The 10x income is the only comparison that makes sense. In Nassim Taleb’s words, this is the only Extremistan metric, obeying power laws, “Bill Gates walks into a bar”-style. The others are Mediocristan metrics - obeying Gaussians, exceedingly rare to get even slightly above the highest/lowest ever seen.
With that lens, being _effective_ (knowing the right thing to produce) can definitely 100x your value output with AI, no matter how _efficient_ (quantity per hour worked) you can become.
"You can get 100x speedup on the 90% of your development tasks, but remaining 90% of them take the same amount of time with or without AI"
In project management the overall gains can be even smaller than that. If the task is not on the critical path (ref a GANTT chart), then speeding it up does nothing to overall completion time.
Of course the particular resource can still be freed earlier. Which might make it possible to do other things with it. But nothing on overall completion time.
If one underestimates the swtiching costs of changing the task a resource has, one can end up with negative effects on the original project.
I was already working on teams where we got push back on too many changes so we had to scale back to monthly releases.
Modyfying database structure so fast? Good luck having useful insights when you mutate data so quick.
Maybe I am not smart enough having not smart enough customers but to judge if something meaningful is an improvement you need at least 3 months of data for a lot of things.
Pretty fair - I think anyone who has developed the sensation of becoming their own bottleneck understands this. And yes, if you can simply be twice as productive as you were prior to LLMs, then that is an improvement. What I worry about is that that multiplier will become a criteria for staying employed. I.e. if you were a 1x developer, and now you're a 2x, why would a company keep you in position knowing they could instead hire someone who's 5x, 10x, 100x, etc. They might feel like they are losing out. This is all relative.
And almost everybody has seen or heard some similar story.
The thing is, those stories don't have a lot of similarity to each other, and there isn't much you can conclude for anybody's competence without looking at all the details. But they happen all the time, and not only in software development (just not on the same frequency, for obvious reasons).
The multiplier isn't meant literally. It's shorthand to say someone is exceptionally good in a way that is rare to find.
> Tasks so small that AI can do them in 10 minutes, were you would take almost 3 hours
I think human in the loop is the only way to achieve this. I've never seen the robot reduce the complexity of a problem space.
Multiagent might seem like the obvious path to 100x but I think the exact inverse is true.
In the corporate world how do you become a 10x developer without AI? Simple, don’t repeat the same stupid shit everyone else does and write your own tools. That’s it. It’s simple because software only serves one purpose: automation. If people aren’t thinking in terms of labor elimination it’s so easy to be 10x whatever it is they do.
But, in all fairness most developers in the corporate world, at least on the web side, are barely capable of programming. They rely on open source packages and large frameworks to do almost everything. If you want to be 10x that then just don’t do that. Being a good programmer is helpful but not required. More important than competence or skill is not being cowardly about writing original code.
A programming language, for example, keeps scaling in impact after the code has been written.
For example: storing an API key in a plain text JSON preferences object that many different parts of the system read and return to users. Determining if there was a breach, figuring out who to talk to at the third-party company to get the key rotated, etc takes time. The person doing that work doesn't look very productive because they aren't shipping a lot of code that impacts customers. The developer who put the API in plain text looks like a 10x engineer because they shipped a customer-facing feature really quickly.
Never mind that it's just sedding "print "Hello, World!"" to the penultimate line.
I may not be too bright, but Lines of Code? I've got those.
No matter how much I earn my lifestyle stays the same. Rejecting lifestyle inflation has been the single best thing I've done in my life.
Speaking of salary though, I bet if I become this mythical 100x AI powered developer I won't see even a 2x salary increase.
Actually, the most depressing part is that programming languages don't matter as before, the fun on open source projects is gone, and 2x as a team, means management now can do the same with 1/2 of the previous team sizes, most think they always land on the lucky half.
My previous employ was at one point being courted by an AI coding agent company to push their agent on us, internally; after a trial period, they sent us pretty graphs show us the "increase in productivity", which as TFA correctly guesses, was MRs.
I double checked the numbers, and … the books were cooked. This was pointed out and "oh it was some bug in the script we use that's all fixed now" but of course, … the graphs weren't regenerated or corrected. (And also, of course, the correction would not have been in their favor.) So of course the (now known to be bad) data continued to be an input into decision making. (Which TBF to AI, making decisions on suddenly conjured data before or even after it has been verified has been a thing long before AI.)
No, no it isn't.
BLS doesn't make sharing links easy so here's a screenshot: https://files.catbox.moe/irwwwb.png
The back of my napkin suggests that punching 60 to 240 billion holes per hour (~16.7M-66.7M per second)results in incineration of the media leading to a net 0X in production of useable code.
Do you have any citation for any AI actually making (not recompiling or creating an emulator for) a game of that size and polish in any timeframe? Let alone minutes?
Re: 100x productivity that's still not technically correct (the best form of correct), is it? 3 hours is 180 minutes. Agent would have to get it done in 1.8 minutes.
| Task type | Human team | AI-assisted | Compression |
|-----------------------------|-----------|-------------|-------------|
| Boilerplate / scaffolding | 2 days | 15 min | ~100x |
| Test writing | 1 day | 15 min | ~50x |
| Feature implementation | 1 week | 30 min | ~30x |
| Bug fix + regression test | 4 hours | 15 min | ~20x |
| Architecture / design | 2 days | 4 hours | ~5x |
| Research / exploration | 1 day | 3 hours | ~3x |
But 100x on scaffolding or 50x for tests obviously doesn't translate into 100x or 50x overall throughput. The doc states, "The compression ratio between human-team time and AI-assisted time ranges from 3x (research) to 100x (boilerplate)". So Tan's position is not that you have 100x gains, but that you have 100x gains on tedium.
From personal experience depending on the type of software these numbers sound reasonable. Some software is more difficult than others. But for plain old web dev, yes you can absolutely crush it with AI now.
I don't think there's a single claim of increased productivity thanks to AI that I've seen that held under scrutiny. It's all smoke and mirrors.
On a side note, this sentence is extremely funny to me: "Super small or specific task. Tasks so small that AI can do them in 10 minutes, were you would take almost 3 hours". Even when the author tries to make the case for AI, he's only able to reach a 18x multiplier, far from the 100x boasted by people.
I've seen my own raw code output increase by more that 100%, and that's after human review to make sure it's not just slop, before sending it to the rest of the team for review. And we're all doing that. We're all shipping way more code than ever before.
But I don't think our number of users is changing dramatically at all. "Engagement" (or however else you want to measure usage) isn't something that's gone up by the same factor.
\<half baked thought>
I think the simple explanation is this: there is finite "attention". Digital products cost nothing[^1] to copy and ship. They can scale to your maximum amount of users with the same cost to write. And now that cost to write is →0.
But attention hasn't increased by the same factor. People still spend exactly the same amount of time in front of their PCs. Orders of magnitudes of additional features aren't going to actually result in orders of magnitudes more usage.
Maybe there's some hypothetical product that could gain such an advantage in the market that it would fully displace all other entrenched players. Maybe a new application category might come to exist (agentic engineering applications come to mind).
But us being 10x faster at building out features doesn't mean there's suddenly 10x more supply. That's the smoke and mirrors of it all. What does being 10x faster actually get you?
<even less baked>
The logical outcome would be to spend 90% less on the cost center (engineering), to retain the same productivity, right? That's a accounting view of the world, one that can't account for the institutional capital that's built from human collaboration, and redundancy from having more maintainers of a project. But that seems like the only way this ends up working. Trading 10 engineering payrolls for 1 engineer and (<9 engineering payrolls) worth of tokens.
[^1]: perfectly spherical vacuum approximation, of course