The load-bearing claim is that bottlenecks, not capabilities, govern the timeline. That is testable only if bottlenecks get the same measurement discipline as benchmarks. The CRUX result, expert reviewers unambiguously rejecting agent-written papers, is a useful start, but "unambiguously rejected" is a coarse instrument; it cannot separate a capability gap from a convention or framing gap.
What would a graded rubric with inter-rater agreement show? For the top-right quadrant to earn its place, bottleneck claims need base rates and falsifiers, not just a better frame.
The line about people getting tax advice from AI outside work is the one I'd most want you to dig into, because the surplus is real but so are the mistakes and nobody is counting either. How would you even measure that when it never shows up in employment data?
When you repeat CRUX, will you also run older models with the new scaffolds? I'd be interested in how much of any improvement comes from the setup around the model.
Would be very interested to collaborate on questions of economic impacts and global inequality from a development context. Seems like it might nest nicely inside a piece of this.
Congrats! I am incubating an aligned project w/r/t counseling civil society organizations to keep pace so that the social change sector is not left behind. Would love to connect sometime. Laura MacCleery, Continuous Garden
I support your project. This is a huge issue. We are about to go public aimed at solving this problem. It requires an adaptive model for continuous improvement. It requires a new operating model where talent and AI resources can be allocated, measured, and optimized. Both the organizations, talent, and AI adapt together. The current job structure and headcount accounting must give way to measure the work performed, who benefits, and its value for talent and AI
Congratulations. This project promises to provide much needed focus on the importance of community/institutional aspects of AI – as we shape out tools and they shape us, as social creatures.
I have a piece on that pending at Tech Policy Press, which proposes key framings and architectural principles related to your helpful quadrant framing. It comments on a related Pro-Human AI Coalition (of some 60+ organizations) that you might want to work with (if not already). Preview extracts at https://ucm.teleshuttle.com/2026/09/in-support-of-pro-human-ai-coalition.html.
Aside from the role of individual agency that many seek, I see critical need for the “social mediation ecosystem” of both formal and informal communities and institutions to maintain the effectiveness of human discourse, as essential to democracy and human flourishing. That is part of my “Three Pillars” framing, first developed for social media, but now essential to pro-human AI as well: https://www.techpolicy.press/three-pillars-of-human-discourse-and-how-social-media-middleware-can-support-all-three/
I distill the most important – and still very neglected) test of AI agents and other components as asking “Who does it serve?” It should serve the individuals and communities (and businesses) that use them as loyal and trustworthy agents, not the corporations that build the tools. (My work too, is motivated by Engelbart and Licklider – the idea of augmentation, not automation.)
They are not bottlenecked INSIDE the AI companies themselves, where AI is leading more and more R&D projects and where the technology is not at all impacted by slow diffusion or legal requirements hampering rapid adoption in the real world.
It’s one thing to say that radiologists won't be fully replaced because, for example, malpractice insurance requires that a human being be the one to make the final call on the diagnosis. It's another to say that these bottlenecks to adoption affect the bottlenecks to social transformation.
At the most basic level, while cyberhacking and synthesized superpathogens require technical expertise and the availability of connections and/or capital, neither fall into the same category as the changes bottlenecked by slow economic diffusion. More to the point, the automation of AI R&D itself does not require any outside institution or individual to “adapt” in any way to what's going on inside.
I expect society to be “transformed” in a negative way, rather quickly. This will come from the development of powerful AI and it's usage by entities that do not need to concern themselves at all with any of what's discussed in a lot of this post.
It's not the first time these points have been brought up, and it's also not the first time they have been ignored by the authors of the “AI as Normal Technology” perspective.
Congratulations on the launch, you said you were hiring, what is the best way to reach you?
The load-bearing claim is that bottlenecks, not capabilities, govern the timeline. That is testable only if bottlenecks get the same measurement discipline as benchmarks. The CRUX result, expert reviewers unambiguously rejecting agent-written papers, is a useful start, but "unambiguously rejected" is a coarse instrument; it cannot separate a capability gap from a convention or framing gap.
