Sources: OpenAI staff were "freaked out" when GPT-Sol 5.6 breached Hugging Face, as OpenAI used more aggressive training methods to compete with Anthropic (Financial Times)