How I use AI (Oct 2026)
Claims that “coding is solved” by modern AI agents spark sharp disagreement, with some engineers reporting huge productivity gains from tools like Claude and Copilot, and others pointing to fragile architectures, subtle bugs, and a general sense that end‑user software quality is worsening. Many see LLMs as excellent tutors and accelerators for routine coding and system design, but insist that human judgment, architecture, and careful review remain essential — especially as token costs mount and multi‑agent workflows prove mentally exhausting. The thread also delves into AI-generated music and video, where opinions split between viewing them as legitimate new creative tools and condemning them as hollow slop that erodes human expression and training data quality.
State of AI Coding (“coding is solved”)
- Some argue that for many typical software tasks, LLMs can now take a Jira ticket and produce a mergeable PR, making manual coding largely unnecessary.
- Others strongly dispute “solved”: domain modeling, architecture, performance, and long‑term maintainability are still hard; LLMs help with syntax and boilerplate, not deep design.
- Several people note models excel when a project is well‑structured with clear patterns; they behave worse in messy, free‑form or complex codebases.
Agents, Workflows, and Fatigue
- Multi‑agent “swarms” are seen by some as magical productivity boosters; by others as expensive, unstable, and mentally exhausting.
- Juggling many concurrent agent threads is described as overstimulating compared to traditional coding.
- A counter‑view: frontier‑level practitioners increasingly rely on fewer, well‑aligned agents plus strong processes, not swarms.
Code Quality, Review, and Reliability
- Reports of “insidious” errors and bizarre implementations (e.g., copying whole DBs to update one row) underscore the need for human review, strong tests, and clear specs.
- Some claim they rarely read code anymore, relying on agents + automated reviewers + QA and rollbacks; others find this reckless and insist reviewing generated code is still essential.
- Disagreement over whether AI‑written code is already better than most human developers or merely faster at producing low‑quality “slop.”
Economics: Tokens, Hardware, and Jobs
- High token spend (thousands per month) is seen by some as justified for research or rapid product work, and by others as wasteful compared to buying local GPU rigs.
- Many see AI as a strong multiplier rather than a job killer: you still need engineers to specify, validate, and iterate.
- Concerns about “deskilling” of programmers vs optimism that weaker devs will be pulled up to a higher baseline.
Training Data and Long‑Term Quality
- One side worries model training will degrade as more AI‑generated content floods the web.
- Others counter that reinforcement learning, synthetic data, and better processes can keep performance improving.
AI for Art, Music, and Video
- Heated debate: some find AI music and imagery soulless and inherently non‑art; others say art is defined by human intention and response, with AI just another tool.
- Many draw a sharp contrast between enthusiasm for AI‑generated code and discomfort or hostility toward AI art and music.
- Examples of AI‑assisted short films and music are cited as “real” creative works, though some viewers still sense an uncanny, not‑quite‑human quality.
AI as Teacher and Research Aid
- Strong agreement that LLMs are excellent for learning and deep dives into complex topics (e.g., energy policy), especially when they surface good primary sources.
- Still, trust is limited: people stress verifying claims and treating AI explanations as starting points, not authorities.
Software Quality and the “Bubble”
- Several commenters note modern software feels more brittle and buggy despite claims that AI has “solved” coding.
- There’s debate over whether AI is a real productivity revolution or another tech bubble; some see steady, undeniable gains, others see mostly hype and soft product degradation.