The news, 365 days behind — on purpose Delayed live · replaying 2025

One Year Ago.AI

Remember how fast this is.

19JUN2025replayed
one year on
communityAndrej Karpathy

Karpathy proposes new versioning for software in the age of AI

Karpathy argues that neural network weights and prompts represent distinct categories that deserve their own version numbers alongside traditional code.

Andrej Karpathy on Thursday released a new talk titled ‘Software Is Changing (Again),’ published to YouTube and quickly rising to the top of Hacker News with over 1,400 points. The HN discussion sorts software into traditional coding, neural network weights and prompts, while Karpathy says the versioning analogy is ‘a bit confusing’ because it ‘usually additionally implies some kind of improvement.’ Commenters wrestled with the taxonomy, with some arguing that versioning implies replacement when in reality neural networks and prompts will complement code. Others seized on the practical implications, with several calling structured outputs a bridge between paradigms that remains ‘criminally underused.’ The thread swelled to 783 comments as developers debated where to draw the boundaries between tools — and whether a Software 4.0, perhaps involving machines generating their own protocols, could follow.

G
gchamonlive

Argues versioning implies replacement, but these paradigms are extra tools that will coexist with traditional code, not panaceas.

M
miki123211

Calls structured outputs / JSON mode a criminally underused tool that bridges neural networks and traditional code without needing labeled data or GPU training.

A
abdullin

Highlights that structured outputs allow forcing LLMs through a predefined task-specific checklist before answering, boosting accuracy and auditability.

B
BobbyJo

Finds the versioning intuitive: each new tool unlocks a new problem space, and no previous tool has disappeared.

G
gyomu

Warns that neural network approaches face training data scarcity and legibility constraints that limit their applicability compared to traditional code.

One year later — open only if you can handle spoilers

The talk's categorization proved influential, with 'Software 3.0' entering common parlance among AI engineers. However, by mid-2026, many practitioners adopted a more pragmatic view that the paradigms are complementary rather than sequential, echoing early HN skeptics.

Replay thisPost on XRedditHNLinkedIn

The Weekly Replay · free by email

This week, one year ago — every Sunday.

One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.

Free · double opt-in · unsubscribe anytime · privacy