5 ms·
Has ANYONE been able to articulate what the "safety" concern is or more likely is it just companies trying to build an artificial moat around a statistical mode
by kylebenzle 3y ago
Has ANYONE been able to articulate what the "safety" concern is or more likely is it just companies trying to build an artificial moat around a statistical model built on free/stolen data?
Currently the for-profit models are Nerfed to not allow bigoted or sexist (against women) text to be generated but that seems to be it. What kind of text are we being protected from?
- hollerith 3y agoThe main danger (but certainly not the only danger) is that the AI kills us all for reasons that have been articulated for 20 years including in books by respected academics like Nick Bostrom. The techniques used to 'steer' or 'aim' the AI stop working once the AI becomes better than humans (and human institutions like the FBI or Microsoft) at certain cognitive capabilities like making intricate plans that will succeed in spite of determined human opposition. Researchers employed full-time at the Machine Intelligence Research Institute have been trying to solve this 'control' or 'alignment' problem continuously for 20 years, but progress has been very slow compared to 'progress' on just building dangerous AI.
- TheOtherHobbes 3y agoLLMs have the potential to be one of the most toxic technologies in human history. Sexist etc content is a footnote compared to the potential for an automated, proven, and tuneable technology of mass persuasion and impersonation. So - massively addictive adtech. Reliable and effective political and ideological propaganda. Bot farms and "credible" fake media content produced at scale. Imagine those being used against a relatively defenceless population to creative division, undermine personal and collective confidence, and encourage stochastic terrorism and other kinds of violence.