Hacker News Reader

I think you have two common misconceptions here.

1) Wages have very little to do with value/productivity in a free market. They are almost entirely determined by supply and demand. Value simply places a ceiling on compensation. Thus, if far more people can perform a "programming" job because of GPT-X, unless the demand for those jobs rises significantly the net result will be wage reduction.

2) There's this weird thinking on HN that since a developer's job involves [insert your % estimate of time spent actually coding/bug fixing] and the rest is figuring out requirements, dealing with weird requests, planning, etc. that means developers can't be replaced. However, I don't see a whole lot of discussion around what the difference is between a developer and a competent business analyst in a GPT-X world. The latter can be had for significantly less money, requires less training, and if the actual programming part is largely automated away or concentrated in the hands of fewer "real" developers, those roles start to look awfully similar.

It's not that more people can do programming with GPT-X around, because the AI will only solve the problems that have already been solved thousands of time in the past in slightly different ways. What GPT-X cannot do is left to real CS people. So instead of coding CRUD systems, we can do real algorithms research again, except fewer people are capable of it.

> competent business analyst

with the danger of invoking the "No True Scotsman" fallacy, I'd say that competent business analyst are even more difficult to get hold of than a competent programmer.

I've had so few managers that were competent at managing people and projects.

Business Analysts aren't usually (project) managers

The developer can prompt for a solution with specific storage/performance requirements by specifying an algorithm, and specific scalability requirements using by specifying an architecture. Imagine a business analyst prompting for an app, and getting a ruby on rails monolith with O(N^2) performance for the core behavior for a service that is expected to have millions of requests daily.

> Imagine a business analyst prompting for an app, and getting a ruby on rails monolith with O(N^2) performance for the core behavior for a service that is expected to have millions of requests daily.

I see this as the main argument against "we will just have tools that allow managers and ba's to do what devs do now". I think folks often forget that there are two sets of requirements for every app: business requirements and technical requirements. Non technical folks might understand the business requirements very well and may even be able to write code that satisfies those requirements, but the real value in a dev is squaring those business requirements with technical ones. A BA might look at a DynamoDB table and say "yeah lets just run a scan for the items we need", whereas a dev will look at the same problem and say "yeah we can do that but it will cause issue A, issue B and sometimes issue C". And the dev knowing those gotchas is why you have them there in the first place, a dedicated person that knows all these gotchas and makes sure you organization avoids a footgun in prod.

The follow-on prompts would be to refactor the existing system to solve the scalability issues. You'd need to be able to feed in the existing codebase for that, though.

1) Economic theory says that marginal product of labor (value) is equal to wages, at least in simple models.

2) With GPT-4 you still need to know how to program. A product manager can't replace you.

1) The real world is not a simple economic model. The wage rate is roughly equivalent to the rate it costs to replace an employee, not their marginal value. If your argument was true, company profits would tend toward zero as wages rise.

2) I specifically did not say GPT-4. If you think v4 is the peak of what will be possible when looking at how far we have come in just 2 years then I don't know what to tell you. Also, a product manager is not a business analyst.

1) I think it's approximately true in the real world. If you can hire an employee who adds substantially more revenue than what their wage is, you keep hiring employees until value created by an employee is close to their wage.

When discussing economy, it's good to start with understanding the situation through the lens of economic models and than look to what extent are the conclusions applicable to the real world.

Using your argumentation - above you said that xyz is true in a free market, but real-world markets are not free...

2) And I specifically said that I'm assuming no dramatic improvement beyond GPT-4. The 2 misconception I supposedly have... I didn't even make that claim.

I completely disagree because of how this batch of AI learns. It learns by studying what is currently available as opposed to supplying something new.

Bad developers on the way of becoming obsolete by AI will see drastic improvements from the thing that’s about to replace them. Otherwise it’s slightly better intellisence integrated in your editor.

