Rob Jefferies

In 2018 I started a journey to incorporate gratitude into my daily life. To see the small positives around me. To focus less on the negatives that felt like an albatross hanging around my neck when it was just a small inconvenience of life.

This got me thinking the other day, I feel like we are living in the “Good Old Days” of AI. The early days before “Enshittification” as defined by Cory Doctorow in his book Chokepoint Capitalism.

Enshittification as defined by Claude.ai

Stage 1 - User Acquisition: Platforms start by being generous to users, offering excellent service, useful features, and good user experience to build a large user base. They operate at a loss or minimal profit to grow quickly.

Stage 2 - Business Customer Focus: Once they’ve locked in users, platforms shift focus to extracting value from business customers (advertisers, sellers, content creators). The user experience begins to degrade as the platform prioritizes revenue from these business relationships over user satisfaction.

Stage 3 - Self-Extraction: Finally, platforms extract maximum value for themselves and their shareholders, squeezing both users and business customers. Quality declines significantly as the platform prioritizes short-term profits over long-term sustainability.

I am a user of multiple AI platforms: ChatGPT, Claude, Gemini, and Copilot. They all seem magical when you put the effort in. They are helpful and patient. They aren’t trying to make me buy anything. They aren’t pushing an overtly biased agenda. They aren’t sending notifications asking me to come back to the app. I’m not following other users. There aren’t algorithmic feeds when I open the app. The switching costs are low. It’s just a useful tool.

But I can see where this is going. In the future these platforms will try to learn more about me. Keep those details in their “memory” and make my experience richer across interfaces (apps, speakers, voice, etc.). They will find a way to lock me in by being useful.

Soon we will be in Stage 2. I don’t know when but it will come. So right now I’m enjoying this time. I’m being more productive. I’m learning to have a childlike inquisitiveness again.

I’m grateful for Stage 1.

Great post highlighting the ups and downs of the recent Tesla Robotaxi rollout in Austin

The launch of Tesla’s robotaxi service in Austin is a major step toward full autonomy. But the Austin launch also makes it clear that Tesla hasn’t discovered an alternative path for testing and deploying driverless vehicles. Instead, Tesla is following the same basic deployment strategy Waymo pioneered five to seven years ago.

What I learned watching 78 videos from Tesla’s Austin robotaxis

Dwarkesh Patel recently shared his thoughts in a video about the design and operation of a company run entirely by AI employees.

In the video, he explores the idea of a single dominant company emerging, scaling rapidly, and outcompeting others by using superior algorithms to process ideas and manage workloads.

This raises some interesting questions. If the goal was to create a society of “haves” and “have-nots,” one way to achieve this would be to keep such a company private for an extended period. Imagine a future “Amazon” that eliminates white-collar jobs, dominates its market, but whose growth remains inaccessible to the general public. What if all the wealth generated is held exclusively by private investors?

Could we see regulations requiring AI-driven companies to become publicly traded, or to otherwise share the benefits of their wealth creation more broadly?

Google has integrated its Gemini AI model directly into formulas in Sheets.

There’s now an =AI() formula that lets you enter a prompt and apply it to a specific cell—very cool. Here’s a video showing how it works: YouTube - # Master 85% of Google Gemini in 12 Minutes (2025) by Jeff Su

Chat based LLMs like ChatGPT and Google Gemini reward the curious, much more than traditional search engines.

Let’s say you are looking to purchase an air purifier for your home. Google Search is not built for people to be extremely curious. Typically Google Search is a one shot tool. You may type “What’s the best air purifier for a single family home of less than 2500 square feet.” But what comes next is a research project. Clicking on the links, reading the click bait ad pages, trying to navigate and remember the different pros and cons and manually comparing and making notes.

LLMs are different. LLMs reward curiosity.

Today I researched a home air purifier using ChatGPT 4o. The experience was very different than a typical Google Search. Here are the questions I asked:

• I’m trying to find the best air purifier for my home.  Something that is free standing that can be plugged into the wall.  What should I take into consideration when I look at options?
• I’ve got a 4 bedroom that is a total of 2200 sq feet.  It’s a semi-open floor plan where the kitchen, living room, dining room, and entry are fairly open and connected. The 4 bedrooms average about 15x15 feet.   What are good options for me.
• How much space needs to be given around these air purifiers? Can they be tucked behind a couch or in a corner?
• Can you get too much air purifier for a space, and would that have any negative consequences?
• Can you perform a deep comparison between the Coway and Alen, including a 5-year total cost of ownership?
• Are there any mistakes that people make when purchasing these filters?
• What other common mistakes or regrets do customers typically make that can be easily avoided?
• Why are people most unhappy with the model XXXX?

LLMs like ChatGPT is much more powerful when you think of it as an advisor and not just an answer machine.

I know there are clearly some things that I have missed and deeper ways I could analyze this problem, but I hope this provides an example of how the power of generative AI really comes from being more curious.

One powerful yet often overlooked application of generative AI is creating test data samples. Instead of manually crafting data sets, you can leverage tools like ChatGPT to generate realistic sample data quickly and efficiently.

This approach is especially useful when you’re learning frameworks like Python’s pandas and need a structured dataset to experiment with. By using a well-crafted prompt (or even not well crafted), you can produce diverse data samples tailored to your learning objectives.

For example, here’s a prompt I used with ChatGPT 4o to generate sample data:

I want to generate code for a test python DataFrame. The table has two columns Flavor and Scoops. The flavor column is persons favorite flavor. Number of scoops should stay between 1 and 3 scoops. Create 5 common flavors to be used for the test sample. Create 30 examples in the dataframe.

And this was the result.

import pandas as pd
import random

# Define common flavors and scoop range
flavors = ["Vanilla", "Chocolate", "Strawberry", "Mint Chocolate Chip", "Cookie Dough"]

# Generate sample data
data = {
    "Flavor": [random.choice(flavors) for _ in range(30)],
    "Scoops": [random.randint(1, 3) for _ in range(30)]
}

# Create DataFrame
df = pd.DataFrame(data)

Great example of using ChatGPT Deep Research for understanding zoning laws by state. Deep Research: Supermajority laws around the states

A friend’s son is applying to a private high school, and as part of the application process, the school requests links to any social media accounts he owns. I’ve heard of similar requirements in sorority recruitment and other selective social organizations.

With AI, it wouldn’t be difficult to analyze social media for potentially negative traits using image recognition models. These models could be trained to detect behaviors such as smoking, drinking, or making offensive hand gestures. For example, a student holding a red Solo cup in a photo could be flagged as a potential concern.

Beyond identifying negative behaviors, AI could also assess social tendencies. It could analyze how many photos feature the applicant alone versus in a group, or how many depict them playing sports, which might indicate teamwork and social engagement. This type of analysis could influence how schools or organizations evaluate an applicant’s personality and fit.