Converging Futures: A New Technique for Arriving at a Shared Vision

General
62

Yep, this is what I’m thinking about - what kind of experiments?

63

For example, to test the proposition “philosophy eats AI”, I took some ideas from this thread and put them into Copilot. Here was the prompt:

”Compare Confucius on the rectification of names with Wittgenstein’s notion of a language game.”

Then I drew my own conclusion from the results, put it back into the this thread, and checked out the reaction it got. All of my major ideas are getting bounced around like that in multiple Internet fora, against AI, in classes, and in other conversations. If I’m really wrong about anything, it’s sure to come out in the wash! Beyond just straight up “right/wrong”, I’m also keenly interested on the rhetorical praxis level of “works/does not work”. “Works/does not work” varies with audiences and interlocutors, of course. Also, “works/does not work” helps to define which channels I should be using (because they work for me) versus which channels are best left in other hands.

BTW, on the Copilot session above, my main takeaway was process, process, process. Confucious also was working a process, although the outcome of that process may sound more like a dogma for the ages. But the process of how that dogma got formulated in the first place is what interests me.

64

How will we save the world or improve our prospects by testing propositions like “philosophy eats AI”?

65

It’s flattering to imagine that “we” might save the world! Depends, I suppose, on how inclusive that “we” is. (Mine is pretty inclusive - it extends to the full tissue of cosmic manifestation + the divine substrate. That “we” can certainly save whatever it wants to save …)

As for improving our prospects, that’s a bit more prosaic. The acid test in my lifeworld is can a “philosophy eats AI” approach work better than AI alone, sans philosophy? My students seem most attracted to AI with alignment - which is a fancy way of saying ethical AI - which is one application of philosophy to AI governance. My students really don’t have much use, in general, for cutthroat tech bro, devil-take-the-hindmost, AI startup competitions. They would like to earn decent livings, however. It strikes me (and not just me - there is a lot of YouTube opining about all this) that well-rounded humans are going to do better than less well rounded humans in the emerging AI-driven economy. We’ll have to look back in a generation or so to find out if that is how things actually do turn out, however.

1 Like
66

I was hinting at what grounds philosophy. Embodied practice is maybe where we need to be?

67

I’d say so. A couple years ago I read Habermas’s Also a History of Philosophy (vol 1) with Brendan Graham Dempsey and a few others. To set the stage for the question, “why did philosophy not become a religion?”, Habermas delved into the origins of language and culture in general. Long story short, he focused attention on ritual. Another book I like for exploring that same topic is Huizinga’s Homo Ludens: a study of the play-element in culture. To roughly summarize all that, there is a reason even the most primitive tribal groups favor song, dance, masks, play acting, narrative tales, etc. All that is foundational for human culture. More abstract linguistic formulations - like philosophy - need to tap into all that or they lose their tether.

In the case of ancient Greek philosophy, the Platos of the world kept their distance from civic ritual. That created a void, that Christianity later moved in to fill. In the Chinese and Indian counter-examples, abstract thinking stayed in better contact with the respective social orders. The Hebrew tradition was mostly about setting up a very particular social order, with a degree of abstraction emerging by the by. Christianity and Islam both followed in that vein.

So whatever comes next - be it a new religion, an amalgamation, or a renewal of one or more existing religions - unless thought can find it’s footing in feeling, we won’t be able to do much anything that AI won’t be able to do better. “Embodied action” is a good generic phrase for what any viable social formation needs to focus on.

1 Like
68

Now that we got through all that contextual framing involving language games, process linguistics, and grounding, I took a good hard look at the protocol itself. As usual, I’m on the hunt for applied cases - especially cases students can sink their teeth into.

The case that intrigues me is “Job Seeker”. Job Seeker has the intention to earn money. Job Seeker makes promises to the market regarding intended work performance offered in exchange for money. Job Seeker offers various types of evidence (resume, transcript, cover letter, recommendations, portfolio product, etc.) to support the likely veracity of performance promises.

Questions - can Promise Protocol reduce friction in the job search process? Will it get better results than current methods? How can this be tested?

69

This is an interesting use case that clearly demonstrates the ‘unfair advantages’ a job seekers’ market built using Promise Protocol would have over a conventional marketplace.

In a jobs marketplace, there are seekers and offerers, both of whom have to signal their intentions clearly and honestly in order for the market to be healthy. Seekers could make promises like the ones you listed, as well as specific promises about their capabilities. This could look very much like LinkedIn endorsements, but the main difference would be skin in the game; claiming capabilities one doesn’t have could lead to assessments by others—who could remain anonymous, but necessarily with their own skin in the game—leading to forfeited stake. Making accurate claims about one’s own abilities or assessing those of others would be rewarded.