The difference is that bad developers are primarily concerned with literacy and their primary intent is chasing easy. Better developers are comfortable reading code no differently than an attorney reading a large contact and their primary intent is chasing better automation. It’s the difference between designing a car engine versus hammering pieces together.

I suspect this will open wage distinctions. Those who can write software will be identified from those who can’t. Those who can will be fewer in number and their wages will increase. Those who can’t will be a lower skilled commodity with depreciated wages, like the way data entry was once a skilled profession many decades ago but no longer.

I think a lot of "good" developers making this very argument will be in for a nasty surprise when they find out which bucket they're actually in

I don’t doubt that at all. “Good” is relative to something but what that thing is often makes for a poor quality of measure with few samples compared.

The are objective ways to measure competence in software but nobody wants to use them. Most people don’t want to discover just how good they aren’t.

> - I expect developer productivity to go up 1.5x - 15x over the next several years assuming GTP-4 based tooling is integrated into IDEs.

Hardly. 10% of the time spent is in writing code and the other 90% goes behind debugging and fixing things.

Even if AI shrinks the time spent in that 10% in writing by another 90%, you still have have to take full time for maintaining that code. Changing, debugging, testing, deploying, profiling, log analysis.

Worse, its now someone else's code you need to understand and debug, and not the one written by yourself.

It's not just writing code, GPT and GPT-based tools can help with fixing bugs, solving configuration problems, understanding existing code, suggesting variable names, etc.

ChatGPT is better at that 90% than the 10%.

> On the other hand, people working in agriculture have become much more effective, which lead to much fewer people working in agriculture, maybe because people need to eat as much calories per day as they did 100 years ago.

Except population exploded also, and the only reason we could support such a population has been thanks to advances in agriculture (it could be argued that's not a good thing, as the larger population is helping fuck our planet up in other ways, but that's for a different thread). So there has been an increased demand for food.

However, there's only so much arable land on this planet. The planet is finite in size. So when basically the entire planet is already being farmed, you can't really add more farmers, so further efficiency is just leading to less farmers.

Software is not so finite. It technically is, as we only have so many servers or whatever, but we're nowhere near the upper limits of what all we can handle or the appetite for software that companies have (which seems about infinite). Additionally, we have a habit of rebuilding and replacing software every few years, so even if we reach capacity limits, there's still demand for new software to replace the old, or working on new features for existing software.

So it's a bit different of a situation and not really comparable.

Also, much like that law of bureaucracy stating how work expands to fill the time available, every employer I’ve worked for has been able to envision and request vastly more software functionality than any dev team has been able to produce, which might not extrapolate forever, but has sustained the field so far.

On the other hand, people are much more finite in the amount of food that they require

> assuming GTP-4 based tooling is integrated into IDEs.

And how, exactly, is this going to work?

Is every IDE going to pay some license fee to OpenAI? Will that be up to the companies or—even worse—individuals who use the IDEs?

What happens when OpenAI decides that GPT-4 is passé and it's time to move on to something else? Will the existing IDEs that are designed to interact with GPT-4 stop working?

Will GPT-5 or whatever provide the same kinds of assistance, or will there be regressions? Will they require a completely different license for it?

See, these are some of the important questions that come with assuming that a fairly new for-profit company will act as the backbone for your marvelous programming renaissance—particularly when the service in question is one that takes quite a lot of resources to continue to operate.

Personally, I would be interested in some kind of LLM-assisted IDE, but no way in hell am I going to make any significant portion of my job dependent on an organization that decided, as soon as it was clear there was real interest in what they were doing, that they'd just toss the "Open" part of their name and principles out the window and just go for closed-source proprietary product and maximum hype for more profit.

> up 1.5x - 15x over

I wish! I still often find out my team members don’t know basic text/Unix tools. Heck one guy took our list of enums in a JSON file and wrote on his keyboard by hand each of them into a class. It took me maybe a minute in Sublime Text when he wasted at least a day

Supply generally doesn't decrease cost fyi