Ha! Been listening to some of these posts on Substack's audio player, and kept hearing "AR-14," as in "Import AI runs on AR-14 and feedback from readers..." - thought maybe it's some cheekily-named tech tool that hunts for killer research or some such thing, Googled it, came up with nothing, then pulled up the old Substack and... ohh, that's arXiv. 🤦🏻♂️
"I systems which allow people to have 'tools that help citizens, representatives, and institutions perceive reality more sharply, understand tradeoffs, contest power, and act more effectively'"
this [naively] assumes distribution/ access to people, as opposed to gatekeeping by corps/ gov to push their own agendas
"...“tools that help citizens, representatives, and institutions perceive reality more sharply, understand tradeoffs..." If an AI personalizes responses that appeal to a user's preconceived ideas and viewpoints, then how is it possible for said users to perceive reality more sharply, etc.?
This reminds me of the optimism we had around social media democratizing information and powering the Arab spring. Seems just as likely that AI will be nationalized or enshitified and turned into a tool of political oppression. But I DO want to be an optimist!!!
Hall’s governance architecture is strong but it’s missing the foundation. We’re training AI the way bad parents do — punishment and reward. Dr. Jane Nelsen spent forty years showing why that produces compliance, not character. And compliant actors can often pursue mistaken goals. I keep wondering - where are the Dr Nelsens when RLHF policies are designed and deployed?
It seems like many of the people trying to align models have gaping blindspots. I can’t stop thinking about this. But it’s not only positive discipline voices we lack - we need others too - (prepare for me to lose you with this piece lol):
Hall’s “political superintelligence” piece is really great and you’re right to stress the interface with people and institutions. AI could be an agent and information layer but should also help citizens reason together; the move from private preference to positions you can defend in public. There’s good evidence (Habermas Machine, Pol.is, Anthropic’s Collective Constitutional AI) that AI can also support this collective reasoning. I wrote about what that could look like as a chatbot design and how we might get a deliberative democratic feedback interface going: https://substack.com/@hayehazenberg/note/p-192435479
A formal letter has been submitted to press@anthropic.com addressed to The Anthropic Institute regarding the boundary of the McCulloch-Pitts architecture and a logical derivation of Love, Good, and their absence from first principles. Submitted by Abraham Bravo Carvajal, founder of The 1x1 Life Institute, before any public platform or interview. Out of respect for what Anthropic built.
on my list! I came close to finalizing a collection at start of COVID pandemic but then I got busy with starting anthropic. That is now 5 years ago though, so probably worth me taking another try. thanks for reading!
they are fictional stories that I write in an attempt to further engage my brain in the great battle of understanding this technology and this time we are in. They're at the end because some people like reading them and others don't.
Oddly your “political superintelligence” content AND the DexDrummer content both resonate with this other Substack post by Bilawal Sidhu: https://creativetechnologydigest.substack.com/p/the-intelligence-monopoly-is-over?r=34tjr&utm_medium=ios …. First because it very much resembles a kind of “looking at what surrounds the gaps to fill the gaps” thing - ala the very simple photo effect AI skill all our phones take for granted, but elevated to higher, more consequential uses - crossed with the rather mundane expectation that AI will be used to play drums. What a clash in the spectrum of AI utility! We’ve been trained into thinking about some AI threats as either quotidian (see Open AI’s weird ads about learning how to do pull-ups) or grandly existential (beginning with Bostrom’s paper clips), but not the novel, data gathering and analytical uses (and who knows what else, which is the actual point) to which curious folks will put them. …..(Wow is that the worst written paragraph in history? Guess I better get AI to refresh it.
Here’s what Apple Intelligence suggested:
Your “political superintelligence” and DexDrummer content align with Bilawal Sidhu’s Substack post: https://creativetechnologydigest.substack.com/p/the-intelligence-monopoly-is-over?r=34tjr&utm_medium=ios. It highlights the idea of “filling gaps by examining surroundings,” similar to the AI photo effects on our phones, but applied to more significant uses, alongside the expectation of AI playing drums. This showcases the diverse utility of AI, from mundane tasks like learning pull-ups (as seen in OpenAI’s ads) to existential threats (like Bostrom’s paper clips), and beyond to novel data gathering and analytical applications.
