Andrej Karpathy · Independent · 2017-11-11
Eight years on, this is still the cleanest statement of the central bet behind the last decade of AI: that large parts of code will be replaced by weights, and that this is a shift in kind, not in degree. Almost every later argument about AI engineering, including "vibe coding," is downstream of this short post.
The value for a learner is that Karpathy lays out the trade-offs of the new stack — homogeneity, agility, but also brittleness and uninterpretability — before the field had vocabulary for any of it. It reads now as a checklist of problems we are still working through, which is the highest compliment an essay can earn.
> Neural networks represent the beginning of a fundamental shift in how we develop software.