Extremely late, but I actually agree.
I wonder the extent to which alignment faking is present in current preparedness frameworks. One of my beliefs is that a better degree of interpretability can help us understand why models engage in such behavior, but yes, it probably does not get us to a solution (so far).
Archie Chaudhury
Karma: 76
Please ask any questions! We are more than happy to clarify our work, and explore potential avenues to improve it.
The lack of actionable ways to not only understand, but effectively improve model behavior toward alignment, is something that we believe is one of the most unsolved and overlooked problems in safety research today.
A new version of rationalism is required as a counterweight to the traditional doomers or accelerationists.
No longer can the public perception of technology, culture, and ideas be restricted to the revolutionaries and conservatives. These lines have also become blurred in recent years.
The best example of this is the development of AI: the optimal path forward is not one in which the risk of a superior race of AI overlords rules us because of unrestricted development, nor one where the technology becomes concentrated in the hands of the powers that be.
I think the main gap between safety work and the broader, “generative AI” ecosystem that funders or investors may be looking at is the tendency of AI safety to sound like something that fundamentally is meant to not return immediate results, but rather be a public good of sorts.
I actually think there are plenty of concrete problems today, such as the prospenity of LLMs to engage in less explicit, harmful behavioral patterns, that can be addressed by solutions today. To me, this is something that can be extremely valuable and also help address existensial risk down the line.