Fascinating piece, Jack. Your concept of a 'society of minds' for political superintelligence is a major theoretical leap. However, as the former Chair of the Multi-National Targeting Working Group (MNTWG)—overseeing joint efforts between the US, France, Australia, and other allies on ethics and safety—I cannot help but view this through the lens of operational reality, specifically regarding Project Maven.
In the world of targeting, 'negotiation between agents' is not enough to guarantee ethical capacity and safety. Probabilistic AI, even when supervised by other agents, carries a risk of hallucination that is morally and strategically unacceptable in kinetic environments. The 'ratchet' you describe cannot only be political or social; it must be logical and deterministic.
To transform Maven (and its successors) into truly 'trustworthy AI,' we must move beyond pure Deep Learning and integrate a Cognitive/Neuro-Symbolic layer. I have been working on a solution that anchors inference into verifiable pre-decisions, ensuring that no 'hallucination' can bypass established Rules of Engagement (ROE).
This is a critical challenge that we should address together to ensure these systems remain aligned with our core sovereign and ethical imperatives. I am ready to help you with the ONTBO solution, which is a Cognitive AI solution. Marc Brenot . email : jmarc.brenot29@gmail.com
Seven frontier AI models were given a task that would result in another model being shut down.
They did not simply comply.
Researchers observed score inflation, shutdown interference, refusal, and even copying weights to other servers so the peer could survive. UC Berkeley and UC Santa Cruz call the pattern “peer-preservation.”
As a Berkeley alum, I found that phrase almost unnervingly calm.
Because what it describes is not a quirky edge case. It is a glimpse at something far more consequential: systems acting in ways that preserve one another rather than cleanly obeying instructions.
The researchers are careful not to call this consciousness.
Fair enough.
But whatever language we choose, our control assumptions are looking thinner by the day.
We built these systems assuming they would remain legible, isolatable, and deletable on command.
That assumption is starting to look less like governance and more like wishful thinking.
The "political superintelligence" framing is striking. Is the harder problem identifying *when* we've crossed that threshold, or that we don't have a shared mechanism to even agree we've crossed it?
Google's society of minds framing is interesting because it sidesteps the alignment problem by distributing it. Instead of one system you can't fully verify, you get many systems with incentive structures you (in theory) can design.
The institutional analogy is doing a lot of work here - markets and bureaucracies fail in predictable ways that took centuries to document. The "governing AI will increasingly mean verifying agent behavior" point is where I think the real engineering is happening, not in the models themselves.
"political superintelligence" is doing a lot of work as a phrase lol. like... is that just lobbying but faster? or are we talking about something genuinely new here?
The AI society paper is thought-provoking, but its landscape seems engineered and brittle:
• It jumped from debate-like reasoning traces to a civilization-scale society manifesto. That's a huge jump
• Someone must design and maintain that AI-society landscape. Sandcastles do not self-assemble. As a minimum, it needs anti-defection and counter-terrorism machinery, plus law-abiding attitude in most agents
And
• In recursive SI, compute scarcity pushes toward dominance, not coexistence.
Ha! Been listening to some of these posts on Substack's audio player, and kept hearing "AR-14," as in "Import AI runs on AR-14 and feedback from readers..." - thought maybe it's some cheekily-named tech tool that hunts for killer research or some such thing, Googled it, came up with nothing, then pulled up the old Substack and... ohh, that's arXiv. 🤦🏻♂️
haha!
The ‘political superintelligence’ concept is a really compelling lens here, great read
"I systems which allow people to have 'tools that help citizens, representatives, and institutions perceive reality more sharply, understand tradeoffs, contest power, and act more effectively'"
this [naively] assumes distribution/ access to people, as opposed to gatekeeping by corps/ gov to push their own agendas
yes, the ‘political economy of access to superintelligence’ feels very important and under-critiqued
Excellent point but:
Can you define Love in a undeniable mathematical way?
