto-tag 100+ pieces of customer feedback, extract key action items from meeting transcripts, build dashboards that categorize spending data.

Ideate (and Brainstorm)

Generate innovative concepts, think "outside the box" and solve complex problems.

Use Cases: Explore unconventional solutions to a recurring bottleneck and stress-test your strategic plan by identifying second order effects.

Learn

You now have a new personalized tutor—ingest complex topics quickly (and based on how you learn best.)

Use Cases: Learn a new industry (e.g. semiconductors, insurance, cleantech) by simulating a back-and-forth Q&A with an expert and role-play business scenarios (e.g. client objections, stakeholder negotiations) to sharpen your communication.

The importance of “model jiu-jitsu”

Repeat after me.

AI is not “Google on Steroids.”

(And it can do wayyyyyy more than crafting and summarizing emails.)

I urged the team to abandon the "Google metaphor" in favor of:

1. Long and detailed prompts that can take 20+ minutes to write

2. Extensive follow-up questions (and asking ChatGPT to ask you clarifying questions)

3. Using the right tool to solve the right problem

Basically, they needed to Learn Model Jiu-Jitsu.

Here’s how people first start using LLMs.

They write a simple prompt, use the default model and once they get their result they’re pissed!

It didn't give me my perfect answer. Sad face emoji. And that must mean:

AI is overhyped!

But here’s what really happens.

• Query Grok or Perplexity to get best practices for your case

• Write up a detailed prompt in Google Doc (so that you can edit or reuse it later)

• Figure out which model to use (potentially, with help from Grok/Perplexity)

• Crap, I need to add more context via files, but the model won’t accept PDFs

• Search for a tool to convert the PDF into a giant block of text

• Once you enter the prompt, ask “Given the task at hand, do you have any clarifying questions for me”

• Realize that the prompt was missing some key context (i.e. you didn’t specify the audience for the output)

• Return to your prompt in your Google Doc and modify it

• Re-run the conversation

• Realize that it’s missing a key conclusion (that you intuitively know)

• Challenge the model about why it made that mistake

• Confirm that the output is correct, but crap, it’s in the wrong format

• Take different sections from the output and paste them into another Google Doc

• Start a new thread with a different model (i.e. one that is better at writing, like ChatGPT 4.5)

• Grab another prompt snippet (from yet, another Google Doc… or better yet, a Text Expander) that has your voice dialed-in

• Receive your final document, perfectly to your liking

That, my friends is Model Jiu-Jitsu. It’s messy. It crosses platforms. It’s highly tailored to your needs and preferences. It requires iteration. And there's no playbook.

And with time, it becomes a part of your muscle memory.

A quick side-bar: The differences between the models

The Masters of the Universe didn’t realize that with one little click, they could open up a world of possibility.

In an overly simplified way, there are 2 types of models: Standard and Reasoning.

Standard "Chat" Models (4o) are work-horses. They’re fast and cheap and apply a brute-force approach to getting you your answer and they’re the default model for this reason.

I think of standard models as Graduate Teaching Assistants. They’re smart, they get the job done and are very accessible. They’re also quite cheap.

Then there’s the Reasoning Models (like o1, o1-mini and o1-pro). You can think of this as meeting with Your Professor.

Your professor is harder to reach and you’re not going to waste their time with a basic question like:

“Can you explain how tariffs work?”

When you get that rare time with a professor, you’re going to really tap into their wisdom and their unique insights. You recognize that they might ramble. Also, since many professors are so specialized, they might not be as good at answering the simple questions.

Continuing with our tariff question, you might ask instead:

“Can you explain if you think the re-shoring of manufacturing in the US is possible given the strength of the dollar, tighter immigration controls and the centralized nature of China’s economy? Argue both the pro and con case.”

You may then notice a little smoke come out of the professor’s ears as their wheels spin to give you this delicious answer.

So what does this mean for you?

Most of the tasks in GRAIL (with the exception of Ideation and to a lesser extent Research) will probably work just fine using a standard Chat model. This is because Generate, Automate and Learn are often “pattern matching” activities.

If I’m honest, I have not found that many use cases in my own work where a Reasoning Model strongly outperforms 4o. Much of the LLM research shows that o1 models dominate in domains like coding, math, science and legal analyses.

However, I have found some instances where they outperform, such as crafting outlines (for presentations, books and websites) using a wide variety of disparate sources and for some business strategy brainstorms.

A Future-Proofing tip for you:

I really encourage you to try using both a Reasoning Model and a Standard model on the same query and then comparing the results.

Over time, you'll start to see different results for your specific use cases.

After all, you gotta develop that muscle memory for Model Jiu Jitsu.

Using the “Projects Feature” to share knowledge

If you’re constantly using ChatGPT, you’ll inevitably run into the challenge of organizing all of your chats, prompts and uploaded files.

This is where Projects become a critical (yet underutilized) part of your workflow.

Let’s say you’re an Investment Analyst responsible for covering five specific companies in the retail sector. You’re expected to know these companies inside-out and will have a bunch of documents associated with each company.

Here’s how to use a project to manage each company:

Step 1. Create the project

In the left side-bar, hit the plus sign to create the project. Name it after the company you’re covering (i.e. Lululemon)

Step 2. Upload the relevant materials

Click on Add Files.

This enables you to upload the following file types.

• Text Files / Google Docs

• PDFs

• Images

• Copy + Pasted information (into a text file)

(Following the Model Jiu-Jitsu Principle, you might need to extract an MD&A section from a 10-K and paste it into a Google Doc. Same goes for that Deep Research query you did on the peer group)

Step 3. Chatting within the project

Now every time you start a chat (inside the Lululemon project) ChatGPT will grab the relevant context from the uploaded files. Behind the scenes, it’s not pasting everything in the context window — it’s likely just grabbing what it deems to be the most relevant information.

It’s important to note that each chat cannot reference the contents of other chats within the project. That cross-pollination is reserved for files you have uploaded.

Does ChatGPT have a "hive mind?"

The last question was around how often ChatGPT "remembers" things. (In AI parlance, persistent memory.)

You may have seen these examples where ChatGPT stares deep into your soul and shocks you with its holistic assessment of you.

This would also mean that for your work, it would understand your industry, decision-making criteria and team dynamics.

First, you can store memories by saying things like:

Remember that I'm an investment analyst focusing on the consumer sector in the US

or

Remember that I manage a team of 10 individuals. I value radical transparency, direct feedback and empathy. My leadership style is strongly influenced by the book Radical Candor.

Once you do this, you'll see a little notification that says "Memory Updated."

You can see all the memories by going to the top right corner and clicking on:

Profile Settings → Personalization → Manage Memories

Gosh, that feels quite vulnerable. (There are about 80 in total.)

I'll let you have a little laugh at my expense.

See you next week!

Khe

PS Earlier in the week, I did a deep-dive on ​how prompting can make you $10,000 per month​. And our ​NotebookLM primer​ went viral on Substack.