Degrees

My continued experimentation with spec-driven development has made me examine this AI revolution/circus with fresh eyes.

If we equate the broad commercialization of Gen AI and LLMs to other seminal events - like the broad commercialization of the internet, the invention of the A-bomb, the implementation of agriculture, or any other wildly hyperbolic comparison - then the tooling creation we're seeing now is the greater refinement of the engineering practice around the base concepts.

It's this mode of thinking that makes me a little more sympathetic to the possibilities of AI than the hyper-capitalist push towards a Utopian Dystopia with 100x increases in productivity every year, where everyone is happily unemployed, waving to their billionaire patrons (while averting their gaze) who sail by on yachts of ever-increasing length.

I'm skipping over a few innovations (or baubles, depending on your perspective), but the wide adoption of coding (or other) agents is significant, not because it further approximates or effects a thinking machine, but because it uses deterministic tools (now commonly known as a harness) to manage the bullshit generated by LLMs, and thus make their net output more productive.

However, even with the agents in place, they're still apt to write nonsensical or redundant code and execute instructions against the explicit permission of the user, thus reducing their capacity to improve productivity.

SDD is another step in the evolution of improving the utility of coding agents.  It's an ostensibly boring layer gating the agent even further, effectively quelling its fevered, digital pseudo-mind.  It forces the agent to explain its reasoning more directly and give the user digestible steps for project execution prior to implementation.  The user has more control over what can be generated at any given step.  Any one of those given steps can still generate unadulterated bullshit (and, in my experience, has).  But the gating mechanism effectively allows you to fact-check a lie at a time rather than turn a pathological liar loose for 10 years and determine which of the whoppers most needs to be refuted.

This incremental gain seems to show up more in what's actually reported by corporate earnings than the exponential panacea we were promised.  I recently stumbled across a report where a company was reducing the number of teams focusing on one of its core functionalities from 4 to 3 and using AI to make up the difference.  At any other point in history, this would be huge (a 25% gain in productivity!), but when compared to the promises of today, it seems downright picayune.  

When the dust settles on the current AI hype, hopefully people will recognize that the tools have utility and promise via incremental improvements without needing to bet all of civilization in order to cover their debts.

Until next time, my human and robot friends.