What I don’t understand is that how is this automatically good for startups?

Say I have a startup that vibe codes “AI for real estate”. What about customer acquisition?

On the other hand, if I’m Zillow, why can’t I just throw a developer on the same feature and automatically have a customer base for it?

If you look at most of the YC funded startups these days, they are just prompt engineers with no go to market strategy and some don’t even have any technical people and are looking for “technical cofounders” that they can underpay with a promise of equity that will statistically be meaningless.

Infact, the opposite is true. There is tech inflation. Coding will get easier, but that doesn't mean tech is getting easier. It will only get more complex and will drive more segregation.

Tech is dividing the society and driving a wedge deeper. There is a huge population that are being thrown wayside by the high-speed tech highway. Which means the tech is getting more and more unreachable.

AI assistants are only going to make this worse, by removing the direct touch between users and the tech. Tech becomes just unmanageable for average person.

Just like how you were able to do all repairs for your bike, as a kid. But you can't do the same for your car now. Tech never gets easier or reachable.

Technology has always been deflationary. But you don't put off buying a computer because it will be cheaper next year. Nobody seems to be putting off buying GPUs despite scary depreciation and a blistering pace of new product introductions that are ever cheaper faster and better.

"But building the same functionality has undoubtedly become simpler."

I disagree with this statement. It has become simpler, provided you don't care about it actually being correct, and you don't care about whether you really have tests that test what you think you asked for, you don't care about security, and other things.

Building the same thing involves doing the things that LLMs have proved time and again that they cannot do. But instead of writing it properly in the first place, you now need to look for the needle in the haystack that is the subtle bug that invariable get inserted by llms every single time I tried to use them. Which requires you to deeply understand the code anyway. Which you would have gotten automatically (and easier) if you were the one writing the code in the first place. developing the same thing at the same level of quality is harder with an LLM.

And the "table stakes" stuff is exactly the thing I would not trust an LLM with for sure, because the risk of getting it wrong could potentially be fatal (to the company, not the dev. Depends on his boss' temperament) with those.

In the end, the article says:

> writing functioning application code has grown easier thanks to AI.

> It's getting easier and easier for startups to do stuff.

> Another answer might be to use the fact that software is becoming free and disposable to your advantage.

For me, the logical conclusion here is: don't build a software startup!

The impulse to make the comparison makes sense. The reality is a bit different, probably leans in the right direction but is buffered by learning. I’ll explain. There is no upside to delaying purchasing if you think things will get less expensive. There is however upside in building today even if you have to rebuild tomorrow, and that upside is in learning the problem space. Specifically, what is likely to be trivialized and what truly requires domain knowledge. Horizontal apps? Little domain knowledge to encode. Vertical app? More domain knowledge to encode. Separately, there are more ways to differentiate than distribution alone, see verifier’s law. Problems that are challenging to verify are challenging for AI to trivialize.

Good explanation for why hiring stopped. If AI is improving rapidly why hire engineers now that you might not need in 6-12 months?

The models themselves represent the biggest deflation cases I've ever seem.

The charged cost of a frontier model is ~200x lower than 2 years ago, and the ones we are using now are much better - although measuring that and how much is challenging. Building a "better than GPT-4" model is also vastly cheaper than building GPT-4 was... perhaps 1/100th?

I'm still ahead of a company starting later:

I have the legal structure, i know my collegues, i have potentially employees and more capacity.

The problem is not that a startup is starting after you but you do not give yourself time to keep an eye on AI and not leveraging it when its helpful.

We leverage AI and ML progress constantly and keep an eye on advances. Segment Anything? Yepp we use it. Claude? Yes Sir!

>> First, models getting better makes AI-based applications easier to build, because they can be simpler.

Don't conflate easy with simple. I'd argue they are actually easier and far more complex.

With re. to the first two paragraphs, it's crazy how someone can be so massively brainwashed.

> when prices go down instead of up. It is generally considered harmful: both because it is usually brought on by something really bad (like a severe economic contraction)

Or, you know, technological improvements that increase efficiency of production, or bountiful harvests, or generally anything else that suddenly expands the supply at the current price level across the economy. Thankfully, we have mechanisms in place that keep the prices inflating even when those unlikely events happen.

"One of the main problems is that if people expect prices to keep going down, they'll delay purchases and save more, because they expect that they'll be able to get the stuff for less later."

That is why we are all waiting to buy our first personal computers and our first cell phones.

Economists have managed to be ludicrous for a very long time and yet we still trust them.

> Desktop app.... though Electron and Tauri have made it easier

Ugh. I don't like that kind of 'desktop' apps. Huge bloat with a blip of actual app.

There’s a fun version of this in futurist space travel speculation.

Let’s say you have the a fusion rocket and can hit 5% the speed of light. You want to migrate to the stars for some reason.

So do you build a generational ship now, which is possible, or… do you wait?

Because if you build it now someone with a much better drive may just fly right past you at 20% the speed of light.

In this one the answer is to plot it out under the assumption there is no totally undiscovered major physics that would allow, say, FTL, and plot the curves for advancement against that.

So can we do this with software? We have the progress of hardware, which is somewhat deterministic, and we know something about the progress of software from stats we can make via GitHub.

The software equivalent of someone discovering some “fantasy” physics and building a warp drive would be runaway self-improving AGI/ASI. I’d argue this is impossible for information theoretical reasons, but what if I’m wrong?

Does anyone else agree with this the premise of this article? Is it sensible to put off building things now because it will get even cheaper and faster later?

Maybe the time value of time is only increasing as we go.