Artifice
Naval Ravikant tweeted today:
“We’re now in the era of slop code.”
Nader Dabit replied:
“Unpopular opinion - AI-generated code will be of higher quality than 99.9% of human-created code in a very short time frame (< 6 months).”
“The average programmer vastly overestimates their ability to avoid stupid mistakes or waste time on things that don’t matter. Arguably the bulk of work in the field has been entirely about preventing human programmers from doing things AI programmers won’t.”
That last point lands differently when you think about it. Type systems, linters, code review, test suites, CI pipelines - guardrails built over decades to constrain the chaos humans introduce. What if the chaos was the human part all along?
Horses and petroleum
Horses evolved to run across open spaces to escape wolves. Selection pressure made them a certain size, which gave them a certain strength. Eventually we realised we could use them for riding, for pulling carts, for agriculture.
But this was adaptation, not design. Horses didn’t come with a label saying “use me as a transport system.”
Likewise, sometime in the past, dead animals and plant matter got compressed over millions of years into petroleum. It didn’t come with instructions saying “burn me just so, and I’ll push vehicles along.”
For centuries, there were tasks you simply couldn’t do without livestock. Now we’ve covered the world in roads. There are 1.6 billion vehicles on Earth - not quite one per person globally, but in places like the US it’s close. NVIDIA’s market cap is larger than the UK’s GDP. We’re doing it again.
Things got out of hand. What started as jimmied-together artifice became a paradigm that stuck, scaling to a degree that would have seemed mad at the moment of discovery.
Next-token prediction
It’s similarly non-obvious that next-token prediction, given enough compute and training data, would lead to models prompting humans to add end-to-end tests - then adding them themselves, then spotting category errors in the user’s original assumptions.
The first computers were built to calculate artillery trajectories. Now we’re running neural nets on them. Von Neumann himself drew inspiration from McCulloch and Pitts’ model of neurons when designing the architecture we still use. Full circle.
You could argue we weren’t even meant to be here. The connectionist approach - modelling the brain’s diffuse, distributed logic - was proposed early, then buried by symbolic AI for decades. Minsky and Papert’s critique of perceptrons in 1969 sent neural networks into winter. Only now has the original intuition scaled to dominance.
And here’s the strange part: these models might be converging on something real. The Platonic Representation Hypothesis suggests that as AI systems get larger and more capable, they represent data in increasingly similar ways - regardless of whether they started with images or text. Different architectures, different training data, same underlying structure emerging. As if they’re all approximating the same statistical model of reality. We jimmied together next-token prediction and it might be revealing something about the territory, not just the map.
The liminal period
Naval could be on the wrong side of history here. The “slop code era” might be a liminal period - the models already pretty good, scaling laws still holding. The METR results on task length keep climbing. There doesn’t seem to be an end in sight.
We’re not entering a dark age where code regresses because it’s easier. We’re in a brief window where human limitations are still visible before they’re papered over entirely.
That’s the thing about artifice. It doesn’t announce itself. It just scales until the original seems quaint.