Devs are famously bad at estimating how long a software project will take.

No, highly complex creative work is inherently extremely difficult to estimate.

Anyway, not shocked at all by the results. This is a great start that begs for larger and more rigorous studies.

As someone that has had to double check peoples code before, especially those that don't comment appropriately, I'd rather just write it all again myself than try and decipher what the fuck they were even doing.

Megacorp LLM death spiral:

1. Megacorp managers at all levels introduce new LLM usage policies.
2. Productivity goes down (see study linked in post)
3. Managers make the excuse that this is due to a transitional period in LLM policies.
4. Policies become mandates. Beatings begin and/or intensify.
5. Repeat from 1.

ahahaha holy shit. I knew METR smelled a bit like AI doomsday cultists and took money from OpenPhil, but those "open source" projects and engineers? One of them was LessWrong.

Here's a LW site dev whining about the study, he was in it and i think he thinks it was unfair to AI

I think if people are citing in another 3 months time, they'll be making a mistake

dude $NEXT_VERSION will be so cool

so anyway, this study has gone mainstream! It was on CNBC! I urge you not to watch that unless you have a yearning need to know what the normies are hearing about this shit. In summary, they are hearing that AI coding isn't all that actually and may not do what the captains of industry want.

around 2:30 the two talking heads ran out of information and just started incorrecting each other on the fabulous AI future, like the worst work lunchroom debate ever but it's about AI becoming superhuman

the key takeaway for the non techie businessmen and investors who take CNBC seriously ever: the bubble starts not going so great

I have an LLM usage mandate in my performance review now. I can’t trust it to do anything important, so I’ll get it to do incredibly noddy things like deleting a clause (that I literally always have highlighted) or generate documentation that’s more long-winded than just reading the code and then go to the bathroom while it happens.

5% "coding"

95% cleanup

I just want to point out that every single heavily downvoted, idiotic pro-AI reply on this post is from a .ml user (with one programming.dev thrown in).

I wonder which way the causation flows.

Anyone who has had to unfuck someone else’s work knows it would have been faster to do the work correctly from scratch the first time.

Software and computers are a joke at this point.

Computers no longer solve real problems and are now just used to solve the problems that overly complex software running on monstrous cheap hardware create.

"Hey I'd like to run a simple electronics schematic program like we had in the DOS days, it ran in 640K and responded instantly!"

"OK sure first you'll need the latest Windows 11 with 64G of RAM and 2TB of storage, running on at least 24 cores, then you need to install a container for the Docker for the VM for the flatpak for the library for the framework because the programmer liked the blue icon, then make sure you are always connected to the internet for updates or it won't run, and somehow the program will still just look like a 16 bit VB app from 1995."

"Well that sounds complicated, where's the support webpage for installing the program in Windows 7?"

"Do you have the latest AI agents installed in your web browser?"

"It's asking me to click OK but I didn't install the 1GB mouse driver that sends my porn browsing habits to Amazon..."

"Just click OK on all the EULAs so you lose the right to the work you'll create with this software, then install a few more dependencies, languages, entire VMs written in byte code compiled to HTML to run on JAVA, then make sure you have a PON from your ISP otherwise how can you expect to have a few kilobytes of data be processed on your computer? This is all in the cloud, baby!"

@dgerard What fascinates me is *why* coders who use LLMs think they're more productive. Is the complexity of their prompt interaction misleading them as to how effective the outputs it results in are? Or something else?

From the blog post referenced:

We do not provide evidence that:

AI systems do not currently speed up many or most software developers

Seems the article should be titled “16 AI coders think they’re 20% faster — but they’re actually 19% slower” - though I guess making us think it was intended to be a statistically relevant finding was the point.

That all said, this was genuinely interesting and is in-line with my understanding of the human psychology that’s at play. It would be nice to see this at a wider scale, broken down across different methodologies / toolsets and models.