The intern and the coach

A recent post from Wharton professor

Ethan Mollick

shared details about an impressive new study done in conjunction with Harvard Business School. The entire post is worth reading:

Centaurs and Cyborgs on the Jagged Frontier

But in a nutshell, the study found that BCG consultants who used GPT-4 were up to 43% more effective at tasks vs. consultants who didn’t. They were more effective along several dimensions, including throughput, speed, and quality:

Consultants using AI finished 12.2% more tasks on average, completed tasks 25.1% more quickly, and produced 40% higher quality results than those without. Those are some very big impacts.

But it also exposed some pitfalls in how we use AI. Namely, workers who got too complacent were at risk of getting things wrong. They “took their hands off the wheel” and didn’t catch GPT-4’s mistakes. And in the study, researchers identified two distinct ways to avoid these downsides: centaurs and cyborgs.

Centaurs (half-human/half-horse) clearly separate user and AI. Centaurs decide and delegate tasks based on whether the human or the AI will be better at them.

Centaur work has a clear line between person and machine, like the clear line between the human torso and horse body of the mythical centaur. Centaurs have a strategic division of labor, switching between AI and human tasks, allocating responsibilities based on the strengths and capabilities of each entity.

Cyborgs (man/machine hybrids) constantly switch between user and AI. They work in real-time with AI as a partner, blurring the line between who is responsible for what.

On the other hand, Cyborgs blend machine and person, integrating the two deeply. Cyborgs don't just delegate tasks; they intertwine their efforts with AI, moving back and forth over the jagged frontier. Bits of tasks get handed to the AI, such as initiating a sentence for the AI to complete, so that Cyborgs find themselves working in tandem with the AI.

Reading the paper made me realize that I’m already using both approaches with ChatGPT, but I’ve been framing them differently. Rather than centaurs and cyborgs, I've been thinking about interns and coaches.

Artwork created with Midjourney.

The intern

When I use AI as an intern, I care about a final product, a deliverable. Often I’m asking for extremely narrow tasks without context, or I’m providing a ton of context to the model to try and get what I want. I’m checking for hallucinations or errors, and I’m planning on editing the final version regardless.

This approach very clearly feels like delegating/managing, like I’m giving instructions and checking someone’s work. And this is, I believe, how most people approach using AI. As a tool to automate white-collar work - where the output is all that matters.

Example: Short writing tasks

Given the following sentences, write a few different short bios/introductions.

• Charlie Guo

• Author of Artificial Ignorance, a newsletter on generative AI projects and news.

• CTO at Crowdmade, an influencer merchandise company.

• Author of Unscalable, a collection of interviews on "doing things that don't scale" with the founders of companies like DoorDash, GitHub and Zenefits.

• Over 15 years of software experience and over 10 years of working in Silicon Valley. Founded multiple startups backed by YCombinator and other prominent investors.

• Graduated with a Computer Science degree from Stanford, with specialties in human-computer interaction and machine learning.

• Lives in the Bay Area.

Write some marketing one-liners for a back-to-school campaign meant to get users to design new merchandise.

Given the resume below, summarize the person’s experience and write a compelling blurb to provide for a job posting.

Example: One-off coding tasks

Write a Django expression to query Users who have more than 5 periods in their email address.

Write a Python script that takes a string and splits it around any anchor tags present in the text.

Summarize the below list of git commits into an outline of the features and bugfixes added to the codebase.

Artwork created with Midjourney.

The coach

When I use AI as a coach, I don’t particularly care about the outcome - rather, the value comes from the process. I want it to help me think through a strategy, or refine my problem-solving approach. Often, I’ll have ChatGPT ask me follow-up questions before answering in order to 1) clarify my own thinking and 2) give better, tailored responses.

This approach feels much, much more collaborative. Clearly, ChatGPT (specifically GPT-4) isn’t a world-class expert - but it doesn’t have to be. It’s a good enough mental sparring partner that I can walk away with far clearer ideas than if I had just tried to think through things myself. And we can keep collaborating and iterating, adding constraints or digging deep into a specific topic, ad nauseam.

Example: Brainstorming paid perks

Let's think through paid perks that I can offer my newsletter subscribers. Before writing a response, ask me relevant questions to help guide your answer.

Of course! Let's dive into understanding your newsletter, audience, and what value you can provide. Here are some questions to get us started:

Content & Audience: What is the primary focus or subject of your newsletter? Who is your target audience?

Current Offerings: What content or features do you currently provide for free? Do you already have any paid offerings, and if so, what are they?