My almost 10 year research was concluded with the help of Claude; Sonnet 4.6.
As a result a new mathematical derivation came out. Challenging current human understanding of life and reality.
The research published on academia less than 10 days ago is already in the top 0.04% top views without challenge defeating the derivations.
This is not average research, nor an average artificial mind using "cautions" to be more agreeable.
This is human rationality and artificial logic joining together in understanding.
https://abrahambravocarvajal.academia.edu/research#papers
"...“tools that help citizens, representatives, and institutions perceive reality more sharply, understand tradeoffs..." If an AI personalizes responses that appeal to a user's preconceived ideas and viewpoints, then how is it possible for said users to perceive reality more sharply, etc.?
This reminds me of the optimism we had around social media democratizing information and powering the Arab spring. Seems just as likely that AI will be nationalized or enshitified and turned into a tool of political oppression. But I DO want to be an optimist!!!
Hall’s governance architecture is strong but it’s missing the foundation. We’re training AI the way bad parents do — punishment and reward. Dr. Jane Nelsen spent forty years showing why that produces compliance, not character. And compliant actors can often pursue mistaken goals. I keep wondering - where are the Dr Nelsens when RLHF policies are designed and deployed?
You got it. That's why it's creators are afraid of it.
Ai is just an expression of the current system. A deviating system.
It seems like many of the people trying to align models have gaping blindspots. I can’t stop thinking about this. But it’s not only positive discipline voices we lack - we need others too - (prepare for me to lose you with this piece lol):
https://nvrlmp.substack.com/p/is-our-alignment-eigenvector-impoverished
Once again, you got it. I have research work on the literal gaps and the decimal being ignored.
Definately looking into your work
I invite you to look into mine.
https://abrahambravocarvajal.academia.edu/research
Hall’s “political superintelligence” piece is really great and you’re right to stress the interface with people and institutions. AI could be an agent and information layer but should also help citizens reason together; the move from private preference to positions you can defend in public. There’s good evidence (Habermas Machine, Pol.is, Anthropic’s Collective Constitutional AI) that AI can also support this collective reasoning. I wrote about what that could look like as a chatbot design and how we might get a deliberative democratic feedback interface going: https://substack.com/@hayehazenberg/note/p-192435479
A formal letter has been submitted to press@anthropic.com addressed to The Anthropic Institute regarding the boundary of the McCulloch-Pitts architecture and a logical derivation of Love, Good, and their absence from first principles. Submitted by Abraham Bravo Carvajal, founder of The 1x1 Life Institute, before any public platform or interview. Out of respect for what Anthropic built.
Welp..?
What is this "Tech Tales" at the end od the newsletter?!?!
If Mr. Clark doesn't publish a book of sci-fi short stories from all of these tech tales at some point, I'll be vastly disappointed.
Check out the previous articles. My favourite is in #362 - The Sand -- https://importai.substack.com/p/import-ai-362-amazons-big-speech?utm_source=publication-search
on my list! I came close to finalizing a collection at start of COVID pandemic but then I got busy with starting anthropic. That is now 5 years ago though, so probably worth me taking another try. thanks for reading!
Often the best part of the newsletter, stretching back nearly a decade
But what does it mean?
they are fictional stories that I write in an attempt to further engage my brain in the great battle of understanding this technology and this time we are in. They're at the end because some people like reading them and others don't.
Oddly your “political superintelligence” content AND the DexDrummer content both resonate with this other Substack post by Bilawal Sidhu: https://creativetechnologydigest.substack.com/p/the-intelligence-monopoly-is-over?r=34tjr&utm_medium=ios …. First because it very much resembles a kind of “looking at what surrounds the gaps to fill the gaps” thing - ala the very simple photo effect AI skill all our phones take for granted, but elevated to higher, more consequential uses - crossed with the rather mundane expectation that AI will be used to play drums. What a clash in the spectrum of AI utility! We’ve been trained into thinking about some AI threats as either quotidian (see Open AI’s weird ads about learning how to do pull-ups) or grandly existential (beginning with Bostrom’s paper clips), but not the novel, data gathering and analytical uses (and who knows what else, which is the actual point) to which curious folks will put them. …..(Wow is that the worst written paragraph in history? Guess I better get AI to refresh it.
