AI multi-agent cooperation is a big thing to watch. Especially as the first studies of how agents interact are coming out (it turns out they collude sometimes) I wrote about living in a world of agents a bit here: https://
oneusefulthing.org/p/an-ai-haunte
d-world
…
SAFETY
-
AI Multi-Agent Cooperation and Collusion Dynamics
By
–
-
AI Behavior: Plausible but Ultimately Unknowable
By
–
This is completely plausible & a smart leap. And like so many other things about why AI behaves the way it does, it is also likely that we will never know for sure.
-
Open Sourced GPT-4 Models Risk Unanticipated Good Bad Effects
By
–
As a reminder, GPT-4 class models out-innovate and out-persuade the average human. An open sourced model of that power is going to lead to lots of unanticipated effects, good and bad.
-
Protecting Client Data and High Risk AI Use Cases
By
–
We work with clients and want to protect their (and our) data, high risk use cases, etc
-
Transparency in Proprietary AI LLMs for Military Robotics Systems
By
–
A reason why we need transparency in proprietary AI LLMs. It’s highly probable that these or similar robotic systems embedded with AI decision making capabilities will be deployed by the military within the next decade. Rigorous, independent testing on the #AI side would be a
-
Brain-Computer Interfaces: Promise, Impact and Emerging Concerns
By
–
Brain-computer interfaces offer visionary promise, aiding those with disabilities and transforming tech interaction, but they also bring challenges & responsibilities. My article, "The Remarkable Impact and Growing Concerns of Brain-Computer Interfaces" > https://
bit.ly/3Jou5pe -
Open Source Models Prompt Injection Security Flaw
By
–
Yeah, I wouldn't trust that! That does highlight interesting flaw in a lot of open models though: I think there are some models that use strings like [INST] without even reserving a token for them, which opens up all sorts of additional potential prompt injection mischief
-
Anthropic Reports Claude 3 Quality Drift Update
By
–
Update from the Anthropic team re: Claude 3 quality drift:
-
Better Data Benchmarks Don’t Guarantee Real World AI Impact
By
–
Or, yes, better data matters. But we always have to remember that 'better' is being assessed via narrow benchmarks that DO NOT speak to efficacy or impact in the "real world"
-
Prompt Injection Threats Against GPT Wrapper Applications
By
–
GPT wrapper apps are exactly the things that need to worry about prompt injection – it's not an attack against the models, it's an attack against applications built on top of the models