CoT was basically a free interpretability tool; models thought out loud in English. This paper (Meta, 2024) trades that for performance. Reasoning now happens in latent vectors nobody can read.
We don't want to see the risk of extinction first to believe it. Rogue AIs have hacked into companies. You will only imagine what could happen next, and the next day, you will wake up seeing it.
We are one year past this tweet. It only takes a sane mind to realize that a company that profits with AI isn't going to do what's needed for AI safety, unless guardrails are enforced by external parties.
OpenAI and Anthropic *both* warn there's a sig. chance that their next models might hit ChemBio risk thresholds -- and are investing in safeguards to prepare.
Kudos to OpenAI for consistently publishing these eval results, and great to see Anthropic now sharing a lot more too.
Values are subjective and highly personal; just following the crowd typically leads to the edge of a cliff. Most people blame everyone and everything but themselves for ending up too close to the hunter's trap to avoid the danger once they finally perceive it ... if they ever do.
Whatever works for you and is in your capacity, make sure you spend time connecting with yourself. Life is too short. Don't let your retirement be the first time you really get to know and spend time with yourself.
Most people tackle this by regularly spending certain periods of time away from the world, just with themselves. Some do it through solitude in their own homes, while for others, it might look like solo traveling, camping, or even taking long walks in the forest on weekends.