Five Small Models Beat a Frontier Benchmark, Karpathy Rescues Buried Video, and Anthropic Draws a Hard Line

A GitHub project called Memdoor demonstrated that a coordinated team of five Qwen models — ranging from 3.8 to 27 billion parameters — can collectively outperform Fable 5 on the LiveCodeBench Hard benchmark. The system does not run a single large model but orchestrates a workflow in which each smaller model holds a specific role: one decomposes the problem, another writes initial code, a third critiques it, a fourth handles edge case analysis, and a fifth arbitrates conflicts. The aggregate result apparently surpasses what was previously considered state-of-the-art on that benchmark, earning 257 points and 119 comments on Hacker News, many of them skeptical in ways the community found productive.

Legitimate questions arose about whether multi-agent systems are solving problems or optimizing against benchmark structure — a concern that applies to single-model results as well. The more durable insight, however, may be economic rather than scientific. Running five 27-billion-parameter models sequentially for a complex coding task remains substantially cheaper than running a single frontier model with hundreds of billions of parameters. For teams building developer tools or coding assistants, that cost differential matters at production scale regardless of the benchmark's imperfections.

The highest-scoring item of the day — 1,425 points and 385 comments — was Andrej Karpathy's HackerTube, described as YouTube built from videos that Hacker News users have recommended over the years. Karpathy scraped HN comment threads and submissions for YouTube links that received meaningful upvotes, assembling them into a curated video platform. The result is reportedly an extraordinary collection of technical lectures, conference talks, and educational content that algorithm-driven YouTube recommendations have largely buried — videos invisible through YouTube's own search because their channels had low subscriber counts or their titles were not keyword-optimized. The project implicitly argues that social curation by a focused community outperforms engagement-optimized algorithmic recommendation for educational content, and the comment thread was filled with users reporting rediscoveries of lectures they had been unable to locate through conventional means.

On the platform governance front, OpenAI merged ChatGPT and Codex usage limits into a single unified system — a change framed as administrative simplification but interpreted by professional developers and API users as a resource squeeze. Codex, OpenAI's code-specialized model offering, had previously operated on a separate usage tier, allowing heavy Codex workloads to run without drawing down ChatGPT capacity. The merger eliminates that separation. The 92-comment thread on Hacker News skewed toward developers whose production workflows had relied on Codex having distinct headroom, with several characterizing the change as a revenue optimization measure pushing heavy users toward higher tiers.

Anthropic generated sharper tension after reports emerged — amplified by a tweet from researcher Jia Hao Shao — that the company was terminating API accounts whose usage patterns suggested training larger models on Claude's outputs. The policy has existed in Anthropic's terms of service, but enforcement appears to have escalated significantly, with some users receiving account terminations for queries that did not obviously violate the rules. Anthropic's concern — that training competing models on frontier AI outputs creates capability leakage undermining both competitive position and safety-conscious development — has a coherent rationale. The 217-comment thread nonetheless revealed a fault line between those who view model output protection as legitimate intellectual property enforcement and those who argue that pattern-matching on API usage is too blunt and too prone to false positives, particularly for academic researchers whose work can resemble training data collection.