Jonathan Blow – If AI solved coding, why is my software still the same? [video]
Claims that “AI has solved coding” collide here with the reality that most software people use still feels sluggish, buggy, or unchanged. Commenters argue that coding was rarely the true bottleneck: organizational incentives, product vision, architecture, testing, and maintenance remain hard, and current LLMs mainly accelerate boilerplate and personal tools while often adding technical debt and low-quality “slop.” Alongside that, participants debate whether prolific indie output and custom one‑off apps count as real progress, and how much weight to give prominent game developers who are skeptical of AI’s impact on software craftsmanship.
Perceived Software Quality and Change
- Many argue everyday software is worse: more bugs, slower (e.g., modern OS releases, Outlook reminders), and more “AI slop” bolted on.
- Others say core quality hasn’t changed much; only hardware and surface tooling improved while UX and reliability stagnated or regressed.
- Some note that dominant commercial apps change slowly, whereas AI-driven improvements are mostly invisible internal tools and personal utilities.
Coding vs. Organizational Bottlenecks
- Strong theme: coding was rarely the true bottleneck; communication, planning, legal, politics, and coordination dominate timelines.
- LLMs compress the “coding + component test” phase, exposing that requirements, architecture, deployment, and feedback loops now dominate.
- Large orgs with meetings, process, and legal review cannot easily convert raw coding speed into faster shipping.
AI Coding Productivity: Hype vs. Reality
- Enthusiasts claim 10–100× output for certain tasks, rapid creation of whole apps, engines, and personal clones of commercial tools.
- Skeptics report negative or modest net gains: AI produces over-complex, brittle code, huge PRs, and more maintenance/verification work.
- Multiple commenters stress selection bias: side projects and greenfield demos look spectacular; business context, edge cases, and long-term maintenance reveal limits.
Architecture, Requirements, and “What to Build”
- Broad agreement that AI doesn’t “solve” product vision, prioritization, or high-level system design.
- Good requirements, domain modeling, and abstraction boundaries still require human judgment; LLMs often violate encapsulation and create tangled architectures unless heavily guided.
- Some foresee future workflows where humans specify goals and constraints while AI dictates implementation details, but this remains aspirational.
Personal and Indie Use Cases
- Many report dramatic success building personal software: custom media tools, productivity apps, indie games, local-first replacements for SaaS.
- These tools often suit a single user well, but would be fragile or unscalable as general products; QA, UX polish, and edge cases remain hard.
Employment, Power Dynamics, and Culture
- Several note rising job anxiety among developers; some managers are openly pleased that AI has “cut devs’ tails” and made them more compliant.
- Concern that AI amplifies managerial focus on artifact count (tickets, docs, lines of code) rather than outcomes, and accelerates enshitification rather than quality.
Overall Sentiment
- Consensus: AI has significantly changed how code is written but not yet what mainstream software is or how organizations operate.
- Enthusiasm for AI as a powerful tool coexists with deep skepticism about quality, maintainability, incentives, and long-term impact.