7.28.26 | Our position on open-weights models, Benchmarking Opus 5 on SlopCodeBench, Using an open model feels surprisingly good

**SPEAKER_1** (0:00)
Welcome to the Hacker News Highlights, where we explore the top 10 posts on Hacker News every day. Today, we dive into OpenWeights Model's national security concerns, benchmarking Opus 5's coding progress, and experiencing the freedom of using an open inference model. Let's get into it.
Title, Our Position on OpenWeights Model, source, anthropic.com. The post by Dario Amadei, CEO of Anthropic, explains the company's stance on OpenWeights Models amid discussions about bans and regulations. He emphasizes that Anthropic has never supported a ban on OpenWeights models, but advocates for regulations like mandatory safety testing for models above a certain capability level, regardless of their origin. He expresses concern about authoritarian governments developing more powerful models for military and repressive purposes, and argues that protectionist bans won't address these risks, which he views as primarily related to chip access and industrial distillation. The post also criticizes the idea that banning open weights would slow China's progress and asserts that open models even from China foster competition and innovation. Overall, the article focuses on contrasting the company's regulatory stance with the risks posed by powerful models, especially those from state actors. In the comments, the community overwhelmingly viewed the post as hypocritical and self-serving. Many criticized Amadei for advocating restrictions that would limit competition and innovation while claiming to prioritize safety and ethics. Several highlighted the inconsistency in opposing open source models from China while benefiting from open instructions and distillation methods. Debates centered on whether regulation could effectively contain risks, with many asserting that such measures mainly serve corporate and geopolitical interests rather than genuine safety concerns. The dominant sentiment was that the post reflected attempts at regulatory capture and nationalist protectionism, with little trust or support for the company's motives or narrative.
Title, Benchmarking Opus 5 on SlopCodeBench. Source, GitHub. The post discussed testing Opus 5 on SlopCodeBench, a benchmark that measures a model's ability to handle incremental real-world coding tasks over multiple checkpoints. Opus 5 achieved a strict pass rate of about 24%, slightly better than Opus 4.6's 17%, but still far from perfect. The writer highlighted that most models showed increased code complexity and verbosity through the checkpoints, implying current models can't be trusted to manage long-term software projects without guidance.
Overall, the results showed incremental progress, but reaffirmed that models still struggle with maintaining code quality over time. In the comments, the community mostly expressed skepticism about Opus 5's improvements and questioned whether benchmarks like SlopCodeBench reflect true coding ability. Some argued that Opus 4.8 was better than Opus 5, and others discussed how the models tend to produce more verbose, complex and duplicated code as they progress.
Several users debated how to better measure a model's maintainability and how to improve models' self-editing and refactoring skills. The broader consensus was that current models are not yet reliable for long-term, lights-off software development, but SlopCodeBench offers a useful signal for future progress.
Title, Using an Open Model Feels Surprisingly Good. Source, matthewsaltz.com. The writer shared their experience of running open code models on their own inference endpoint. They found it to be freeing and satisfying because it gave them full control over their data and infrastructure. They described the process of quickly setting up open models on their own hardware, which felt like a breath of fresh air compared to relying on commercial providers. In the comments, the community largely supported the idea of open models for privacy and control, with many praising how easy and cost-effective it became to build personalized AI tools. Users debated whether open models truly matched the performance of large commercial models, but generally believed they were good enough for hacking and small projects.
Several discussed the importance of harnesses and infrastructure for open models, and noted that many saw these models as a foundational technology similar to the Internet with the potential to lower barriers for individuals and small teams. Dissenters questioned whether the post was just marketing for the author's company, or noted that open models are still more expensive and less capable than top-tier closed models. Overall, the community reflected excitement about the freedom and privacy open models offered, balanced with realistic views on their current limitations.
Title, Top Stories and Discussions from Hacker News Issue Hash 9 Source, PagedOut.Institute. The Hacker News Issue Hash 9 covers a wide range of topics including innovative cryptography methods like anamorphic cryptography, new hardware exploits on old and decommissioned devices, advances in reverse engineering techniques such as analyzing flirt signatures and entropy shifts in x86 instructions as well as recent vulnerabilities in popular libraries like PIPDF and attack vectors involving TLS memory extraction. Additionally, it features projects on building a computer with simple tools, exploring sub-pixel rendering challenges, and discussions on the reliability of online assessments versus real world conditions. Community comments express fascination with the depth of technical content, curiosity about the practical applications of complex algorithms, and a shared fondness for the humor and creativity in posts such as Baby Steps in Sea and the Sub-Pixel Zoo, highlighting a broad enthusiasm for security, reverse engineering, and programming in unconventional ways. In the comments, the community largely praised the mix of serious technical explorations and humorous takes, with some debate around the practicality of certain cryptographic schemes and reverse engineering methods. Discussions also touched on visual display complexities like sub-pixel renderings' impact on text clarity and the implications of security flaws like the Android screen reader bug or the Wi-Fi patching techniques. Overall, the tone was highly engaged, with experts and enthusiasts exchanging insights, highlighting a curiosity-driven and playful attitude towards intricate systems and security research.