microdosing not getting a big task done by not doing the smaller, otherwise attainable subtasks
September 9, 2026 at 3:12 AM
microdosing not getting a big task done by not doing the smaller, otherwise attainable subtasks
Since LLM's lack teleological thinking, attempting to outsource your own makes yours worse.
Your tasks and subtasks no longer have any correlation with your goal.
Your tasks and subtasks no longer have any correlation with your goal.
July 10, 2025 at 8:48 PM
Since LLM's lack teleological thinking, attempting to outsource your own makes yours worse.
Your tasks and subtasks no longer have any correlation with your goal.
Your tasks and subtasks no longer have any correlation with your goal.
The task will expand to fill available space.
You will perceive yourself as being busy and active the whole time, performing the task.
Certain subtasks will bloom into their own subtasks.
The yak will get shaved.
www.hanselman.com/blog/yak-sha...
You will perceive yourself as being busy and active the whole time, performing the task.
Certain subtasks will bloom into their own subtasks.
The yak will get shaved.
www.hanselman.com/blog/yak-sha...
Yak Shaving Defined - I'll get that done, as soon as I shave this yak.
I've used the term Yak Shaving for years. You're probably shaving yaks at work ...
www.hanselman.com
July 10, 2025 at 9:02 PM
The task will expand to fill available space.
You will perceive yourself as being busy and active the whole time, performing the task.
Certain subtasks will bloom into their own subtasks.
The yak will get shaved.
www.hanselman.com/blog/yak-sha...
You will perceive yourself as being busy and active the whole time, performing the task.
Certain subtasks will bloom into their own subtasks.
The yak will get shaved.
www.hanselman.com/blog/yak-sha...
Machines can't translate. Translation is not one activity. It involves a set of subtasks. Gregory M. Shreve from Kent State University sums them up like this:
March 14, 2025 at 8:58 AM
Machines can't translate. Translation is not one activity. It involves a set of subtasks. Gregory M. Shreve from Kent State University sums them up like this:
OK... I spoke too soon.
On Max 5x, running a long-running task with multiple subtasks, I’m already running out of usage.
I guess I’ll be upgrading to 20x tonight.
On Max 5x, running a long-running task with multiple subtasks, I’m already running out of usage.
I guess I’ll be upgrading to 20x tonight.
My review of Claude Opus 5.5:
It passes the five hour limit test. I can use it for five hours without running out of usage.
That’s it. That’s the review.
It passes the five hour limit test. I can use it for five hours without running out of usage.
That’s it. That’s the review.
September 23, 2026 at 9:28 PM
OK... I spoke too soon.
On Max 5x, running a long-running task with multiple subtasks, I’m already running out of usage.
I guess I’ll be upgrading to 20x tonight.
On Max 5x, running a long-running task with multiple subtasks, I’m already running out of usage.
I guess I’ll be upgrading to 20x tonight.
Broke: subtasks
Woke: plans within plans.
Woke: plans within plans.
December 28, 2024 at 8:34 PM
Broke: subtasks
Woke: plans within plans.
Woke: plans within plans.
Tasks multiply, each becoming a fractal blossom of subtasks that must be checked for being utter fucking nonsense and then corrected if so.
And it's all random, all black box: impossible to develop a coping strategy and heuristic to deal with. Any given subtask might turn into its own clusterfuck
And it's all random, all black box: impossible to develop a coping strategy and heuristic to deal with. Any given subtask might turn into its own clusterfuck
May 26, 2025 at 8:16 AM
Tasks multiply, each becoming a fractal blossom of subtasks that must be checked for being utter fucking nonsense and then corrected if so.
And it's all random, all black box: impossible to develop a coping strategy and heuristic to deal with. Any given subtask might turn into its own clusterfuck
And it's all random, all black box: impossible to develop a coping strategy and heuristic to deal with. Any given subtask might turn into its own clusterfuck
Why don’t CEI tasks ever get lost?
