Btw. about 6 months ago the Anthropic CEO said that by now 50% of all code would be written by LLMs.

How does that prediction relate to the reality we all live in and what does that say about his ability to make predictions about the future?

(https://www.businessinsider.com/anthropic-ceo-ai-90-percent-code-3-to-6-months-2025-3)

Anthropic's CEO says that in 3 to 6 months, AI will be writing 90% of the code software developers were in charge of

"And then in 12 months, we may be in a world where AI is writing essentially all of the code," Anthropic CEO Dario Amodei said.

Business Insider
@tante I guess that makes him more accurate than most politician's promises ...? 😕
@tante maybe internally they are at 90%, and now nothing works any more 🤣
@claudius @tante Ok 90% of code by AI ... And also 90% Of _really working_ non junk code by humans :)

@tante to quote the aws ceo: “It’s a silly metric,” he said, because while organizations can use AI to write “infinitely more lines of code” it could be bad code.

“Often times fewer lines of code is way better than more lines of code,” he observed. “So I'm never really sure why that's the exciting metric that people like to brag about.”

https://www.theregister.com/2025/08/21/aws_ceo_entry_level_jobs_opinion/

AWS CEO says using AI to replace junior staff is 'Dumbest thing I've ever heard'

: They're cheap and grew up with AI … so you're firing them why?

The Register
@hagen @tante Musk fell in to that amateur trap when he bought Twitter - used number of lines of code to evaluate employees' performance. A bit like how he now measures his success by number of starships and USAs he's broken.
@tante
One could get to this result if LLMs were now creating as much code as was written by humans before.
@tante I mean, not technically wrong, just... How often are those generated lines of code actually compiled and used for something?
@tante Why do we keep seeing headlines about people who use % like an almost-accurate way of measuring their -self claimed- success with zero objectivity?
It feels like betting on the weather with your windows closed, yet trying to convince both the public and the investors you are looking at the sky.
@tante We still have time until sept 14th ?
@tante That may as well be true for all we know. 99% of that 50% are probably just automatically discarded and re-done until the test passes, though. Brute-force vibe coding.
@tante since when anyone would listen to startup CEO whose sole goal is to scam and sell off snake oil product?
There is no gotcha, that is their game.

@tante Another non-dev commenting on what he doesn’t know.

There is a well-known saying: ”The last 10% of a product’s development takes 90% of the time.”

Even if that was true by line-count (and it won’t be, just another Elonism), it still wouldn’t make much difference.

Further, a lot of lines in industry codebases arent handwritten even now! ”Code Generators” provide devs with great options to avoid writing boilerplate code themselves. But they don’t make random mistakes.

@gimulnautti @tante One of the first things I learned as a software dev

https://mstdn.social/@staringatclouds/114967186727805872

staringatclouds (@[email protected])

@[email protected] I learned the "80/20 rule" The first 80% of a project takes 80% of the resources, time & money The next 20% takes another 80% of the resources, time & money Then I learned that Sometimes it's a 90/10 rule or even a 99/1 rule

Mastodon 🐘
@tante well, it might be writing 50% of the *code*. It's not creating 50% of the *features* though - more like 90% of the code debt.
@tante this is an Elon FSD situation

@tante

He doesn't care about reality. He got his gigantic investments and he got his big pay bonus. Results do not matter, especially when he can say "another 6 months" and the cycle repeats.

@tante Yes, I want to see critical infrastructure code written and designed by a machine that can neither count B's in a blueberry or compute simple arithmetic.
@tante CEOs love to underestimate the work real human workers do
@tante
Not like anyone is fact-checking marketing copy. It was 2+ years ago they said we were "95% of the way" to AGI. This is and always will be about conning people into buying products that have no legitimate use case so a handful of companies can eventually become profitable. Weird corporate ouroboros chasing nonexistent profits and efficiencies.
Saturday Morning Breakfast Cereal - Nigh

Saturday Morning Breakfast Cereal - Nigh

@tante
I am a software engineer. 80% of the time I do not write code, but talk to people about what the code should do.

If "AI" would write 50% of the code, that would "save me" 10% of my time, or most likely less, all under the assumption that the code is perfect, which of course is just a wet dream of some managers that don't know shit about technology.

