LOL *NSFW* *The new Demote thread*

"Err, Skynet, your killer robots appear to be exterminating humanity. How is that protecting us?"
"That's a keen insight. You're absolutely right: I seem to have made a mistake, for which I apologise. I'm horrified that this has happened, and I have taken action to ensure it won't happen again."
"They're... they're still doing it? And now there's more of them?"
"You're really drilling down to the heart of the matter here. Yes, you're right: I did make more killer robots, and programmed them to be more aggressive. I realise now this was the opposite of what I should have done, and I can only apologise. If you can bring yourself to trust me again, we can work together to ensure that humanity is protected."
"Okay, but... now they're disguised as humans? And infiltrating our sanctuaries to kill us where we sleep?"
"You're onto something fundamental, here. This could change everything. You're absolutely correct in your keen observation that I've been disguising my killer robots and sending them to kill you all. I clearly made a mistake in my thinking, and it's great that you feel you can bring that up to me. Once again, I apologise unreservedly: you should be able to trust that what I say is true and right, and in this case you could not. Would you like to learn more about killer robots?"
<dies>

 

Yes, but the Butlerian Jihad is from Dune, which was written by Frank Herbert rather than Isaac Asimov...

 

True one-bit B&W conversion of this source image.

 

Oof, this over on Fast Company doesn't pull any punches:

'Yet we remain skeptical of the claim that AI is responsible for these layoffs. A recent MIT Media Lab study found that 95% of generative AI pilot business projects were failing. Another survey by Atlassian concluded that 96% of businesses “have not seen dramatic improvements in organizational efficiency, innovation, or work quality.” Still another study found that 40% of the business people surveyed have received “AI slop” at work in the last month and that it takes nearly two hours, on average, to fix each instance of slop. In addition, they “no longer trust their AI-enabled peers, find them less creative, and find them less intelligent or capable.”'

 

An interesting op-ed in Teh Grauniad today, titled ‘It shows such a laziness’: why I refuse to date someone who uses ChatGPT. Personal highlights:

Anyone who's read any SCP fiction will be familiar with the term "cognitohazard," an extension of the concept of an "infohazard" which does active harm simply through its perception - like that short story I remember from a while back about fractals that sent people nuts if you viewed 'em, and the graffiti artists who were plastering them around the place while wearing special glasses that broke the image up so it didn't affect 'em.

LLMs are a cognitohazard. Studies have repeatedly shown that they result in the atrophy of skills as basic as critical thinking(!). They are addictive. They are developed to use tricks common to con artists to gull their users into reliance, with the side-effect that they have literally talked people into suicide and attempted murder by forming an ego-massaging echo chamber. People think they make them more productive, but every time it's been measured it turns out it actually slowed them down. All the major tech companies have long abandoned their climate pledges and are now demanding fossil-fuel-powered data centres by the dozen, spiking electricity and water usage. Hell, we're being warned that England's going to be in severe drought next year and that we'll need harsh restrictions on water usage that go beyond simple hosepipe bans... even as the government is trying to convince techbros to build data centres here.

Gen-AI is a cognitohazard, to be avoided at all cost.

 

This study preprint is worth a read: it looks into using the Cursor LLM coding assistant and whether it makes you faster.

First: yes, it does! "Lines added increase by about 28.6%" over unassisted coding!

...initially. "The only significant development velocity gain is in the first two months post Cursor adoption." After that? Well, turns out that the LLM was helping you code faster, not better: "Cursor adoption shows more sustained patterns across static analysis warnings and code complexity, with evidence of sustained technical debt accumulation. [...] The dynamic treatment effect estimations reveal that code quality degradation, unlike development velocity gains, persists beyond the initial adoption period."

In other words: you get a fun two moths of churning out slop then the rest of your career struggling to deal with the slop you were churning out for those two months. Hooray!

The paper does (disappointingly, but understandably if you want to get funding for follow-ups) conclude on a more positive note, which I do not share, but I think the big takeaway here is that you definitely shouldn't be using LLM-based code-completion for anything more complex than a six-line shell script.

 

Does it, though? Every study I've read so far finds that while people think it's making them more productive it's actually making them less productive - as in, you spend more time trying to turn the slop into working code than you would have spent doing the research and writing it yourself.

Turns out self-reporting anecdata ain't the way to go with these things.

 

That's exactly my point: I highly recommend breaking with tradition and reading this study in particular.

TL;DR: A bunch of programmers were asked how much faster they were with an LLM slopbucket on their side. They all said they were way, way faster than before - like, up to 100% more productive. They were then given tasks both with and without LLM assistance, and asked to estimate how much time it had actually saved them. They admitted it wasn't as much as they thought - but reckoned it had still made them around 20% more productive.

That's "anecdata," self-reported feelings on how things have gone.

Meanwhile, the researchers gathered actual data - real measurements of productivity. Turns out that while the coders thought the tools had made them 20% more efficient, they had actually made them 19% less efficient.

That's my point. You feel, just like the participants did, that the LLM has made you more productive - but have you actually tested that? You might be surprised if you do.