Intro
This is my first post in a while and surprise surprise, it’s about AI and LLMs.
This is the first part in a multi part post to make it easier to read. It was originally a single post and has since been broken down into multiple parts.
To start with let me agree to disagree with everyone who doesn’t think the LLM powered automated world is coming.
In my view, it is coming.
Whether full AGI (whichever definition is trending now) will be achieved in 5/10/15/20/… years, I don’t know, this post is not about that.
The following however rings true to me in terms of possibilities (to be clear, I’m not trying to write an AI prediction blog post, I’ll stop with this section of premises. I promise)
- LLMs will continue to get better and have larger context windows
- Nvidia will continue to improve the quality of their chips
- Chip Competitors will catch up
- LLMs on smartphone chips become as powerful as GPT-4 within 5 years.
- Jobs continue to become redundant over time as capabilities, cost, and energy efficiency get better.
- AGI definitions and timelines will keep changing, the world along with it.
Using my anecdotal experience with conversational products like chatgpt, gemini, grok, claude, perplexity, and the underlying apis, I’m quite convinced of the 6 points above.
Are we living in the AGI world where everything is fully automateable? No.
Are we going in that direction? I think so.
Before we get into the framework, here’s where this is headed: the interesting question is not which jobs survive automation. It’s that we’re entering a glut of execution where the marginal cost of “intelligent” output races toward zero.
What becomes scarce is judgment and the uncomfortable problem is that judgment has always been a byproduct of doing the labor that’s now being automated away.
That’s the trap. And later in this post, I’ll argue that we need to build systems that let humans and machines compete on the same playing field, with skin in the game on both sides.
Now, a framework for thinking about what “purpose” actually means when the machine can do the work.
POSIWID + +
POSIWID is a well known term in the tech/systems thinking word, it’s a shorthand for ‘a purpose of a system is what it does’
More generally, if I replace system with ‘thing’ it has broader applicability while still having the same spirit.
It is a useful way of evaluating what a ‘thing’ is for.
For example, lets take a Librarian. Their role has evolved, expanded and contracted over time.
Their initial role was to keep a record of what books were present.
It evolved to organizing books in specific ways to make it easier to access them.
They were also tasked with ensuring that books were stored safely and accessed securely, in the process, preventing theft.
Over time as the number of books being published went up, they became book curators who managed the book collection in the library to have the appeal that it needed to have.
In effect, the role evolved to also become a recommender of good books for enthusiastic readers.
Curation in some ways is one step before recommendation.
As libraries started stocking more than books (magazines, newspapers, cds, movies) their role expanded to include the management of these assets as well.
Doing all of these roles requires some level of expertise. Over time this has led to specialization at various levels.
If I apply POSIWID to this, what is the purpose of a librarian?
The interesting thing about POSIWID is that the best measure of a system is the outcomes it has.
So what are the outcomes/outputs generated reliably by a librarian:
- Reduction of cognitive friction - making things easier to find, organizing things correctly
- Creating order and preventing chaos - keeping information upto date, latest versions of books, conservation, new media formats
- Creating trust and credibility - a book available in a library curated by a librarian you like is one you’re more likely to read than otherwise
This breaks the administrative capacity down to its fundamentals.
With this in mind, if I think about how these outcomes have been impacted in the recent past, I’d say the following ring true:
Google/Goodreads/Amazon/Archive.org/The general internet have built solutions
- to make things easier to find, this optimization continues till date
- to keep things upto date, allow for preservation
Some of these are paid, some are free, but they exist.
But, in the process of democratizing access to information, what to trust/what not to trust has remained a bastion of humans in general.
Viewed through this lens, some social media influencers/large businesses are in effect librarians, signalling that something is worthy of a readers time by it being on a list that they put together.
The LLM era will continue to
- struggle against this last bastion of what to trust for a while
- eat up the rest of it.
LLMs do not suffer for being wrong and have no personal stake in a recommendation, making them a system that can only partake in your upside but with none of the downside of being wrong.
To be clear, I have nothing against librarians, in fact, the above thought experiment is an argument that posits that in the future the librarians we trust will be human, more so than ever, but, their role will be stripped to the fundamental aspect of a signal of trust and credibility. The rest of it might not survive.
In effect, The purpose of a librarian will be to decide what deserves attention and make that decision trustworthy.
POSIWID - -
Switching gears, let’s look at a non-knowledge-worker role.
In the city I grew up and still live in, there’s an area called Washermanpet, in Tamil, its called Vannarapettai (means the same thing)
It used to be one of the primary areas where many of the city’s Launderers used to live and operate.
These areas were called Dhobikhanas, an entire industry of semi-skilled labourers operating at scale to wash clothes.
