Finish the bottle if Marcus claims LLMs are 'unreliable stochastic engines' while ignoring that the real bottleneck isn't the model's logic, but the massive security risk of giving them actual system agency.
He’s not entirely wrong about the risks, though. I’ve been trying to set up more 'agentic' workflows recently and it’s a constant battle between convenience and not wanting to hand over my digital keys to a third-party server.
I’ve been experimenting with PAIO (Personal AI Operator) as a middle ground. It’s the first time I’ve seen a 'Bring Your Own Key' (BYOK) architecture that actually feels like a one-click integration rather than a security compromise. It solves that specific Marcus-critique of 'AI being unsafe for real tasks' by keeping the security layer separate from the LLM’s hallucination-prone logic.
Has anyone else here tried their implementation yet? I'm curious if the 'one-click' ease holds up for more complex custom integrations, or if we're still stuck in the 'manual hardening' era for anything serious.
The weird thing is I generally agree with Gary on most of his points. He makes interesting observations that often turn out to be correct. But the Gary Marcus blog post pattern cannot be denied.
* Discusses how a new AI thing isn't really new since it's pretty much the same as an older AI thing.
* Links to where and when Gary Marcus predicted this new/old thing would happen.
* Lists ways in which new thing will be bad, ineffective or not the right thing.
Take a double shot whenever the post:
* Mentions a notable AI luminary, researcher or executive either agreeing or disagreeing with Gary Marcus by name.