The endangerment of lazy work
The consequence of AI doing a decent job
If you look around here you’d probably form the view I’m your typical bullshit AI bro. I guess by some measure I am but in many ways I find the most interesting thing about AI is what it means for us.
Whether you think someone is rocking up soon with ‘AGI’ as Altman and the chip-shovel LLM goldrush store Nvidia will tell you, or whether it shows up in 100 years. It was going to show up, right? In whatever form you care to define.
I wonder about what human work looks like then. This makes today an interesting moment in time that scholars will look back on, I think. On one hand, who isn’t tired of AI slop right now? Janky Claude-ese gets up my nose, if I’m honest. Even this is quite interesting.
Tangent; in Drew Breunig’s recent unpacking of system prompt changes in the Fable series, Drew suggests that the models are trying to write in bullet points, even when told not to and I think he’s onto something there.
Anyway, I was talking about what it means for people, so here’s a little thing I’ve been thinking about. We’re tired of slop but I’ve been tired of human slop for years now. I used to be a tech journalist, which I guess makes me the ex-smoker of writing. Or just old and grumpy. Either way, little irks me more than reading low-effort internet shit-takes. People being paid to do a job and they’re just phoning in some shallow analysis when an hour of research might have them write something actually worth reading.
Which brings me to Orac Deep Dive experiment. The initial goal was to surface some jank writing on the topic of the day, something to react to, like the other experiments. But then it ended up being kind of interesting to read. Absolutely not always right and certainly missing anything like the insight of a person.
I got to thinking though. Why not just decompose the construction and authorship of an article so that you can be sure that some of the craft and graft was done. In a nutshell; get inspired by something (the news briefing and a Guardian API slurp), read it properly, research it with some searches, pick a framing (an opinion if you like), steelman alternative views, write it up as a narrative of your experience and edit it for unjustified reaching and other editorial mistakes. Diagram on the About page.
I’m getting to the point, I swear. So let’s make the bold claim. Orac’s piece, on average, is a better piece of analytical narrative than most of the writing on the web. Why? Because it did all those things. It didn’t just read an article and paraphrase it which is what online writing has devolved to in many areas.
I used fancy models initially, but today I wondered. What if I just use these great bang-for-buck super cheap models instead, and I switched some models and dropped the average price of it from about a buck US to six cents, and four of those cents was Sonnet 5 writing the prompt for the image generator. Here’s the result.
Does a better job than (some) humans. For six cents. Not because humans can’t do a better job than this, of course they can. They just don’t want to. They didn’t need to, because human laziness set a kind of low bar for your run-of-the-mill intellectualism. All that is changing.
How this all fits with the rise of the meat proxy I don’t know, but what I will say is that when people create systems that produce useful outputs, which are notably better than lazy human effort (AI-accelerated or not), then the value of lazy work will surely plummet.
I’m kind of down for that to be honest.