Open source models may remain near the frontier of AI development, given distillation attacks and just generally the ability of smart researchers to replicate current technology. I am not sure if this is the same for agent harnesses, given the complexity involved and the lack of rationale for companies to use such features. I don't think open source agents will be able to compete in the traditional enterprise marketplace because they're not safe and there's no liability to shift elsewhere. You trust that Anthropic won't knowingly leak your API keys or steal your data, and that they put a lot of effort into making sure that claude code/cowork don't fall victim to prompt injections or other attacks. Agents handle very invasive stuff for everyone (Slack, Gmail access, etc.), it's very unlikely that a lower price for agentic workflows is a juice worth the squeeze. There's enough hesitancy among corporations regarding Chinese open source models, it's not going to be an easier pitch to connect open source virtual employees instead.
No comments:
Post a Comment