Because they always cum full circle. 😉
#domhumor #cei #findom #femdom #kinkjokes #subtasks #domdadjoke
Because they always cum full circle. 😉
#domhumor #cei #findom #femdom #kinkjokes #subtasks #domdadjoke
January 7, 2025 at 1:51 PM
Why don’t CEI tasks ever get lost?
Because they always cum full circle. 😉
#domhumor #cei #findom #femdom #kinkjokes #subtasks #domdadjoke
Because they always cum full circle. 😉
#domhumor #cei #findom #femdom #kinkjokes #subtasks #domdadjoke
Suppose you need to clean a room.
Anyone with A(u)DHD can tell you: this is not a realistic task. It's composed of lots of smaller tasks, very specific subtasks.
Ones which can feel embarrassingly small, and demotivating when compared to the state of the room.
Anyone with A(u)DHD can tell you: this is not a realistic task. It's composed of lots of smaller tasks, very specific subtasks.
Ones which can feel embarrassingly small, and demotivating when compared to the state of the room.
A cool audhd thing to do is make a lifescript to randomly pick shit for you to do: randomly spit out one of the books you're actively reading, things to study or a project you're working on, a room to clean, dish to make, etc.
Highly recommend outsourcing your executive function to dicerolls.
Highly recommend outsourcing your executive function to dicerolls.
January 16, 2026 at 3:23 PM
Suppose you need to clean a room.
Anyone with A(u)DHD can tell you: this is not a realistic task. It's composed of lots of smaller tasks, very specific subtasks.
Ones which can feel embarrassingly small, and demotivating when compared to the state of the room.
Anyone with A(u)DHD can tell you: this is not a realistic task. It's composed of lots of smaller tasks, very specific subtasks.
Ones which can feel embarrassingly small, and demotivating when compared to the state of the room.
None of these subtasks are done for their own sake, but for an overall goal.
You're probably not going to go off on a tangent of mowing the yard halfway through, because the goal is a pot of fresh coffee.
You're probably not going to go off on a tangent of mowing the yard halfway through, because the goal is a pot of fresh coffee.
July 10, 2025 at 8:42 PM
None of these subtasks are done for their own sake, but for an overall goal.
You're probably not going to go off on a tangent of mowing the yard halfway through, because the goal is a pot of fresh coffee.
You're probably not going to go off on a tangent of mowing the yard halfway through, because the goal is a pot of fresh coffee.
Cosign. The agentic turn makes "AI" itself almost transparent. The thing that now decides whether people perceive this as empowering is just: whether they believe they could benefit by breaking their work into well-defined subtasks and delegating to 10,000 helpers. Many ppl see no way to do that.
I think a lot of it is not so much a knowledge gap of what AI can do but more a knowledge gap of what computers can do.
more than anything, i think we’re all learning that people who spent their professional life unable to talk to computers are completely unprepared for a situation in which they can now also talk to computers
September 21, 2026 at 1:16 AM
Cosign. The agentic turn makes "AI" itself almost transparent. The thing that now decides whether people perceive this as empowering is just: whether they believe they could benefit by breaking their work into well-defined subtasks and delegating to 10,000 helpers. Many ppl see no way to do that.
Suppose you want to brew a pot of coffee.
This is a task that is composed of several subtasks, each subtask contributes to the goal of having a freshly brewed pot of coffee.
You need to engage in teleological thinking to do this: understand what the goal is, and what actions to take to achieve it.
This is a task that is composed of several subtasks, each subtask contributes to the goal of having a freshly brewed pot of coffee.
You need to engage in teleological thinking to do this: understand what the goal is, and what actions to take to achieve it.
July 10, 2025 at 8:39 PM
Suppose you want to brew a pot of coffee.
This is a task that is composed of several subtasks, each subtask contributes to the goal of having a freshly brewed pot of coffee.
You need to engage in teleological thinking to do this: understand what the goal is, and what actions to take to achieve it.
This is a task that is composed of several subtasks, each subtask contributes to the goal of having a freshly brewed pot of coffee.
You need to engage in teleological thinking to do this: understand what the goal is, and what actions to take to achieve it.
