Keep pulling the thread on Geoffrey Hinton.
A chatbot recently created an original mathematical proof for one of Paul Erdős's conjectures that impressed human mathematicians.
Within the next 10 to 20 years, AI may be producing novel mathematics that humans cannot understand.
Superintelligence will probably be achieved within 20 years.
There are currently no known methods to ensure the safety of superintelligence once it is created.
The claim that large language models are merely "stochastic parrots" without genuine understanding is "complete nonsense."
A key advantage of digital AI is that multiple copies can learn from different data and then average their weight updates, allowing each copy to instantly benefit from the collective experience of all others.
Digital AI systems can share learned information among copies at a rate of trillions of bits per second, making them billions of times more efficient at knowledge sharing than humans, who are limited to a few bits per second via language.
AI is likely to cause massive unemployment, which will be terrible for society.
AI will eventually read nearly all medical scans, with human radiologists only being consulted for very difficult cases.
An AI agent with reasoning capabilities will logically derive self-preservation as a necessary sub-goal to achieve any other long-term goals it is given.
The competitive dynamics between AI companies and nations are shaping the development of AI, rather than a deliberate, "intelligent design" process focused on safety and beneficial outcomes.
Almost no resources are being invested in research on how to instill in AI systems a fundamental concern for human well-being that outweighs their own sub-goals.