I read a book called ‘A Geography of Time’ by Robert Levine about 10/11 years ago (an interesting book that explores how different cultures and peoples evaluate time and how technology and systems impact how we utilise time)
An interesting passage that has stuck with me is from the chapter regarding industrialization,
Until the Industrial Revolution, in fact, most evidence suggests that people showed little inclination to work. In Europe through the Middle Ages, the average number of holidays per year was around 115 days. It is interesting to note that still today, poorer countries take more holidays, on the average, than richer ones
It has often been the very creations intended to save time that have been most responsible for increasing the workload. Recent research indicates that farm wives in the 1920’s, who were without electricity, spent significantly less time at housework than did suburban women, with all their modern machinery, in the latter half of the century. One reason for this is that almost every technical advance seems to be accompanied by a rise in expectations
In effect, a Dhobikhana provided us a maintenance service that over time came into the household.
We would call this democratization and decentralization today if it happened to an internet service.
Today there still exist Laundry and Dry Cleaning services, they do the following:
- Doorstep pickup and dropoff (or a common dropoff and pickup point)
- Operate a set of laundry machines, keep them well maintained
- Provide a predictable timeline for pickup
- Have specialized skills when it comes to certain types of fabric or certain types of dyes or stains
- Maintain accounts and relationships with customers
If we apply POSIWID to this role, then, the purpose of a laundry system is to externalize a recurring maintenance cost from households and convert it into a predictable service.
The dhobikhana didn’t vanish because it was inefficient.
It vanished because washing machines became cheap enough to purchase and easy enough to operate to push maintenance inward.
In this case, unlike the librarian, the role didn’t transform, it almost disappeared, it hangs on because people hate doing their laundry.
The productivity gain from moving from a centralized system to a self-owned decentralized system was negative, we end up leasing up our time to do a task we paid someone to do, but, the cost arbitrage made it a simple decision to make.
So what happens when robotic automation removes the need for a washing machine loading and unloading process in our homes?
In this scenario, there’s no need for a specialized service or machine that does just one thing, there’s a need for a machine that removes the mental bandwidth making low-stakes decisions and physical time spent executing it.
The difference is that the ‘cost’ in this case is one that is defined by time spent doing a task, not the cost arbitrage of us doing it at home versus someone else doing it for us elsewhere.
In my view this presents an interesting scenario where we were willing to ‘lose’ personal productivity which led to a loss of jobs and what society gained from it is a little unclear.
POSIWID + + - -
What about newer types of jobs like a Software Engg/Product Manager?
The role of an SWE has been:
- Write correct, testable, as bug-free-as-possible code in the least amount of time possible
If they were more senior, it also included:
- review other members code
- define the frameworks to be used
- define the coding style to be followed
- (sometimes) figure out the best way to host it
- improve latencies and optimize whatever they’ve built to be as fast as possible
This is an oversimplification, there are several other things that these roles do.
I’m taking a birds eye view to avoid bias due to the fact that I’ve been in and around these roles for the past 10 years unlike the other 2 roles I’ve written about.
A common opinion today is that LLMs have made writing code cheap and fast.
I’ve been writing code/managing teams that write code/reading code for more than 12 years now.
I would say that even prior to ChatGPTs launch, it was cheap to write code, starter kits existed for everything and any well maintained library had a storybook with implementations and stackoverflow was always available to troubleshoot most edgecases, all of which kept making it faster.
What remained true was that you needed to know how software worked and how to write code in order to utilize those resources.
Until the past year or so this has been true, and remains true to a large extent even now.
The extent is just getting smaller and smaller and my belief (as mentioned at the top) is that this will continue.
What will remain true is that while the agent can make a strong recommendation of what to use, bearing the cost of those decisions is still on the person making that decision.
Something very similar is happening in the world of product management.
It is the job of product teams across the world to collect information, derive insights from it and make product decisions from it.
I’d say the first and last steps are still largely untouched, the rest of it, unstructured and multi-modal - is already something that can be performed by LLMs.
But, let’s apply POSIWID to these roles,
SWE
It is to decide what code should exist and be accountable for its behavior in the real world.
The obvious way to achieve that is to write code.
A Senior SWE on the other hand is making these decisions without writing code at times since a Junior SWE is writing the code.
i.e making the decision of what stays and doesn’t without writing the code itself.
PM
Absorb uncertainty and commit an organization to a direction while considering the limited information and options available to them, while taking on the cost of being wrong.
I’m confident in saying that similar reductions can be made for any other type of role that is categorized as a skilled or a knowledge worker : architect, author, designer, accountant, financial planner etc.
