Are LLMs good software engineers?
summary
Argues LLMs are not reliable software engineers: they are probability machines that rush into implementations, often reinvent the wheel, and ignore rules; dataset and benchmark issues lead to misleading scores — SWE-bench results inflate performance because models memorize repositories (SWE-bench memorization issue), Claude claims 93.9% on SWE-bench but obfuscation slashes success. Observed facts: 72% of AI PRs not merged, 80% of non-owner AI PRs merge without human review, a toolchain includes the ultra-fast analysis engine (Wily tool rewritten in Rust) and a “spray bottle” of linters, tests and multiagent reviews; one agent finds an ARM hashing trick that can improve Git CPU performance by ~50% but still needs human verification.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 2:17LLMs are not good software engineers
- 6:49LLM-generated code traits
- 11:00AI pull requests impact
- 18:30Spot issues in AI content
- 21:18SWE-bench flaws
- 24:53Why use AI despite flaws
- 31:18Back pressure for AI output
- 33:44Case study: AI pull request
- 38:18Measuring AI patch complexity
- 43:36Multi-agent reviews & gates
- 47:52Projects rejecting AI code
- 49:56Summary & best practices
- 50:18Q&A
related talks
32mWhat can we expect of LLMs as software engineers?
01h 01mLike Having an Intern? The Impact of LLMs on Software Engineering Careers
29mGenerative-AI: Usecase-Specific Evaluation of LLM-powered Applications
24mWill an LLM replace me? A Systematic study of Gen AI's impact on data jobs
38mNobody Pages the LLM: Engineering Rigour for Vibe Coding
14mSAG Expert Interview #03: Impact of AI in Software Architecture
25mEvaluating An LLM's Ability To Code
44mProduction-Grade LLMs & AI Agents