In short, the ability of jobseekers to signal their credibility effectively would be one of the main benefits. A worker with real merit could display this with a merit profile much more effectively than through a resume, which can be faked, or must be increasingly tailored to AI consumption just to be noticed. You can think of this in terms of deadweight loss in microeconomics: there are transactions that should happen but don’t because participants have no effective means of signaling their credibility. Also, transactions that shouldn’t happen do, because the appearance of credibility is easily faked.

Most of my evidence is rational as opposed to empirical because the system is rather complex and I haven’t been able to pilot it. I have agent based models showing it works quite well in theory.

One aspect of this system that I think could be particularly impactful is coalition resistance. If a group of agents decides to try to game the system by faking assessments, they have more to lose than to gain.

I do believe it will get better results than current methods, but those agent based models are the closest I have to empirical evidence (if you can call it that). I also have purely theoretical evidence in the form of Coq proofs demonstrating that various properties hold in the system, mentioned in the yellowpaper. If anything, I probably went overboard with the pre-empirical validation.

70

To mashup a few considerations -

  • there is an emerging view that AI is going to push more workers into the gig economy (like it or not). The current job search process (AI-screened resumes) is pretty broken by all accounts. If the proposed system proves efficient at matching gig workers to opportunities, it could start out in the gig world and then migrate eventually to the salaried HR world.

  • I see a family resemblance here to the Superforecaster model. Superforecaster - Wikipedia The gist of superforecasters is they aim for precision, keep score, and work on process improvements to get better. Placing a candidate into a job based on performance expectation is a species of superforecasting applied to a particular vertical.

  • This could all spawn a new HR model in which HR is mostly about forecasting, and there are tracking metrics and feedback loops to incentivize getting better at it.

71

Yes, this seems inevitable. This is another force we can attribute to the pull side of the equation (returning to Marshall Poe’s Innis-inspired media adoption framework) that would reduce the adoption friction for a proposed solution. It’s easier to convince people that something like Promise Protocol is necessary when entire industries come into being just to deal with the inefficiencies engendered by unfettered AI use.

72

This came to my attention today. Seems pretty relevant.

73

I don’t have time to listen to this or read the paper for now, but I did ask GPT-5 to evaluate it and list some of the challenges of this approach. It listed the following (TL;DR - I think Promise Protocol could really help!):Here’s the short list of “hard parts” and what’s still unsolved, mapped to the stack the paper proposes.

#1 Modeling values (TMVs) so they’re usable by machines

  • Disentangling values from preferences/behavioral noise. PMV bundles addictions, social pressure, and fads with genuine values; TMVs claim to separate them, but a generally accepted operational test is still missing. The paper names the desideratum (robustness to distortion) but not a settled procedure.

    full stack alignment

  • Picking a concrete formalism. Options include “values as attentional policies,” moral graphs, and contractualist structures; the paper surveys them but leaves open which formal core is best for broad deployment and how to compose them.

  • Collective values, not just individual ones. We lack scalable ways to elicit, aggregate, and update shared norms across diverse populations without collapse to slogans or preference polls. The paper frames this as a desideratum rather than a solved pipeline.

    full stack alignment

  • Generalization to novel contexts. TMVs promise better transfer than PMV/VAT, but we lack benchmarks showing reliable cross-domain application under shift.

    full stack alignment

#2 Evaluation & benchmarks (what counts as “aligned”?)

  • “Moral help” vs. manipulation. How to tell when an assistant’s guidance reflects genuine, user-endorsable values rather than subtle steering remains an open evaluation problem the paper calls out directly.

    full stack alignment

  • Value reliability under distribution shift. Do model sizes/training schemes improve stability of value-consistent behavior across contexts? Unknown; listed as an open question.

    full stack alignment

  • Auditable justifications. We need standard, checkable artifacts that trace decisions to explicit values/norms—especially for democratic/market settings operating at “AI speed.” The need is noted; concrete audit formats are not yet specified.

    full stack alignment

#3 Learning normative competence (beyond rules)

  • From text policies to structured norms. Translating natural-language standards into machine-actionable structures (and proving compliance) is open research.

    full stack alignment

  • Training methods. Early ideas include norm-augmented Markov games and resource-rational contractualism, but we don’t know which training/fine-tuning regimes yield robust norm following, adaptation, and reasoning.

#4 Integrity, honesty, and incentives among agents

  • Value-based commitments. We lack protocols that let agents credibly reveal/commit to their values with guarantees against bluffing and incentives to cooperate; integrity-checking and anti-manipulation remain open.

    full stack alignment

#5 Meaning-preserving economic mechanisms

  • Measuring “flourishing” without Goodharting. The paper points to outcome-based, values-guided contracting and “meaning-promoting intermediaries,” but concrete, widely accepted metrics (and anti-gaming defenses) aren’t specified.

