Old and new apps, via modern coding agents
Modern coding agents are enabling domain experts, including top mathematicians, to quickly build educational visualizations, revive old Java applets, and prototype custom tools that would previously have taken days of manual work. Commenters see this as evidence of a huge latent demand for “personal software,” while debating how far LLM-generated code can be trusted beyond low-stakes dashboards and teaching aids. The thread also touches on broader worries about job displacement, the quality and maintainability of AI-written software, and the emerging role of AI in serious mathematical research as well as hobby projects.
Perception of LLM-Coded Software & “Vibecoding”
- Many describe “vibe-coded” apps as quick, fun prototypes: small utilities, games, dashboards, personal trackers.
- Several note these projects would previously have taken days, now done in hours with agents.
- Others argue most such apps are trivial, buggy, and non-representative of serious software, likening them to amateur phone photography vs professional wedding photographers.
Impact on Software Labor, Economics, and SaaS
- Some predict a huge increase in software supply, potentially depressing wages unless demand explodes; Jevons paradox is cited.
- Others think only low-end or consulting-style work is at risk because production systems still need robust engineering and maintenance.
- A few foresee traditional coding as a shrinking career path and stress “retooling” to agentic workflows, especially for younger developers.
- There is a strong thread about personal, in-house tools displacing SaaS for many workflows, with an eventual capitalist “outsourcing → insourcing → outsourcing” cycle.
Use in Education, Visualization, and Personal Tools
- Multiple educators describe using LLMs to rapidly build simulations, visualizations, and teaching computers that they’d wanted for years but never had time to implement.
- Visual aids are seen as “nice to have,” not mission-critical, so AI-generated bugs are more acceptable.
- Individuals report highly customized tools for tiny businesses or hobbies (e.g., niche trackers, gear planners, Java applet ports), often much more useful to them than off-the-shelf software.
Trust, Reliability, and Appropriate Use Cases
- Strong consensus that AI outputs must be reviewed; “generally not to be trusted” is a recurring phrase, especially for production or safety-critical code.
- Some argue trust should be framed as choosing the right tool for the right job, not full autonomy.
- Others say LLM code is typically “atrocious” and unmaintainable, acceptable only for toys and low-stakes experiments.
AI in Mathematics and Research
- The blog author’s use of coding agents for interactive math supplements is seen as a pragmatic, time-saving choice, especially for teaching materials.
- Commenters note broader use of AI in formal verification and suggesting ideas in pure math and theoretical physics.
- Some mathematicians reportedly worry about professional displacement; others see AI as primarily a force multiplier.
Broader AI Trajectory & AGI Debate
- One camp insists current models are just “stochastic parrots/token vomiters,” not autonomous or genuinely advanced.
- Another counters that these same systems are already contributing to math research and may scale to far beyond human capability, absent any “magic” in the brain.
- Debate arises over whether universal approximation theorems imply practical AGI, with pushback about energy, data, and architectural limits.
Legacy Code & Modernization
- Coding agents are praised for reviving or modernizing old Java applets and games into JavaScript/HTML, sometimes with surprisingly smooth workflows.
- A question remains whether agents truly handle large, messy legacy codebases well, given their need for context; this is left as unclear.