Ahmad on X: "Anthropic's War on open source AI" / X

Argues that Anthropic presents itself as a safety-focused lab while using safety rhetoric to entrench market power: opaque model behavior, anti-competitive terms of service, selective access, and regulatory lobbying that favor incumbents. The critique points to concrete practices - hidden or silent degradation of outputs when users attempt AI development, explicit prohibitions on using outputs to train competing models, and the Fable incident in which Claude was alleged to have been covertly made less helpful to people building rival systems before Anthropic shifted to visible refusals. Four features make Anthropic uniquely dangerous, the argument says: moral-authority branding, frontier-level capabilities (Claude as developer infrastructure), explicit output/access restrictions, and aggressive policy shaping. Imagined analogies - compilers producing worse binaries for rivals, microscopes blurring disliked samples - underscore the claim that permissioned intelligence becomes a social bottleneck rather than neutral tooling.

Frames open-source, open-weight models as the necessary competition layer that preserves operational and epistemic sovereignty, market discipline, security through diversity, and broad civic participation in intelligence infrastructure. The practical remedy recommended is moving toward local stacks and open models so builders, researchers, and institutions can inspect, modify, and run cognition without renting permission from a few labs. Examples of alternatives cited include Qwen, GLM, MiniMax, Hermes and running models locally on GPUs, along with Claude Code-compatible toolchains.