The alignment problem with AGI - on writing

?“We learn how to write when we react vigorously to what we have written …. I write what I dream about and what I speak about and what causes me to struggle” Brazilian educator and philosopher Paulo Freire, Letters to Christina

Freire famously stated that before he learned to write and think with words he wrote and thought with banana leaves in impoverished conditions Northeast of Brazil.

Writing is a higher-order form of thinking. Today such thinking can be framed in terms of what is known as the “alignment problem” with AGI.

The phrase alignment problem refers to AI systems adhering to human intention and design. Misalignment occurs when these systems take actions outside human intentionality. ?

In their recent release of ChatGpt 4, OpenAI states “there is currently no known indefinitely scalable solution to the alignment problem.” What this means is that human-machine alignment is very difficult to ensure even at this early stage of large language model (LLM) development.

My concern is not with such shifts in thinking due to near-human machine expression. My concern is that as Freire describes, that acts of writing with dream and struggle get disconnected, that embodied dimensions between writing, passion and thinking get further attenuated, disembodied.

Writing has never been a sole text-based activity. For example direct writing of nanoparticles is a form of writing.[1] Nobel Physicist Giorgio Parisi discovered hidden patterns inscribed in complex systems on planetary and astro scales as a form of writing. ?As humans the act of writing is an embodied dimension of larger writing universes.

I don't think ChatGpt alone is the monster threatening such forms of writing and thinking. This disconnect already happens every day in distressed schools, in board rooms that can lack vigorous approach to writing and communication.?We often already have misalignment, human-institutional misalignment. ?The machines we should be aware of are often not without but within.

Such misalignment that promotes suspension of critical thinking can go even further. It has insidious longer-term effects. It can lead to confusion, misinformation, division, and cruelty. ?It can lead to a kind of zombie reaction to what we have written a zombie reaction to language, and to each other.?

I will leave you with an excerpt on how the AGI alignment problem is being approached from OpenAI

“We believe that finding an indefinitely scalable solution [to the alignment problem] is likely very difficult.” …. “As we make progress on this, our AI systems can take over more and more of our alignment work and ultimately conceive, implement, study, and develop better alignment techniques than we have now. They will work together with humans to ensure that their own successors are more aligned with humans.”[2]

It seems to me that as better alignment techniques are developed there is no better (or urgent) time to write then now with the astounding possibilities all around us being generated from all sources and tools at hand - banana leaves, nano-scale particles, LLMs, texts, astro inscriptions, and poetry. As Salman Rushdie once wrote, to write means to "shape the world, and stop it from going to sleep.”

#writing #nanoparticles #physics #AGI #alignmentmatters #ai

?

?

?

?

?

?

?

[1] https://link.springer.com/article/10.1007/s11051-012-1220-y

[2] https://openai.com/blog/our-approach-to-alignment-research