What would a graded rubric with inter-rater agreement show? For the top-right quadrant to earn its place, bottleneck claims need base rates and falsifiers, not just a better frame.
The line about people getting tax advice from AI outside work is the one I'd most want you to dig into, because the surplus is real but so are the mistakes and nobody is counting either. How would you even measure that when it never shows up in employment data?
When you repeat CRUX, will you also run older models with the new scaffolds? I'd be interested in how much of any improvement comes from the setup around the model.
Dear team, I suggest we set up a call as our model my provide insights and perhaps a short cut.
Would be very interested to collaborate on questions of economic impacts and global inequality from a development context. Seems like it might nest nicely inside a piece of this.
Congrats! I am incubating an aligned project w/r/t counseling civil society organizations to keep pace so that the social change sector is not left behind. Would love to connect sometime. Laura MacCleery, Continuous Garden
I support your project. This is a huge issue. We are about to go public aimed at solving this problem. It requires an adaptive model for continuous improvement. It requires a new operating model where talent and AI resources can be allocated, measured, and optimized. Both the organizations, talent, and AI adapt together. The current job structure and headcount accounting must give way to measure the work performed, who benefits, and its value for talent and AI
Congratulations on the launch! I am a fan of quadrant charts, and I appreciate your proposed approach to framing the "missing quadrant"!
FWIW, I just posted a link to this thread on LinkedIn, to a Group focused on AI topics:
https://lnkd.in/p/g-MVQwUG
Congratulations. This project promises to provide much needed focus on the importance of community/institutional aspects of AI – as we shape out tools and they shape us, as social creatures.
I have a piece on that pending at Tech Policy Press, which proposes key framings and architectural principles related to your helpful quadrant framing. It comments on a related Pro-Human AI Coalition (of some 60+ organizations) that you might want to work with (if not already). Preview extracts at https://ucm.teleshuttle.com/2026/09/in-support-of-pro-human-ai-coalition.html.
Aside from the role of individual agency that many seek, I see critical need for the “social mediation ecosystem” of both formal and informal communities and institutions to maintain the effectiveness of human discourse, as essential to democracy and human flourishing. That is part of my “Three Pillars” framing, first developed for social media, but now essential to pro-human AI as well: https://www.techpolicy.press/three-pillars-of-human-discourse-and-how-social-media-middleware-can-support-all-three/
I distill the most important – and still very neglected) test of AI agents and other components as asking “Who does it serve?” It should serve the individuals and communities (and businesses) that use them as loyal and trustworthy agents, not the corporations that build the tools. (My work too, is motivated by Engelbart and Licklider – the idea of augmentation, not automation.)
“Yet we recognize that the pace of societal transformation will be https://knightcolumbia.org/content/ai-as-normal-technology by the speed of human and organizational change more than progress in capabilities.”
They are not bottlenecked INSIDE the AI companies themselves, where AI is leading more and more R&D projects and where the technology is not at all impacted by slow diffusion or legal requirements hampering rapid adoption in the real world.
It’s one thing to say that radiologists won't be fully replaced because, for example, malpractice insurance requires that a human being be the one to make the final call on the diagnosis. It's another to say that these bottlenecks to adoption affect the bottlenecks to social transformation.
At the most basic level, while cyberhacking and synthesized superpathogens require technical expertise and the availability of connections and/or capital, neither fall into the same category as the changes bottlenecked by slow economic diffusion. More to the point, the automation of AI R&D itself does not require any outside institution or individual to “adapt” in any way to what's going on inside.
I expect society to be “transformed” in a negative way, rather quickly. This will come from the development of powerful AI and it's usage by entities that do not need to concern themselves at all with any of what's discussed in a lot of this post.
It's not the first time these points have been brought up, and it's also not the first time they have been ignored by the authors of the “AI as Normal Technology” perspective.