distillation

AI distillation, a technique where models are trained using outputs from other AI systems, is becoming a competitive tool that threatens the profitability of major AI companies. US firms like Anthropic, OpenAI, and Google warn that rivals, including Chinese companies like Alibaba and Z.ai, may use distillation to replicate their models cheaply, undermining investments in data and computing power.

Anthropic is removing hidden code from its Claude Code system that was used to detect unauthorized resellers and Chinese AI competitors by hiding data in system prompts. The code, which used steganography and checks for specific hostnames, was criticized for lacking transparency despite not being malicious. The company cites stronger mitigations like behavioral fingerprinting and access controls as reasons for discontinuing the covert measures.

Elon Musk testified that xAI trained its Grok model using OpenAI's models. The article highlights 'distillation' as a method frontier labs use to prevent smaller competitors from replicating their AI models.

The article explains the process of an oil refinery, detailing how crude oil is refined into products like gasoline, diesel, and kerosene through methods such as distillation and cracking. It provides an educational overview of the technical steps involved in refining.

The US accuses China of conducting 'industrial-scale' AI intellectual property theft through distillation attacks, with companies like OpenAI, Google, and Anthropic alleging Chinese firms cloned their models. China has dismissed these claims as 'slander,' while the US government warns such activities threaten American AI leadership.