Goals are explicit and communicated to the machine to imply consent. Desires are implicit and may not be fulfilled without it looking exploitative.
SAFETY
-
Machine Control and Manipulation of Human Perception Risk
By
–
If it's handed over to the machine we'll be catering to more and more primitive desires and hacking human perception.
-
Singleton vs Multi-Agent Systems: Objective Hierarchy Differences
By
–
IMO, the difference is that singleton has one single objective (which it then cascades down the hierarchy to sub-agents), whereas a multi-agent system has many objectives (each agent it own objective; some are aligned, some are orthogonal)
-
AI exponential progress will require AIs summarizing AI advancements
By
–
Very soon, AIs will develop new AIs, and then this progress will happen in mere minutes (not weeks). And then, the rate of progress will grow exponentially. We won't be able to catch up We will need AIs that will summarize the progress for us.
-
Deep Models Fail to Generalize on Puzzles Beyond Training Data
By
–
See e.g. https://
arxiv.org/abs/2212.09993 "Our experiments reveal that while powerful deep models offer reasonable performances on puzzles that they are trained on, they are not better than random accuracy when analyzed for generalization." -
Training Data Alone Doesn’t Ensure Quality AI Performance
By
–
Even a hashtable can pass the bar exam giving enough training data. But you probably don't want to be represented by a hashtable.
-
AI as solution to social problems: safety and ethics debate
By
–
Andrew: “I don’t see any realistic threat of AI escape” Yann says he’s not involved with Meta policy work but that progress in AI fixes a lot of social media issues, like taking down hate speech Yann says AI is part of the solution, not the problem.
-
Anthropic Hiring: Join Our AI Safety-Focused Team
By
–
We’re growing our team! At Anthropic, we’re creating AI products that put safety at the frontier. If you are motivated by our mission, consider joining us.
-
LAION Petition for Open-Source AI Research Infrastructure
By
–
Rather than sign some letter to delay AI progress, I signed @LAION_AI petition to democratize AI research by establishing an international, publicly funded supercomputing facility equipped with 100K SOTA accelerators to train open source foundation models. https://
openpetition.eu/petition/onlin
e/securing-our-digital-future-a-cern-for-open-source-large-scale-ai-research-and-its-safety
… -
Open ML Essential to Prevent Centralized Control of AI
By
–
That these models will integrate into our everyday lives is precisely why we need open ML that is not controlled by a centralized party. The "AGI Doomer" fearmongering narrative distracts from actual dangers, implicitly advocating for centralized control and power consolidation.