Engagement Metrics: How engaged is your audience? What is the open rate and click-through rate of your newsletters? Are there specific topics or types of content that resonate most with your subscribers?

Resource & Capabilities: What resources (like time, money, software, or other assets) do you have at your disposal to create or offer additional perks?

Feedback: Have your subscribers expressed any wishes or needs? Are there common questions or requests that come up?

Long-Term Goals: Where do you see your newsletter in the next year or two? Do you have any specific business goals associated with it, such as increasing the subscriber count, generating revenue, or promoting other ventures?

Unique Selling Proposition (USP): What makes your newsletter unique or different from others in the same niche? Do you have any special skills, access, or connections that could be leveraged as perks?

Answering these questions will help me tailor suggestions for paid perks that align with your goals and resonate with your audience.

The primary focus is AI news and productivity. Free subscribers currently receive 2 posts per week. There is a paid option that provides access to behind-the-scenes data and code. My open rate is 40%. This is a very new space, and not a lot of newsletters are going in-depth on how products and models work in an accessible way - most are either covering high-level topics or are writing dense, technical articles. I have experience in both computer science and writing, as well as a history of working in startups. In the long term, I want to grow my audience and build a sustainable publication that generates revenue.

Thank you for the detailed information. With your unique background in computer science, writing, and startup experience, combined with your ability to provide in-depth yet accessible insights into AI, you have a solid foundation to offer valuable paid perks. Here are some suggestions based on the information you provided:

Deep Dives: Offer monthly deep-dive articles exclusively for paid subscribers. These articles can dissect a particular AI model, concept, or tool, explaining its intricacies in a digestible manner. This builds on your unique selling proposition of making complex topics accessible.

Exclusive Webinars or Q&A Sessions: Host monthly or quarterly live webinars where you cover a trending topic in AI and productivity. These sessions can also have a Q&A component where subscribers get direct access to ask you questions. This doesn't necessarily require more writing but leverages your expertise.

Curated Resources: Provide paid subscribers with a curated list of the best AI tools, papers, or resources every month. With your background in startups and CS, you likely have insights into the best emerging tools and resources that others might overlook.

Access to an Exclusive Community: Build a closed community forum or group (e.g., Discord, Slack, Circle) where paid subscribers can interact with each other and with you. This becomes a hub for networking, sharing resources, and discussing recent AI developments.

Discounts or Early Access: If you have plans to launch other projects or products in the future, offer your paid subscribers early access or exclusive discounts.

Collaborations & Interviews: Collaborate with other experts in AI and productivity for exclusive interviews or joint articles. This brings fresh perspectives to your subscribers and can also help in expanding your reach.

Templates & Toolkits: Given that you can code but have limited time, occasionally create and share simple templates, scripts, or tools that can help your subscribers in their work related to AI or productivity. This doesn't have to be frequent; even a quarterly toolkit can be valuable.

Job Board or Opportunities: Given the tech-oriented nature of your audience, consider starting a mini job board or listing of opportunities in the AI and productivity space. This can be a win-win, as companies might even pay to list their positions, and your subscribers get access to unique opportunities.

Feedback Loop: Regularly seek feedback from your paid subscribers about the kind of content or perks they'd like to see. This not only tailors your offerings to their needs but also makes them feel valued.

Archived Content Access: If you've been running your newsletter for a while, consider putting some of your older, high-value posts behind a paywall. New paid subscribers get access to this archive as part of their subscription.

With your unique position in the market and a solid open rate, you have a great opportunity to offer valuable perks that will not only retain your current subscribers but also attract new ones. Remember to continually evaluate and iterate based on subscriber feedback and the evolving landscape of AI and productivity.

Example: Launching a podcast

I want to start a podcast, and I want you to help me think through the process. First question - what equipment do I need?

Starting a podcast is an exciting endeavor! Let's go step-by-step through the equipment you might need. Keep in mind that you can start with the basics and then upgrade as your podcast grows or as your budget allows.

1. Microphone:

Beginner: USB microphones (e.g., Audio-Technica ATR2100x-USB, Blue Yeti) are an affordable and easy way to get started.
Intermediate/Advanced: XLR microphones (e.g., Shure SM7B, Rode NT1-A) provide better sound quality but require additional equipment.

2. Headphones:

A good pair of closed-back headphones (e.g., Audio-Technica ATH-M50x, Sony MDR-7506) is essential for monitoring your voice and ensuring there's no background noise.

[...]

Final Thoughts: Remember, the most important aspect of your podcast is the content and your delivery. While high-quality equipment can enhance the listening experience, it doesn't replace engaging, well-researched, and well-presented content. Start with what you can afford, and upgrade as you go!

