What’s This About?
Motivation
I caught myself feeling embarrassed to tell people that I’m interested in LLMs for productivity. That sounds so handwavey!
But alas, I think this area has massive potential, and it’s deeply intriguing on a human level. It’s also oddly under-appreciated in the current LLM hype cycle 🤷‍♂️.
This document attempts to outline my interest. Work in progress!
Hypothesis
LLMs will transform task management more than any technology to date, in a short time (10 years).
I’m 50% on this. The mechanisms seem so obvious, but there are probably obstacles that I don’t see.
Nomenclature
- Task management is just the process of keeping todo lists. This doc is scoped to personal task management, though there is obviously overlap with project management in a team.
- A system is the combination of a task list, calendar, and reference files (notes). You really need all three.
- Getting Things Done (GTD) is a methodology for task management. It’s 20+ years old now and there are some criticisms, but the book and the methodology have been impactful to me. Try to look past the cover.
- Productivity is an amorphous concept of how well you do task management. I dislike this word, but can’t think of a better one.
Productivity Axioms
The concept of productivity is nebulous, so I’m beginning here with some of my personal perspectives on productivity. This section has nothing to do with LLMs.
Having a system is binary
Either you have a system where you can deposit a commitment and forget about it - with confidence it will get done - or you don’t. There is no sorta. This was my main takeaway from GTD.
Systems are highly variable
There are wildly different ways to implement a system.
Some people use a single TODO.txt
file, and some engineers use Github issues + pull requests for replying to texts.
Both are valid.
Productivity maximization is dangerous
If you have a spidey sense about this kind of optimization, that’s totally valid.
There is a distinct psychology when one starts to invest too much in task management systems - you think that all your problems in life can be solved by productivity gains. I think the most elegant summary of this is 80000 hours. This also explains why there has been some legitimate backlash against the GTD mentality.
Productivity systems are part of the human experience
Productivity systems have been around forever, and probably haven’t changed much since the invention of pen and paper. “Beware the barrenness of a busy life” is attributed to Socrates.
Mechanisms
Let’s first start with specific mechanisms by which LLMs can improve task management.
This section is pedantic / you may want to skim.
Automation
LLMs can automate some rote work involved in task management - scheduling appointments, researching things, etc. I put agents like MULTI·ON in this category.
This is the most basic mechanism, and IMO also the lowest value.
Clarify
Conversely, this may be the highest leverage work in a GTD process, and LLMs can excel here.
I define clarify as the process from taking a task from simple note to a task with clear success criteria and next actions.
Imagine you quickly write down a bullet list of 10 things that need to get done. An LLM should be able to convert your scratch list into a clear task list.
New workflows
New task management workflows that aren’t practical without AI. My personal favorite is filling in open time. Suppose you have 10 minutes before your next meeting. An AI assistant can serve you some quick tasks with low cognitive load.
Process management
If you had a program manager for your process, they’d do things like set reminders, update due dates, and surface inconsistencies.
An agent could do much of this work in the background, prompting you for questions when necessary.
Organizing reference data
In GTD, reference data is all the stuff you need to store for fast retrieval later - notes, accounts, contacts, etc. Most people struggle to keep reference data organized.
This could be grouped with process management, but I’m separating out since
Calendar optimization
Given your current tasks and priorities, agents can optimize a schedule.
This mechanism may be the first to reach broad usage, and already supported by some calendar apps e.g. Superhuman. I also think it’s subject to rapidly diminishing returns.
Memory enhancement
Memex is a distinct class of LLM application. It’s probably further out - i.e. not in the next 5 years.
I’m adding here because there may be a subset of memory enhancement that can be used for task management.
Example: suppose Sensitivity analysis for price increases is a task. 6 months ago, you read an article on how to do sensitivity analysis with Python and starred it. This article could be suggested as reference content for the task.
Brainstorming
Some tasks move faster when you brainstorm with another human. A chat agent can possibly replace the other human.
Accountability partner
There are a range of ways that social support can help aid task management.
Some of these can happen with an agent instead of another human - I’m grouping them as an accountability partner.
Narratives
Those mechanisms are pretty dry. These are the higher level stories we can tell about how LLMs will change the nature of task management.
The functional need
It’s obvious that GTD is valuable, but maintaining a GTD process in practice is such high overhead - most real people can’t actually do it. LLMs can eliminate that overhead by automating management processes. That makes GTD finally accessible to real people.
Breadth of the opportunity
Consider how many people in modern society have unwieldy scattered to-do lists, or just a bunch of things in their head that they want to do but won’t. Now think about how many of those tasks could advance, if the human could just find 15 minutes to properly focus on it and identify a clear next action.
That gap is essentially the TAM for LLM task management, and it’s massive! Billions of lives could improve.
Technical fit
Here, we consider why LLMs are actually a great fit for task management:
- Current LLMs were trained on the internet, and thus how to complete this task guides are already in the training data.
- Task management doesn’t require much context. For basically any task on my list, I could explain all the necessary context in a voice note in a few minutes.
- Task management can be fully text-based, and doesn’t require any media.
Part of the plan
Now we get more esoteric. This narrative says that the hypothesis here is obviously true, because AI will guide humans in all walks of life. Productivity is just a small example.
I think this is argued pretty well in Homo Deus, in the parts about Dataism.
Counterarguments
Managing emotions is the real problem
Even with the perfect GTD product, task mangement would still be limited by emotionse e.g. discipline and focus. AI won’t cure your malaise and get you out of bed in the morning.
Assistants will forever feel uncanny
Despite what you see in demos, chat agents are still uncanny and awkward IRL. AI assistants will still be uncomfortable for normal people for a long time.
People should do fewer things instead
This aligns with the message in Four Thousand Weeks.
Landscape
I’m researching this section now. I’m looking for any people, products, companies that are working toward this vision.