Your friend Mary sounds like an optimist! Human-as-amanuensis could be read as the bargaining phase of grieving the end of an anthropocentric intelligence paradigm.
Yes, many current applications of AI are still very "manual" and require a human in the loop. But as the era of agents is upon us, we're seeing more of this:
"In general, we find that AI agents perform better than humans at AI research engineering tasks when both are given 2 hours, but they perform worse at higher time budgets."
There's no shortage of loud voices still working through the denial and anger stages of grief. "Nonhuman entities will never be able to do [X]!" Perhaps some subset of them are correct. But the observed pattern so far is that progress continues to chip away at hubris.
Should AI research engineers hang their hopes and dreams on the improbable outcome that we never progress beyond the "2 hour" complexity barrier? Or might it be wiser to consider having a dialogue about what happens when/if we scale to, say, the "1 year" barrier?
I think you're channeling what was implicit in the article Mark 🙂 – to me the concept of human labor in service of AIs is pretty disturbing, and one that it's becoming easier to envision as this thought experiment shows.
Oh, indeed! I'm just pulling on the thread you so beautifully weaved into the fabric of your narrative. And I'm pulling on it because the point you make is one keeping me awake at night.
You may be interested in reading a study made by Italian Central Bank about “An assessment of occupational exposure to artificial intelligence in Italy”.
A first finding is :
Overall, the different methodologies agree on the fact that occupations requiring cognitive skills are more likely to be exposed to the introduction of AI, a marked difference compared to the introduction of robots, one of the most recent waves of innovation observed.
And moreover:
Observably, AI technologies have brought advancements in all those activities (perception, handling with dexterity, content generation, social interactions) that were previously considered inherently human and at low exposure. In fact, some of the occupations with low exposure to earlier automation technologies now appear in the medium or high exposure groups, albeit with different degrees of complementarity (e.g. mathematicians or human resource managers). Other occupations, mainly related to physical strength, were highly exposed to some of the previous technological waves, but appear only mildly exposed to AI.
Now I'm going to be thinking about this all day! Wondering how it would play out in an organization over time. Like, would the primary agent who presumably is responsible for several categories of work have a separate AI for each? Or one AI that directs the work of several humans (who as you show also use AIs to assist them)? Do those AIs interact with each other? What's the equivalent of a 1:1 meeting when the primary agent needs to touch base with the AI? What percent of time/energy do the human-to-human communications still require?
Thank you especially for the drawings. Very helpful to understand the progression of models.
Thanks Margie! I think there are probably many configurations that are possible -- although as Mark Daley notes in the comments, a good many of these are probably not desirable!
Really enjoyed reading this thought experiment, which especially resonated with us with regard to the creative skills which are uniquely human and relevant for innovation and progress at a societal level. We would love to explore with you what it means to be human in a human-AI symbiotic relationship...
Your friend Mary sounds like an optimist! Human-as-amanuensis could be read as the bargaining phase of grieving the end of an anthropocentric intelligence paradigm.
Yes, many current applications of AI are still very "manual" and require a human in the loop. But as the era of agents is upon us, we're seeing more of this:
https://metr.org/blog/2024-11-22-evaluating-r-d-capabilities-of-llms/
"In general, we find that AI agents perform better than humans at AI research engineering tasks when both are given 2 hours, but they perform worse at higher time budgets."
There's no shortage of loud voices still working through the denial and anger stages of grief. "Nonhuman entities will never be able to do [X]!" Perhaps some subset of them are correct. But the observed pattern so far is that progress continues to chip away at hubris.
Should AI research engineers hang their hopes and dreams on the improbable outcome that we never progress beyond the "2 hour" complexity barrier? Or might it be wiser to consider having a dialogue about what happens when/if we scale to, say, the "1 year" barrier?
I think you're channeling what was implicit in the article Mark 🙂 – to me the concept of human labor in service of AIs is pretty disturbing, and one that it's becoming easier to envision as this thought experiment shows.
Oh, indeed! I'm just pulling on the thread you so beautifully weaved into the fabric of your narrative. And I'm pulling on it because the point you make is one keeping me awake at night.
You may be interested in reading a study made by Italian Central Bank about “An assessment of occupational exposure to artificial intelligence in Italy”.
A first finding is :
Overall, the different methodologies agree on the fact that occupations requiring cognitive skills are more likely to be exposed to the introduction of AI, a marked difference compared to the introduction of robots, one of the most recent waves of innovation observed.
And moreover:
Observably, AI technologies have brought advancements in all those activities (perception, handling with dexterity, content generation, social interactions) that were previously considered inherently human and at low exposure. In fact, some of the occupations with low exposure to earlier automation technologies now appear in the medium or high exposure groups, albeit with different degrees of complementarity (e.g. mathematicians or human resource managers). Other occupations, mainly related to physical strength, were highly exposed to some of the previous technological waves, but appear only mildly exposed to AI.
You find the full paper in English at this link:
https://www.bancaditalia.it/pubblicazioni/qef/2024-0878/index.html?com.dotmarketing.htmlpage.language=1
Thanks Franco!
Now I'm going to be thinking about this all day! Wondering how it would play out in an organization over time. Like, would the primary agent who presumably is responsible for several categories of work have a separate AI for each? Or one AI that directs the work of several humans (who as you show also use AIs to assist them)? Do those AIs interact with each other? What's the equivalent of a 1:1 meeting when the primary agent needs to touch base with the AI? What percent of time/energy do the human-to-human communications still require?
Thank you especially for the drawings. Very helpful to understand the progression of models.
Thanks Margie! I think there are probably many configurations that are possible -- although as Mark Daley notes in the comments, a good many of these are probably not desirable!
Really enjoyed reading this thought experiment, which especially resonated with us with regard to the creative skills which are uniquely human and relevant for innovation and progress at a societal level. We would love to explore with you what it means to be human in a human-AI symbiotic relationship...