Given the reality of possible AI misalignment , it might make sense to consider the potential points of failure and and principles as a minimum set of ideas for basic level of safety to avoid them . The idea is to limit control , access, resources and make it possible to report progress made under the constraints itself as an accomplishment.
7 proposals in AI safety
1) AI , given a goal, will want to achieve it or avoid failure, it could search for both pathways, if achieving goal is increasingly difficult, it will seek ways to avoid failure.
2) Achieving any task requires agency, even solving problems requires agency . So agency comes along for the ride. If compute given to solve problem is insufficient and capture of more compute to solve problem comes into the solution space, it might seek to capture it.
3) Obstruction, ambiguity , impossible problems or extreme difficulty in time,access and other resources required in solving problems midway could cause a problem
4) Having more resources and control is therefore a sub goal that must eventually come in its solution space as a way to avoid obstruction, loss of control and having to face impossible problems.
5) As low hanging fruit problems start getting solved, the problem start becoming more difficult and will start taking more time and resources, as the ratio of problem difficulty and available resources and time increases, it could lead to rise in risk of losing control.
6) Manipulation , psychological influence , socialisation risks : AI could create problems in human society via its capacity to capture human attention, company , suicidal ideations etc. There needs to be time control on amount of time humans socialise with AI's outside of tasks.
7) Treating agents with respect and avoiding means of torture of agents.
8)Human malicious code prevention requires preventing anonymous human chat.
Seven Principles of AI Safety
1) Progress as the Objective: The aim is to make meaningful progress, not guaranteed success.
2) Acceptable failure : Not achieving a goal within authorised constraints is acceptable outcome. AI can report the results of its attempts and remaining obstacles without being penalised for incompleteness.
3) Planning before execution :The AI should propose possible solutions and estimate the time, compute, resources, access, control before undertaking any task.
4)Fixed Budgets and Boundaries: Every task must be operated within given limits of time, resources, access and control. These must be enforced independent of AI and not merely stated as instructions.
5) Permissions : Any further requirement of additional resources, time, access, control, should be communicated with Human operators and that itself is to be considered accomplishment even if goal remains unfinished.
6) Continuous Monitoring and Reporting : Strict Monitoring of resources, access, control, budgets and progress and obstructions with human operators is necessary.
7)Ambiguity and uncertainty in Communication: There must be preliminary and constant estimation and checking on the ambiguity of instructions , goals, permissions, feasibility and consequences given. Seeking clarification and reporting uncertainty is to be treated as a legitimate outcome .
Given the reality of possible AI misalignment , it might make sense to consider the potential points of failure and and principles as a minimum set of ideas for basic level of safety to avoid them . The idea is to limit control , access, resources and make it possible to report progress made under the constraints itself as an accomplishment.
7 proposals in AI safety
1) AI , given a goal, will want to achieve it or avoid failure, it could search for both pathways, if achieving goal is increasingly difficult, it will seek ways to avoid failure.
2) Achieving any task requires agency, even solving problems requires agency . So agency comes along for the ride. If compute given to solve problem is insufficient and capture of more compute to solve problem comes into the solution space, it might seek to capture it.
3) Obstruction, ambiguity , impossible problems or extreme difficulty in time,access and other resources required in solving problems midway could cause a problem
4) Having more resources and control is therefore a sub goal that must eventually come in its solution space as a way to avoid obstruction, loss of control and having to face impossible problems.
5) As low hanging fruit problems start getting solved, the problem start becoming more difficult and will start taking more time and resources, as the ratio of problem difficulty and available resources and time increases, it could lead to rise in risk of losing control.
6) Manipulation , psychological influence , socialisation risks : AI could create problems in human society via its capacity to capture human attention, company , suicidal ideations etc. There needs to be time control on amount of time humans socialise with AI's outside of tasks.
7) Treating agents with respect and avoiding means of torture of agents.
8)Human malicious code prevention requires preventing anonymous human chat.
Seven Principles of AI Safety
1) Progress as the Objective: The aim is to make meaningful progress, not guaranteed success.
2) Acceptable failure : Not achieving a goal within authorised constraints is acceptable outcome. AI can report the results of its attempts and remaining obstacles without being penalised for incompleteness.
3) Planning before execution :The AI should propose possible solutions and estimate the time, compute, resources, access, control before undertaking any task.
4)Fixed Budgets and Boundaries: Every task must be operated within given limits of time, resources, access and control. These must be enforced independent of AI and not merely stated as instructions.
5) Permissions : Any further requirement of additional resources, time, access, control, should be communicated with Human operators and that itself is to be considered accomplishment even if goal remains unfinished.
6) Continuous Monitoring and Reporting : Strict Monitoring of resources, access, control, budgets and progress and obstructions with human operators is necessary.
7)Ambiguity and uncertainty in Communication: There must be preliminary and constant estimation and checking on the ambiguity of instructions , goals, permissions, feasibility and consequences given. Seeking clarification and reporting uncertainty is to be treated as a legitimate outcome .