Report · danluu.com
We should be more ambitious with the software we build
danluu.com
Source published October 24, 2026
Summary: 2 min read
74/100
Summary
Dan Luu argues that increasingly capable coding models let individuals build software projects that were previously too costly in time or expertise, while warning that fast generation does not ensure correctness, performance, or maintainability. He says he built a customizable regex engine, a faster ripgrep fork for his workloads, and a Rust-related build-speed tool in a low single-digit number of hours with AI assistance; he estimates these projects would previously have taken a year or more to reach “works for me” quality. He says newer models have also made it easier to implement techniques from papers and codebases in specialized areas such as board-game AI. Luu reports that GPT-6 Astra passed registered zstd implementation tests more than 99% of the time in his evaluations and cost about $1 per implementation, though he says more ambitious tasks such as IMAP and Pandoc remained useful tests. He advocates building systems with self-checking properties, including fuzzing, property-based testing, consistency checks, and formal verification, rather than relying on one-shot AI output. Yossi Kreining cautions that most users should moderate their ambitions because many problems lack clear pass/fail criteria and AI-generated systems can become opaque, brittle, and hard to maintain. Luu also says pgrust is not yet suitable as a primary production database, citing founder Michael Malis’s September 17, 2026 statement that its query engine still has regressions and unsupported complex queries.
Key Takeaways
- Dan Luu says coding models have reduced some personal software projects from a year or more of work to a few hours, but product-grade reliability still requires substantially more effort.
- Luu reports that GPT-6 Astra passed all registered zstd tests more than 99% of the time in his evaluations, including under conditions that impaired performance.
- Luu argues that testing, benchmarking, fuzzing, property-based testing, consistency checks, and formal verification can make AI-assisted development more reliable.
- Yossi Kreining says most people should be less ambitious with AI-built software when they cannot establish clear correctness criteria or understand and maintain the resulting systems.
- Pgrust founder Michael Malis said on September 17, 2026 that he would not recommend pgrust as a primary production database because reliability evidence, query support, and performance regressions still needed work.
What the discussion said
The discussion largely embraced the idea that AI should raise the ceiling of what individual developers attempt, not merely automate familiar chores. Several commenters described LLMs making long-deferred projects viable: a substantial game remake, an ETL system, vertical-specific commerce tooling, and work across unfamiliar parts of a stack. Recent Claude models drew especially strong praise, with reports of handling difficult code transformations and producing large pull requests that need only small corrections. One participant also described using an LLM inside a formal-verification workflow to establish behavioral equivalence between C and Rust at considerable scale, suggesting a more serious role than disposable prototype generation. The enthusiasm came with an insistence that fast MVP creation is not the same as product creation. Readers stressed that integrations, tradeoffs, maintenance, user variation, and distribution remain stubbornly human problems. Others worried that calls for greater ambition can become a license for low-quality AI-generated software, while benchmark claims around an AI-adjacent rewrite project prompted distrust after accusations of query-specific optimization and caching. A smaller but meaningful thread focused on craft: AI can be faster, yet some developers still find hand-coding more absorbing because they retain a complete mental model of every decision. The room therefore favored ambitious AI-assisted building, but not blind faith in generated output or flashy measurements.
Where opinion split
The central fight is whether AI-assisted speed creates genuinely more ambitious software or simply floods the world with brittle, low-effort output. Optimists argue that capable models let a single developer tackle projects previously blocked by time, missing expertise, or implementation scale; skeptics answer that MVP generation does not solve real product complexity and can reward slop, overfitting, and misleading benchmarks.
Coverage & discussions
Hacker News · danluu.com
Community discussions
Community Sentiment
Positives
- Recent Claude systems are reported to complete demanding code rewrites and complex pull requests with only minor review feedback, making AI assistance feel production-relevant rather than merely demonstrative.
- LLMs are unlocking projects that developers had postponed for years, from game remakes to high-throughput data pipelines, by filling gaps across languages and domains.
- AI-assisted formal verification is being used to generate proofs about C and Rust behavior, pointing toward a role for models in improving confidence in ambitious systems code.
- Rapid implementation lets developers spend effort on neglected ideas and broader technical experiments instead of treating existing scope as a fixed ceiling.
Concerns
- Prompt-built MVPs do not eliminate the difficult work of product tradeoffs, integrations, customization, multi-user requirements, and market communication, so generated code alone rarely creates durable value.
- Some developers find AI coding less satisfying than direct implementation because it weakens their sense of ownership and complete understanding of the system.
- The push for bigger AI-built projects risks normalizing low-quality generated software, especially when speed is mistaken for engineering judgment.
- Claims around AI-adjacent rewrite performance face credibility problems when benchmark results may rely on workload-specific hand tuning and caching rather than general capability.
