montyanderson.net

New Bottlenecks

June 2025

A paradigm shift is occurring in the production of software. A fake "end of history" consensus is about to be upended as old bottlenecks subside and new ones emerge.

Language lies on the borderline between oneself and the other
— Mikhail Bakhtin, 1953

Intelligence too cheap to meter is well within grasp
— Sam Altman, 2025

Last week my cofounder Mikhail, who does not have a background in programming, was able to build something in one day that easily surpasses the output of the best front-end engineer I know. This caught me off guard.

Until last week, my perspective had been programming with AI functions as an accelerant and lowers the barrier to entry to building software. This is unbelievably impressive on its own but insufficiently optimistic – underpriced by at least a factor of ten. The difference is that both developers and non developers can make digital machines faster than anyone could previously.

This will be a shock because, over the last twenty plus years, a deep consensus has been created about how the production of software should be done. These organising principles have been realised through experience and, while they are not random or arbitrary, they are dependent on practical constraints which have until now been constant: namely that developers are expensive, error-prone, and limited by the speed of thought.

Here's three ideas taken from this perspective that we need to reunderwrite:

  1. Code should be simple rather than maximally rigorous
    • Developers are expensive and great developers are even more expensive
  2. Changes should be thoroughly reviewed and tracked in version control
    • Mistakes happen
  3. Maintenance burden should be minimised
    • It's really hard to rebuild a codebase
  4. It's cheaper to run software than it is to build it

I suspect that each of these notions will be considered legacy within the next few years. Tools we hold dear will become our biggest bottlenecks; as the ground shifts we need to interrogate our assumptions and imagine new ways of working.

Existing dichotomies like source versus compiled/generated code will start to fall apart. We'll need to create systems that cater to fuzzy and not just boolean logic. May we scale smoothly, exponentially and uneventfully through superintelligence.