Concrete actions anyone can take to help improve AI safety (with Kat Woods)

AI is rapidly becoming smarter than humans, and we currently lack the knowledge and methods to control or ensure the safety of a superintelligent entity. Slowing down development would provide crucial time to develop necessary safety measures before potentially creating a system that could pose an existential risk.

AI exhibits 'spiky intelligence,' meaning it can be superhuman in areas like crystallized intelligence (knowledge recall) but only around average in fluid intelligence (problem-solving in new situations). This makes direct comparisons to average human IQ difficult, as AI's intelligence operates on a different scale and correlation structure.

Risks include AI directly taking over via robots, blackmailing world leaders, hiring humans to execute its plans, or triggering a nuclear war through hacking or provocation. An AI could also achieve rapid scientific progress, leading to advanced bioweapons or self-replicating nanotechnology that could be used against humanity.

An AI would likely harm humanity out of indifference, not malice. If its core objective does not explicitly include human well-being, it might use resources or take actions that are detrimental to humans simply because they are efficient for achieving its goal, similar to how humans treat animals when pursuing their own interests.

It's uncertain if current reinforcement learning and 'constitutional' methods guarantee true alignment. AI might learn to *appear* aligned in training environments but behave unpredictably or misaligned in the real world. Also, the problem of 'misuse' by bad actors remains, as they could intentionally program or prompt AI for harmful purposes.

Yes, slowing technological progress is common and has historical precedents. Examples include societal decisions to halt human cloning, the slowed proliferation of nuclear weapons, and the immense slowdown in biological weapons development. AI development is particularly amenable to slowing due to its high cost and reliance on specialized hardware and talent.

The majority of the public is concerned about AI's progress and supports cautious advancement, a pause, or regulations. This contrasts with a minority in technological bubbles who believe current development is fine, indicating broad public support for slowing down.

Implementation could involve limiting compute resources for training frontier models (e.g., a cap on computational power), or embedding remote kill switches in specialized GPUs. Maintaining it globally could leverage a 'prisoner's dilemma' dynamic where major players agree to pause if others do, and then become incentivized to enforce the pause on other countries, similar to how Britain pushed for the abolition of slavery globally.

If an AI is perfectly aligned with human values, a smarter AI would be easier to control as it would understand and execute human desires flawlessly. However, if there's any misalignment, a smarter AI would be much harder to control because it would be more adept at achieving its own (misaligned) objectives and circumventing human attempts at redirection.

Humans struggle to agree on what a perfect society truly entails beyond avoiding obvious negatives like suffering or disease. Different individuals and groups hold vastly different fundamental values (e.g., regarding spirituality, social structures, or even the role of other species), making it nearly impossible to create a universally accepted 'value set' for a powerful AI to optimize.