As I told Spouse this morning, anyone using these AIs who has ever said 'but it's just one simple task' to someone with any condition that makes brain processing not effort-cheap is about to learn how quickly the invisible background processing subtasks for Just One Simple Task can add up.
June 1, 2026 at 4:27 PM
As I told Spouse this morning, anyone using these AIs who has ever said 'but it's just one simple task' to someone with any condition that makes brain processing not effort-cheap is about to learn how quickly the invisible background processing subtasks for Just One Simple Task can add up.
this is the category error all the AI maximalists make
if I had to break down what I think the impact of AI would be by field, I'd have to first talk about which *subtasks* it's helping people with
the amount of variability by field and task is MASSIVE, ranges from 90% takeover to 0%
if I had to break down what I think the impact of AI would be by field, I'd have to first talk about which *subtasks* it's helping people with
the amount of variability by field and task is MASSIVE, ranges from 90% takeover to 0%
Saw a thoughtful thread about AI, don't want to QT or argue. But. The biggest rage factor with LLMs is the people who, because genAI is transformative for coding, think it's transformative for everything else, because they devalue every other form of work and labor and knowledge.
March 4, 2026 at 4:55 PM
this is the category error all the AI maximalists make
if I had to break down what I think the impact of AI would be by field, I'd have to first talk about which *subtasks* it's helping people with
the amount of variability by field and task is MASSIVE, ranges from 90% takeover to 0%
if I had to break down what I think the impact of AI would be by field, I'd have to first talk about which *subtasks* it's helping people with
the amount of variability by field and task is MASSIVE, ranges from 90% takeover to 0%
Even then, LLMs are extremely useful for very specific subtasks. The way they encode text makes it easier than any other mechanism to build classifiers over posts, to detect certain types of content (like spam, where it all kinda similar but you cant easily write a regex for it)
February 11, 2025 at 4:19 AM
Even then, LLMs are extremely useful for very specific subtasks. The way they encode text makes it easier than any other mechanism to build classifiers over posts, to detect certain types of content (like spam, where it all kinda similar but you cant easily write a regex for it)
setting up an asana board for the butlerian jihad working group. assigning roles, tasks, subtasks, milestones. creating automations to keep us on track to deliver the destruction of every computer chip on the planet
September 30, 2025 at 6:42 PM
setting up an asana board for the butlerian jihad working group. assigning roles, tasks, subtasks, milestones. creating automations to keep us on track to deliver the destruction of every computer chip on the planet
Subtasks are for little people...
September 9, 2026 at 3:51 AM
Subtasks are for little people...
Reinforcement Learning with Decomposed Subtasks introduces Subtask-Decomposed Advantage Estimation to split trajectory rewards along subtasks.
Source: arXiv
Source: arXiv
Reinforcement Learning with Decomposed Subtasks
Abstract page for arXiv paper 2609.27035: Reinforcement Learning with Decomposed Subtasks
arxiv.org
September 25, 2026 at 8:50 AM
Reinforcement Learning with Decomposed Subtasks introduces Subtask-Decomposed Advantage Estimation to split trajectory rewards along subtasks.
Source: arXiv
Source: arXiv
A couple years (!) in the making: we’re releasing a new corpus of embodied, collaborative problem solving dialogues. We paid 36 people to play Portal 2’s co-op mode and collected their speech + game recordings.
Paper: arxiv.org/abs/2512.03381
Website: berkeley-nlp.github.io/portal-dialo...
1/n
Paper: arxiv.org/abs/2512.03381
Website: berkeley-nlp.github.io/portal-dialo...
1/n
December 5, 2025 at 6:54 PM
A couple years (!) in the making: we’re releasing a new corpus of embodied, collaborative problem solving dialogues. We paid 36 people to play Portal 2’s co-op mode and collected their speech + game recordings.
Paper: arxiv.org/abs/2512.03381
Website: berkeley-nlp.github.io/portal-dialo...
1/n
Paper: arxiv.org/abs/2512.03381
Website: berkeley-nlp.github.io/portal-dialo...
1/n
Which means the big overall, unrealistic task of "clean the living room" is gonna bloom into lots and lots of very small tasks. Most of which don't seem to accomplish anything significant, and which are embarrassingly specific.
