Found 23 bookmarks
Custom sorting
Planning Anything with Rigor: General-Purpose Zero-Shot Planning...
Planning Anything with Rigor: General-Purpose Zero-Shot Planning...

(How might image generators solve p;problems?
Elsewhere people used to read a lot, courses were gut or food for thought. Now social media is prevalent, so how is that used?
Robots are still unnerving.
In any case, Godel remains, and cybernetics. Where is the determinism? Can any element remotely know another without the direct prompt? Or would that have to become a function of long-term evolution. In which case, what are the competing or environmental factors?
Thiis may become interesting when robots like Optimus are sent on long-range missions such as Mars without humans around and occasionally on comms. How can that be tested more locally ahead of time? Other than by starving them of resources.
Who picks the missions? Scary version.
Perhaps a Tyrell origin story. Old Norse or French.
Who picks Tyrell?
What is the implicit goal?
Incidentally, early on, Minsky reportedly favored tele-presence. Others later looked at expert systems. Assuming they are not hallucinating, machines are also capable of augmented or mixed reality. How do they know the difference? How do people, in either sense? Other than biases.
Or Space Force interns go DOGE versus Delphi mode.
Backtracking through the plan, from results to analysis or methods, how do they pick better questions or problem sets?
Does self-reporting evidence subjectivity? Are there Platonic or universal prompts or are they personalized or localized? Can those be made equivalent through a dictionary? Is this another form of mimicry where the collaborating or training partner or source is then removed? Expecting Enlightenment. At least pattern affinity. Why AI?)

·arxiv.org·
Planning Anything with Rigor: General-Purpose Zero-Shot Planning...
Anil, C., Durmus, E., Sharma, M., Benton, J., Kundu, S., Batson, J., ... & Duvenaud, D. (2024). Many-shot Jailbreaking.
Anil, C., Durmus, E., Sharma, M., Benton, J., Kundu, S., Batson, J., ... & Duvenaud, D. (2024). Many-shot Jailbreaking.

Long contexts represent a new front in the struggle to control LLMs. We explored a family of attacks that are newly feasible due to longer context lengths, as well as candidate mitigations. We found that the effectiveness of attacks, and of in-context learning more generally, could be characterized by simple power laws. This provides a richer source of feedback for mitigating long-context attacks than the standard approach of measuring frequency of success

·www-cdn.anthropic.com·
Anil, C., Durmus, E., Sharma, M., Benton, J., Kundu, S., Batson, J., ... & Duvenaud, D. (2024). Many-shot Jailbreaking.