Social Media Briefing

Swyx offers a conceptual essay on Loopcraft, arguing the central skill of the coming era is stacking loops effectively, knowing when to descend a loop for reliability and when to ascend for leverage as models improve.

Mollick demonstrates Claude Code with the Fable model rebuilding the lost SimRefinery game from screenshots and docs, fully playable with a learning mode, contrasting it against an attempt 10 months prior.

Mollick highlights a study finding that frontier general-purpose LLMs outperformed dedicated clinical AI tools like OpenEvidence across three evaluations, with clinical tools performing on par with auto-enabled Google Search overviews, despite 65 percent of doctors reportedly using OpenEvidence.

OpenAI announces the ability to save Codex rate limit resets for later use, starting with one free reset for several user tiers.

Summary of a Carnegie Mellon benchmark called SusVibes testing whether coding-agent output is secure, finding that while SWE-Agent on Claude 4 Sonnet passed 61% of functional tests, only about 10% of solutions were secure and over 80% of working code contained vulnerabilities; prompt-based fixes barely helped.

Swyx explains his motivation for building his own vibecoding platform, frustrated that existing deploy platforms do not close the loop on errors and failure alerts, requiring too much per-project webmaster infrastructure he wants consolidated into one thing.

Mollick cites a paper finding frontier LLMs outperformed clinical AI tools like OpenEvidence across three evaluations, with clinical tools comparable to Google Search AI Overview.

Delangue argues AI evals are broken because they structurally favor closed-source APIs that can route, fallback, and ensemble opaquely, questioning the fairness of comparing one model to two.

NVIDIA introduces AgentPerf, an agentic AI infrastructure benchmark from Artificial Analysis, claiming Blackwell delivers 20x more agents per megawatt than Hopper.

Following yesterday's Reddit buzz about the planned MiniMax M3 release, vLLM details MiniMax M3's features: frontier coding/agentic capabilities, native image/video input, computer use, 1M-token context, and MSA sparse attention scoring 128-token KV blocks, with day-0 vLLM support across NVIDIA and AMD hardware.