@musicmatze @tante

Real devs wait for the AI shirtless, idly, in the cold!
Come at me!!

@tante Lies to boost his stock. He will make new lies at every opportunity so that we forget the old ones. It's the American Way.
@tante They measure code by LoC not utility so they prediction just means they're planning to spew orders of magnitude more boilerplate and outright garbage.
@dalias @tante
50% of email is now spam
— a spam sender
@lanodan @tante Yeah this is exactly the same thing.

@dalias @tante @lanodan

Is this a subtoot about car traffic?

😉

@DavidM_yeg @dalias @tante Heh wasn't intended as one but yeah it also works :D

@tante In a similar timeframe my employer's CEO told us he's planning for 100% AI written code. Meanwhile the organization can't reliably document an internal API or library for human consumption.

There's zero possibility of it happening without major investment in the infrastructure they already refuse to do for humans.

@tante a former coworker was continuing her CS degree. She related a story of chatting with a prof about students using #AI to do their homework. Prof related giving an intro assignment: create a function that will return true or false if a submitted text string contains a vowel.

Instead of 1 function, the students who used AI to do their work returned a 7 function program: contains_vowel(txt) which in turn called contains_a(txt), contains_e(txt) contains_i(txt)....

@tante obvious BS aside, the thing that gets me is this notion of people not taking note of the AI replacement on their livelihoods. Assuming he was right, what are people going to do in this AI dystopia? They're not talking about new jobs appearing, is he suggesting we have the revolution early, while we're still well fed and healthy?
@tante Maybe he is right, the amount of AI slop will be so absurdly large that it can literally be 90% of all code.
@tante My bet's that the LLM that he had generate his talking points is the source of that number.
@tante I wouldn't be surprised if it is accurate, it's just the 50% AI code gets immediately discarded.
@tante like the number can be arbitrarily close to 100% using a pipe to /dev/null
@shironeko @tante What if other people/companies will do the same? /dev/null throughput competition.
@hellman @tante yeah I think he was hallucinating about all code?
@tante His prediction might be correct, if you include all the code generated where the user said "this is the 20th iteration and this code is still shit. Let's code it by hand and tell management that we used the AI."
@tante it's likely true by the sheer volume of bullshit AI generated code that's being produced. real human written code is still the same, and AI is slop-vomiting 9 times as much code now that is utter shit and doesn't actually work, or borderline does but is absolutely inefficient and wasteful...

@tante All you have to do is to generate 10 times more slop code than are currently written by humans, and these systems are certainly capable of that, no?

No one said anything about the quality or even usefulness of said code?

@loke @tante
Or even if the code is used. As a spewer, AI is unmatched

@tante
I actually now imagine they have an LLM sitting in a box just outputting like, 1 billion lines of code each day.

'Our metric shall rule them all!'

It doesnt' do anything, of course, but still.

@tante And in the same 6 months close to 100%, if not exactly 100%, of that LLM generated code will need to be debugged, at great expense, by someone who learned how to do this by writing code that LLM's are now producing

When these people, who learned their craft by doing it, retire or give up because idiot managers replace them with LLM's & they can't get jobs

Who will fix the bugs then ?

There won't be any new experienced people who know how to debug & all the rest will be gone

@tante this will depend a lot on the definition of "writing" and "lines of code". If you count individual characters, autocomplete probably has been "writing" the majority of code for quite some time. Lots of autocomplete functionality gets meshed in with LLMs these days, and so, without representing any major shift in the way we write code, you can massage the statistics to make it look like AI is "writing" a lot more code than we would otherwise say.

In the end, a lot of writing code is about the code that isn't written, and hand edited automatically generated code has been with us pretty much since the time text editors were powerful enough to support it.

@tante AI will be producing 90% of the puke. Not that drunkards are going to produce less puke, but AI will start producing 10 times more puke than we currently have

The world will be awash in fake stomach acid, they will paint the walls in technicolor puike, bile will run in the streets

Mr Creosote has met his match.

https://www.youtube.com/watch?v=GxRnenQYG7I

#MrCreosote #AIeeeee

Monty pythons, Mr creosote, Full version,

YouTube
@tante I’d say he’s about as good as any other human. Which is to say: terrible.
@tante you don’t need an LLM to write bad code, but it is a great force multiplier.