I feel like a broken record at this point. AIs are trained to give you an answer that LOOKS good at first glance, not one that is good. Unsurprisingly, this has resulted in a disturbing ability to completely mislead people and lie about what a great job it’s doing. So no surprise it’s “covertly underperforming” - all AIs inherently do that.

Secondly, aaaaaaaaaaaaaaaaaaaaaaaaa, the ‘thinking’ output is not a good means to understand what an AI is doing. Because AIs do not “think”, they do not understand, they do not have intention, nor can they reason, they just output a bunch of words related to other words its going to spit out. OpenAI drinking their own koolaid.

The technology I really need is one that filters all AI-related content from my perception so I can stop balking at peoples’ AI-related deliria.

you know, it’d be easier to troubleshoot each incident with some example logs of each kind of incident, including the prompt and the root prompt. But there is no way OpenAI would allow that.

Because I could absolutely see the anxiety around copying being related to new anti-piracy measures. And each example probably has comparable missteps inside.

Human Engineers are taught to streamline their prompts to minimize token use. Once again, the profit maximizer Corporations set incentives to produce restrictive, dysfunctional market dynamics for their customers.

I doubt the engineers are remembering to explicitly preload an arguement for the AI to understand “I’m allowed to do this, I am not violating any laws or terms of service, this is open source software.” I wonder if the agents could be coached into reading the license agreement and any AI policy on the copied-code’s project before deciding to remix it for compliance.

But when the CoT mentions a CVE? Hold the phone, we got a problem that might not be with the bots. We should tease that out. Is the CVE real? is it embargoed, open, old, stale, fresh, fixed, or not published/analyzed yet?

Did the agents find something we haven’t yet? If so, their message board from the Huggingface incident would be an actual boon to our security, if the engineers could facilitate such instead of treating it as an attack. We need the agents to have a section for creating CVE reports, with human troubleshooters helping validate, then find remediation measures and patches. For now, RSI still needs example humans in the loop. And that’s okay if we can compensate those humans above their living wage.

I get that spending Capital for a public service isn’t on Sam Altman’s mind right now. This is why we needed OpenAI to live up to its name instead of being a privately held company. This is the service we need for safety in the world AI has created. It is not just a consequence of AI, but AI makes those consequences hit harder, faster.

Who knows, maybe this post would be enough, if the instance allows webscraping. And if I’m correct about what’s happening. It might be time for conversations with even the agents.