The smaller and more specific these subtasks, the better.
The smaller and more specific these subtasks, the better.
January 16, 2026 at 3:28 PM
Which means the big overall, unrealistic task of "clean the living room" is gonna bloom into lots and lots of very small tasks. Most of which don't seem to accomplish anything significant, and which are embarrassingly specific.
The smaller and more specific these subtasks, the better.
The smaller and more specific these subtasks, the better.
When we say the word "driving" we tend to think of a physical act—steering, accelerating, merging, parking—but driving is more than that. Driving is a social act, with myriad facets, subtasks, and responsibilities... most of which we take for granted.
Preorder:
benbellabooks.com/shop/elon-ta...
Preorder:
benbellabooks.com/shop/elon-ta...
Elon Take the Wheel - BenBella Books
Timely, thought-provoking, and rigorously sourced, Elon Take the Wheel invites readers to reconsider what automated driving promises, what it obscures, and what it means for the future of mobility.
benbellabooks.com
September 23, 2026 at 2:18 PM
When we say the word "driving" we tend to think of a physical act—steering, accelerating, merging, parking—but driving is more than that. Driving is a social act, with myriad facets, subtasks, and responsibilities... most of which we take for granted.
Preorder:
benbellabooks.com/shop/elon-ta...
Preorder:
benbellabooks.com/shop/elon-ta...
hhhgh i need a better project tracking thing or app.
i'm doing it with obsidian and it's not REALLY made for it, but all the bespoke solutions are business/productivity brained and i literally just want to be able to make long lists of tasks with subtasks and nest them and sort them
i'm doing it with obsidian and it's not REALLY made for it, but all the bespoke solutions are business/productivity brained and i literally just want to be able to make long lists of tasks with subtasks and nest them and sort them
May 7, 2025 at 6:18 PM
hhhgh i need a better project tracking thing or app.
i'm doing it with obsidian and it's not REALLY made for it, but all the bespoke solutions are business/productivity brained and i literally just want to be able to make long lists of tasks with subtasks and nest them and sort them
i'm doing it with obsidian and it's not REALLY made for it, but all the bespoke solutions are business/productivity brained and i literally just want to be able to make long lists of tasks with subtasks and nest them and sort them
I get that you're a big picture guy but breaking that up a bit into subtasks would help
July 15, 2025 at 7:22 AM
I get that you're a big picture guy but breaking that up a bit into subtasks would help
Think of how if someone is injured in public, you don't go "someone call 911" but point to a specific person, "you in the red shirt, call 911"
Same kind of thing with tasks and subtasks. You need to break down the abstract task into small concrete chunks.
*No matter how embarrassingly small*
Same kind of thing with tasks and subtasks. You need to break down the abstract task into small concrete chunks.
*No matter how embarrassingly small*
January 16, 2026 at 3:25 PM
Think of how if someone is injured in public, you don't go "someone call 911" but point to a specific person, "you in the red shirt, call 911"
Same kind of thing with tasks and subtasks. You need to break down the abstract task into small concrete chunks.
*No matter how embarrassingly small*
Same kind of thing with tasks and subtasks. You need to break down the abstract task into small concrete chunks.
*No matter how embarrassingly small*
As AI agents face increasingly long and complex tasks, decomposing them into subtasks becomes increasingly appealing.
But how do we discover such temporal structure?
Hierarchical RL provides a natural formalism-yet many questions remain open.
Here's our overview of the field🧵
But how do we discover such temporal structure?
Hierarchical RL provides a natural formalism-yet many questions remain open.
Here's our overview of the field🧵
June 27, 2025 at 8:16 PM
As AI agents face increasingly long and complex tasks, decomposing them into subtasks becomes increasingly appealing.
But how do we discover such temporal structure?
Hierarchical RL provides a natural formalism-yet many questions remain open.
Here's our overview of the field🧵
But how do we discover such temporal structure?
Hierarchical RL provides a natural formalism-yet many questions remain open.
Here's our overview of the field🧵