Miguel Conner’s post is not a rejection of AI coding. It is a report from someone who has built AI agents and then chose to slow down for a three-month coding retreat in Brooklyn. The detail that made Hacker News lean in was the framing: “the old way” now means writing and debugging mostly by hand, at the same time many programmers are being told that programming itself is close to solved.

The work described in the post is specific rather than romantic. Conner says he is working through CS336 assignments, training a 17M parameter model on Tiny Stories, moving toward FlashAttention2 in Triton, writing small agents and neural networks in Python, using BASIC on an Apple IIe, practicing terminal security challenges through Bandit, and spending time in Vim. A footnote says he still asked Claude for advice on a few bugs, which makes the experiment more useful: this is calibration, not purity.

The HN discussion quickly became a thread about what developers lose when the feedback loop collapses. One commenter described teaching 18-year-old students 6502 assembly on an emulated Apple II Plus. Another argued that agentic workflows can look competitive in output while weakening codebase memory, debugging pattern recognition, and active recall. Several people treated AI autocomplete as a middle ground that keeps the hands on the code.

That is why the post traveled. The community was not debating whether tools should exist; it was asking what kind of developer is left when every obstacle is routed through a model. Manual work forces contact with syntax, runtime errors, slow searches through docs, and the small experiments that build judgment.

Read alongside the thread, the post lands as a practical warning. AI coding can raise throughput, but the ability to review, steer, and repair AI-written code still depends on the slower skills that hand coding trains.

#craft

HN turned “coding the old way” into a debate about what AI skips