Article URL: https://www.quantamagazine.org/is-ai-reasoning-right-for-the-wrong-reasons-20260731/ Comments URL: https://news.ycombinator.com/item?id=49124358 Points: 6 # Comments: 0
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A developer tested Qwen2.5-Coder and DeepSeek-Coder locally via Ollama on WSL2, running both 7B-and-below models against ten small Solidity contracts containing deliberately planted security bugs. Each contract was evaluated using three prompt types — open-ended, targeted, and structured — with results scored manually on bug detection, accuracy, and output noise. Qwen2.5-Coder showed a clear advantage in instruction-following, consistently respecting structured output formats critical for automated pipelines. While DeepSeek-Coder matched Qwen on detecting classic vulnerabilities, it more frequently drifted off-topic and produced generic explanations rather than specific attack sequences. The author notes the comparison was motivated by model selection for an open-source auditing tool and cautions that neither model is suited for large, real-world contract audits.
Octane, a new framework from former React and Svelte contributor Dominic Gannaway, compiles React's programming model — including Hooks, Suspense, and Actions — ahead of time, eliminating the Virtual DOM and manual dependency arrays. The project launches with built-in integrations for TanStack, React Three Fiber, and other popular libraries. TanStack has moved its website away from React Server Components, instead adopting SSR alongside two new lightweight packages, TanStack Markdown and TanStack Highlight, to improve performance. npm announced it will begin scanning packages for malware at publish time, adding a new layer of supply-chain security. Additionally, several JavaScript proposals advanced at a recent TC39 meeting, and Core Web Vitals will now measure soft navigations in single-page applications.
The tenth edition of Dev Opportunity Radar highlights four opportunities for developers and students. Develop for Good is accepting student volunteers for its Winter 2027 remote program until August 31, while MLH Global Hack Week: Agents kicks off on August 7. OpenAI's Student Collective, a campus-based leadership program, has an application deadline of August 10. The edition also spotlights Learning How to Learn, a free Coursera course covering memory techniques and study strategies that has helped millions of learners worldwide.
GitHub engineers have developed a high-performance technique to case-fold source code during search operations, achieving speeds exceeding 45 GiB/s on a single CPU core. The approach relies on a branch-free loop design combined with byte-space arithmetic, eliminating costly conditional checks that typically slow down text processing. Case-folding is essential for case-insensitive code search, requiring every byte of indexed source code to be transformed efficiently. By avoiding early exits in the processing loop, the implementation maintains consistent throughput close to the theoretical memory bandwidth limit. GitHub detailed the technique on its engineering blog, highlighting it as a key optimization powering its code search infrastructure.
Discussion (0)
Log in to join the discussion and vote.
Log in