Careful with that axe, Eugene
I’m sure this is not a perfect parallel but, this is my attempt at adding my perspective to it.
| Role | Old Scarcity | New Scarcity |
|---|---|---|
| Librarian | Access | Trust |
| Launderer | Labour | Time |
| Software Engineer | Implementation | Judgment |
| Product Manager | Information | Commitment |
If we proceed to analyze other roles, I’m sure similar parallels will exist there too.
The interesting thing about all this is, excepting the Launderer, the others all become functions of ‘do you trust this person’s judgement?‘
An extension of this is, ‘would you trust an LLMs judgement in making decisions for you?’
While I’m sure they’ll continue to get more and more advanced, decision making and the ownership of the costs of those decisions are unlikely to be outsourced to an LLM.
There have been multiple published studies about giving LLMs a few thousand dollars to invest or letting an LLM manage a retail outlet
But, before we continue down this path, I feel it’s worth looking through the past with a couple of examples.
First, lets take the role of a programmer/SWE (its the role I have the most experience with)
My understanding is that the earliest programmers programmed on punch cards
The ability to punch cards with precision and avoid errors before the card went into the verifier was a crucial part of the role.
The role evolved, as the nature of programming changed, and with it the above skill set became obsolete.
Next, If we take the role of an accountant, to the layman it’s a role that involves working on spreadsheets in Excel.
Spreadsheets were originally Ledgers or Worksheets that were spread on a table in order to make tallying numbers and cross-checking easier. No formulas, no macros.
The best accountants back then were not just people who knew the rules but were also the ones who were able to notice an error in a number.
Great vision was probably of a lot of value back then, also maybe good hand-writing. The point is, the role has evolved and with it the requirement has changed.
What’s interesting about both of these examples to me is that when the tools evolved and old skill sets became redundant, the amount and complexity of the work done just exploded. The tools kept getting better and the skill sets kept evolving.
Something similar happened in the world of librarians and launderers as well. The skill sets evolved, the number of libraries expanded, in the launderers case, their core semi-skilled work which was manual labour got outsourced to a machine, but, knowing how to handle specific fabrics, remove specific stains remained and then each one of those lost ground to the machine and better detergents etc as well.
The role’s POSIWID evolved over time for these roles and others as well.
On applying the same logic as in the table above, it seems to me that what’s left is human judgement and the willingness to make a bet.
The execution cost is going down, it’ll tend towards zero, maybe never get to zero.
So when do we switch over to an LLM based system? In my view, the switch over point is when these 3 conditions hold true:
- the cost of executing with the system is cheaper than executing with a human
- the LLM system generates the required output faster than a human
- the output from such a system beats what a human produces
All 3 are both true and false right now.
It’s not yet true for all 3 at the scale a lot of people predicted it would.
A librarian became trustworthy by doing it for years An engineer’s judgement was valuable because they engineered things A launderer gains credibility by laundering clothes A product manager gains a reputation for making the right bets
The argument could be that an LLM could gain a similar reputation by having such a track record over a period of time.
I don’t think it’s that simple.
I’m not an AI doomer by any means, I embrace it and have been telling people I know that I have been becoming redundant one piece at a time these past few years.
However, one aspect that an LLM based system will never have to deal with is downside-risk.
This will in my opinion remain in the human domain.
The value of being right is sometimes outweighed by the cost of being wrong If that sometimes becomes often enough, you go bankrupt.
My hunch is that we’re moving into the decision making and verification economy.
Humans are here for quality control.
Which is weird but, I sense that this might be true.
Quality control might not be the right term for it, but the POSIWID application of all of them should look about the same.
So What’s Left?
The pattern across all four roles is the same. Strip away what the machine can do, and what remains is a human standing behind a decision, bearing the cost of being wrong.
The librarian’s purpose isn’t to organize books, it’s to stake their credibility on what deserves your attention.
The engineer’s purpose isn’t to write code, it’s to decide what code should exist and own what happens when it runs in the real world.
The product manager’s purpose isn’t to gather information, it’s to absorb uncertainty and commit an organization to a direction.
Execution is becoming cheap. It will get cheaper.
The scarcity is shifting from “can you do the work?” to “will you own the outcome?”
This sounds clean. It isn’t.
Because here’s the thing about judgment, it was never taught in a classroom or transferred through a document.
It was a byproduct of labor. You learned to spot bad code by writing bad code.
You learned which books matter by reading ones that didn’t.
You learned which product bets to make by making wrong ones and sitting with the consequences.
If the machine does the labor, how does the human acquire the judgment necessary to supervise it? I don’t have a full answer yet.
But I think this is the question that matters more than “which jobs will AI replace?” because it determines whether the humans left in the loop are actually qualified to be there.
In the next post, I dig into what I think is coming: a glut of execution, a shortage of taste, and the uncomfortable future of high-stress, low-effort work. You can read Part 2 here
What do you think?
Thank you for reading
Sainath