The Slow Death of the project manager

and AI can't play Pokemon.

April 04, 2025

“The Only Thing That Is Constant Is Change -”― Heraclitus

In today’s newsletter:

AI kills the project manager
Reducing cost to train LLM with distributed models
AI doesn't understand Pokemon
Build something with AI this weekend

Pense

And so we're off. I've come out of the closet! ( not in that way) The startup closet is where you go into a dark, hidey hole to get a project developed. It's built, but it needs some beautification. But that's not the point of this pense today. No, I'll mention more about that later. The point is that I couldn't have done it without AI.

Traditionally, startups are built with a small group of friends who set out to work on an idea in a garage, their parents' basements or couches, and so on. Not me. I don't have a network of people who enjoy wonder, adventure, or excitement in that way. So, I found interns—students who can do the work but need some more direction.

The first attempt at this didn't go so well. And that's because I didn't use AI. See after some time, I discovered that AI could take my zoom calls, transcribe them, and created detailed PRDs, tasks, and knowledge docs in minutes.

There was no more trying to get things out of my head and on to paper. I could just talk and ramble to myself or others, and AI would sort through the noise for the vital pieces.

Then the outputs could be added to a task management tool or an AI-generated document for review. Now, though, now, I found that the project manager role is killer. Startups are building tools to make this automatic. And if they can build it this quickly, I can create an entire set of custom internal tools in a few weeks.

Is anyone else's mind blown??? This is wild. Sure, large corps will keep their project managers around for a while, but it is not a field I would tell a new grad to go into anymore. Job security in about 20-30 years is highly suspect.

Now, I'd actually have them consider gaining a trade, craft, or knowledge skill. Because if a project manager is not a few dollars a month, then a solo entrepreneur will be able to act like a team of 10 with the proper setup. Can't write your book? Record yourself talking about it and upload it to the AI. Want to start a newsletter? Have the AI summarize incoming RSS feeds to create first draft runs. I want to be a Youtuber, feed the research to AI, and have it develop a video script. Send it to be chomped into time block tasks on your behalf. Want to make fashion designs? Train your AI to create tech packs so you don't need to hire someone.

Blood, blood, blood in the streets. Sorry, working on startups put me in my villain era.

This is fantastic and scary at the same time—it's kind of like robot bugs. Please don't make robot bugs.

Article

Now, since LLMs and AI agents are the foundation of my business strategy, I decided to see what new research has come out.

Amazon Science doesn't disappoint. A team recently published an article titled "Training large language models more efficiently." The article discusses reducing the resources required to maintain and improve the performance of LLMs. The authors discuss an approach called Distribution-Edited Models (DEM).

Here are the key points:

The Problem: Traditional LLM training requires multiple iterations with different data distribution mixes, which is computationally expensive and inflexible.
The Solution: The DEM approach trains separate models on different datasets and then merges them using weighted distribution vectors.
Validation: The approach was tested on popular benchmarks, including MMLU, BBH, and HELM, across various domains, such as MathQA and Chain-of-Thought reasoning.
Recognition: The method won a special "efficient modeling, training, and inference" award at the EMNLP 2024 conference.