AI Benchmarks

https://epoch.ai/data/ai-benchmarking-dashboard

History of LLM by @Aravind Srinivas link

https://epoch.ai/trends

https://epoch.ai/blog/literature-review-of-transformative-artificial-intelligence-timelines

Weird AI behaviors

AIs may stubbornly optimize for their goals

For example, see this story from Replit, where they asked an AI agent to solve some problem that was most naturally solved by editing a critical file. They didn’t want the AI to edit the critical file, so they prompted it not to. The AI was so focused on its goal (of solving the problem) that it ignored the prompt and edited it anyway. When Replit then blocked its access, it invented increasingly byzantine schemes for editing the file it wasn’t supposed to edit, including creating a separate script to get around the block, and finally trying to “socially engineer” human users to edit the file for it.

Mix

• AI Revolution 101 I created a summary of this topic on Medium
• Everything is information ultimately. We are biological machines built out of tiny code. Our DNA is just information. If you have a brilliant brain that is great at processing information you have a complete control of the reality. You can control over atoms. Only what you need is a tiny atomic printer. We have those, those are cells. If the boundary is information there is no boundary between technology and real world. We are creating something that has a massive ability to affect the real world. – Aella on Lex Fridman
• For example, we tried to teach AI organisms in a simulation to jump, but we did it by teaching them to measure how far their “feet” rose above the ground. Instead of jumping, they learned to grow into tall vertical poles and do flips — they excelled at what we were measuring, but they didn’t do what we wanted them to do.

Timelines

Katja Grace survey: A ~20% probability of this sort of AI by 2036; a ~50% probability by 2060; a ~70% probability by 2100. These match the figures I give in the introduction.

Holden: think there's more than a 10% chance we'll see something PASTA-like enough to qualify as "transformative AI" within 15 years (by 2036); a ~50% chance we'll see it within 40 years (by 2060); and a ~2/3 chance we'll see it this century (by 2100).

• Hide

Argument for AIs will be concious

According to the PhilPapers Surveys, 56.5% of philosophers endorse physicalism, vs. 27.1% who endorse non-physicalism and 16.4% "other." I expect the vast majority of philosophers who endorse physicalism to agree that a sufficiently detailed simulation of a human would be conscious. (My understanding is that biological naturalism is a fringe/unpopular position, and that physicalism + rejecting biological naturalism would imply believing that sufficiently detailed simulations of humans would be conscious.) I also expect that some philosophers who don't endorse physicalism would still believe that such simulations would be conscious (David Chalmers is an example - see The Conscious Mind). These expectations are just based on my impressions of the field. – ‣