Got it. I'm going to be recording these interviews remotely, so I'm guessing I won't need the full studio setup.

Absolutely! If you're conducting remote interviews, the process can be simplified, and the equipment needs can change a bit. Here are the essentials for remote podcasting:

[...]

Tips for Remote Interviews:

Sound Check: Before starting the interview, conduct a brief sound check to ensure both your and your guest's audio are clear.

Wi-Fi: A stable internet connection is crucial. Encourage your guest to be on a stable connection too. If possible, use a wired connection.

Instructions for Guests: Send your guests a guideline on how to find a quiet recording environment and ask them to use headphones. If they have an external microphone, even better!

Backup: Always record a local backup. Ask your guest if they could do the same. This way, if there's an issue with the online recording, you have a fallback.

Remote interviews are convenient and allow you to tap into guests from all over the world. With the right preparation, you can achieve high-quality recordings that rival in-person interviews.

Let's dig deeper into the structure - what are some common elements of a podcast structure? I'm thinking about things like:

- A welcome message for viewers
- An introduction
- Asking the guest where to find them on social
- An outro

- What else am I missing?

Absolutely! Structuring your podcast with clear segments helps guide your listeners and offers a more polished, professional experience. Here's a breakdown of common elements in a podcast structure, along with some additional segments you might consider:

1. Pre-Roll (Optional):
- This is a short sponsored message or ad that plays before the episode starts.
- Some podcasts also use this time for a brief teaser of the episode’s content to pique the listener’s interest.

2. Welcome Message:
- Greet your listeners and introduce yourself (especially important for new listeners who might not be familiar with you).
- Briefly describe the podcast’s general theme or mission.

3. Introduction:
- Introduce the episode's topic or guest.
- Provide some background or context to set the stage. For instance, why did you choose this guest? Why is the topic important?

4. Musical Interlude (Optional):
- A short musical break can transition from your introduction to the main content. It acts as an auditory signal of the shift.
- Ensure you have the rights to use any music on your podcast.

5. Main Content/Interview:
- This is the bulk of your episode.
- If interviewing, after your initial questions, allow the conversation to flow naturally, but keep your structure in mind to ensure you cover key points.

6. Prompt for Social Media or Other Promotions:

- Midway or towards the end of the episode, you can ask your guest where listeners can find them online, about any projects they're working on, or if they have anything they'd like to promote.
- This segment can also be used for your own promotions or announcements.

7. Outro:

- Sum up the episode briefly.
- Thank listeners for tuning in.
- Provide a call to action: encourage listeners to subscribe, leave a review, join a mailing list, etc.

8. Post-Roll (Optional):
- Like the pre-roll, this is another spot for sponsored messages or ads. Some podcasters also use this slot for special announcements.

9. Credits:
- Acknowledge anyone involved in the production.
- Mention any music or sound effects you used and ensure you're giving proper attribution.

10. Call-to-Action Reminder (Optional):
- If there's a specific action you'd like your listeners to take (e.g., supporting a Patreon, filling out a survey), you can reiterate this after the main content.

11. Bonus Content or Teaser for Next Episode (Optional):
- Provide a sneak peek of your next episode to entice listeners to return.
- Alternatively, you could include a short bonus segment related to the episode's topic, like listener questions or fun facts.

12. Outtakes (Optional):
- If it fits the tone of your podcast, you could include funny outtakes or bloopers at the very end, after the official outro. This can add a personal touch and humor to your show.

Remember, while these elements provide a template, the structure of your podcast should resonate with your style and your audience's preferences. As you progress, you can adjust based on feedback and what feels right for your show.

The hybrid

What might be the most valuable aspect of using AI is that these two roles - the intern and the coach - can grow, or even merge. Normally, we wouldn’t ask our interns to coach us, or vice versa. But ChatGPT can seamlessly switch between the two. And rather than “interns” or “coaches,” we may come to think of AI more generally as “engines for the mind” - tools that amplify, refine, and work alongside us.

But while the BCG consultants got better at their jobs when paired with GPT-4, we must remember not to take ChatGPT’s answers for granted. 95% correct automation is more dangerous than 25% correct automation - because it lulls you into a false sense of security.

Of course, no one knows the limits of what LLMs are capable of - not even their creators. And in a world where technology and human creativity merge, we’re going to be constantly evolving, adapting, and improving our strategies. Whether we align with centaurs, cyborgs, or somewhere in between, the potential benefits are vast - as long as we don’t get complacent in our relationship with AI.