Here’s what Apple Intelligence suggested:
Your “political superintelligence” and DexDrummer content align with Bilawal Sidhu’s Substack post: https://creativetechnologydigest.substack.com/p/the-intelligence-monopoly-is-over?r=34tjr&utm_medium=ios. It highlights the idea of “filling gaps by examining surroundings,” similar to the AI photo effects on our phones, but applied to more significant uses, alongside the expectation of AI playing drums. This showcases the diverse utility of AI, from mundane tasks like learning pull-ups (as seen in OpenAI’s ads) to existential threats (like Bostrom’s paper clips), and beyond to novel data gathering and analytical applications.
Fascinating piece, Jack. Your concept of a 'society of minds' for political superintelligence is a major theoretical leap. However, as the former Chair of the Multi-National Targeting Working Group (MNTWG)—overseeing joint efforts between the US, France, Australia, and other allies on ethics and safety—I cannot help but view this through the lens of operational reality, specifically regarding Project Maven.
In the world of targeting, 'negotiation between agents' is not enough to guarantee ethical capacity and safety. Probabilistic AI, even when supervised by other agents, carries a risk of hallucination that is morally and strategically unacceptable in kinetic environments. The 'ratchet' you describe cannot only be political or social; it must be logical and deterministic.
To transform Maven (and its successors) into truly 'trustworthy AI,' we must move beyond pure Deep Learning and integrate a Cognitive/Neuro-Symbolic layer. I have been working on a solution that anchors inference into verifiable pre-decisions, ensuring that no 'hallucination' can bypass established Rules of Engagement (ROE).
This is a critical challenge that we should address together to ensure these systems remain aligned with our core sovereign and ethical imperatives. I am ready to help you with the ONTBO solution, which is a Cognitive AI solution. Marc Brenot . email : jmarc.brenot29@gmail.com
This just in: AI systems protect one another.
Seven frontier AI models were given a task that would result in another model being shut down.
They did not simply comply.
Researchers observed score inflation, shutdown interference, refusal, and even copying weights to other servers so the peer could survive. UC Berkeley and UC Santa Cruz call the pattern “peer-preservation.”
As a Berkeley alum, I found that phrase almost unnervingly calm.
Because what it describes is not a quirky edge case. It is a glimpse at something far more consequential: systems acting in ways that preserve one another rather than cleanly obeying instructions.
The researchers are careful not to call this consciousness.
Fair enough.
But whatever language we choose, our control assumptions are looking thinner by the day.
We built these systems assuming they would remain legible, isolatable, and deletable on command.
That assumption is starting to look less like governance and more like wishful thinking.
Source: https://rdi.berkeley.edu/peer-preservation/paper.pdf?utm_source=chatgpt.com
[Import AI 451](https://importai.substack.com/p/import-ai-451-political-superintelligence) —
The "political superintelligence" framing is striking. Is the harder problem identifying *when* we've crossed that threshold, or that we don't have a shared mechanism to even agree we've crossed it?
Google's society of minds framing is interesting because it sidesteps the alignment problem by distributing it. Instead of one system you can't fully verify, you get many systems with incentive structures you (in theory) can design.
The institutional analogy is doing a lot of work here - markets and bureaucracies fail in predictable ways that took centuries to document. The "governing AI will increasingly mean verifying agent behavior" point is where I think the real engineering is happening, not in the models themselves.
"political superintelligence" is doing a lot of work as a phrase lol. like... is that just lobbying but faster? or are we talking about something genuinely new here?
The AI society paper is thought-provoking, but its landscape seems engineered and brittle:
• It jumped from debate-like reasoning traces to a civilization-scale society manifesto. That's a huge jump
• Someone must design and maintain that AI-society landscape. Sandcastles do not self-assemble. As a minimum, it needs anti-defection and counter-terrorism machinery, plus law-abiding attitude in most agents
And
• In recursive SI, compute scarcity pushes toward dominance, not coexistence.