  • Pay-for-meaning interfaces. How suppliers get paid for contributions to value portfolios (with auditability and low measurement burden) is aspirational, not turnkey.

    full stack alignment

#6 Democratic regulation at AI speed

  • Scaling value aggregation. Can processes like MGE/moral graphs represent large, diverse populations without capture or bias? What guarantees/optimality properties can be proven? Open questions, per the paper.

    full stack alignment

  • Legitimacy under time pressure. How fast-acting systems maintain democratic legitimacy when real-time human deliberation is impossible remains unsolved.

    full stack alignment

#7 “Full-stack” plumbing and interop

  • Preserving value information up the stack. Today’s stack loses signal (person → product → company → market → regulator). The paper shows the problem but not the full data/standards layer (schemas, provenance, custody) to stop the loss.

    full stack alignment

  • Cross-institution handoffs. We need formats so an assistant’s value artifact can be consumed by corporate governance and by regulators without re-elicitation. Still to be designed; flagged as a program, not a deliverable.

    full stack alignment

#8 Adoption & fieldwork

  • From theory to practice. Authors explicitly call for cross-disciplinary, real-world experiments; the approach is a research program, not a deployed standard. Institutional buy-in, incentives, and governance models are future work.

What would move the needle

  • Reference TMV spec + SDK. Pick a first-class representation (e.g., attentional-policy core + moral-graph layer), with tools for elicitation, versioning, diffing, and validation. Pilot it in one assistant and one institutional workflow.

  • Benchmark suite. Tasks that test “moral help vs. manipulation,” norm generalization, and audit quality under distribution shift; publish leaderboards.

  • Two exemplar mechanisms. (1) A values-guided dynamic-contracting pilot with measurable flourishing outcomes; (2) a “democracy-at-speed” prototype that produces auditable, value-grounded recommendations with stated legitimacy guarantees.

Bottom line: the vision is compelling and well-scoped as a research program. The main gaps are (a) a canonical TMV formalism with tooling, (b) rigorous evaluations that distinguish help from manipulation and show cross-context reliability, and (c) end-to-end demos where value information is preserved and audited across multiple institutional layers.

74

What I like about the TMV approach is they are waking up the need for a thick description of values. What this runs into however is the whole Confucius/Wittgenstein line of reasoning we discussed earlier in the thread - any ‘name” (attribute, data point, descriptor, etc.) may need to be “rectified” (corrected, verified, grounded), because any digital artifact no matter how elegant can be spoofed (deployed by threat agents to play a language game of their own). What this points to for me is TMV needs to be grounded in processes that recurse back to the original human (at least from time to time) to refresh and validate the represented values. Or, in AI lingo - “human in the loop”.

Or maybe even multiple humans in the loop. That would be MFA for values attestation. (If fits your post elsewhere about the desirability of collective leadership models). Consider the authentication procedure needed to launch a self-destruct sequence on the Star Ship Enterprise - multiple crew members in the loop, each with their own pass code. The Enterprise computer is basically an AI agent. That computer is so capable is so capable, why do they even need a crew? Why not just send AI on the fiye-year mission? One might surmise, because there are advantages to humans in the loop.

75

I’m on a Zoom call right now with a bunch of senior AI engineers. (I’m the token professor). This seems like a standard reference everyone is working with as a sort of baseline. Any relationship to Agency Protocol?

76

You can think of MCP as a standardized way for AI agents to send and receive messages. Promise Protocol is like an extra layer that could be added to ensure accountability undergirds these communications. Imagine if humans had a general way to ‘put your money where your mouth is’ when discussing things. That’s analagous to what Promise Protocol could be to MCP.

Furthermore, Promise Protocol could help MCP evolve to be better. Suppose there’s room for improvement in the current version of MCP. Well, defining it as an agent in Promise Protocol means there’s an incentive and a meritocratic process in place for improving it. Actually, I already have defined it as an agent here:

A robust integration between Promise Protocol and MCP would enable applications like AI-to-AI services, where an AI agent would be in charge of a budget and be able to outsource jobs to other AI agents based on their provable track record of domain-specific promise-keeping.

77
78

OK, now we’re cooking with juice! Here is a sample of some of the proposed learning objectives from yesterday’s session:

Design multi-agent workflows
Implement agent loops
Establish agent communication protocols

Also …

Define ethical and policy standards to govern agent behavior and data usage.
Implement content moderation and bias mitigation filters to enforce safe outputs.
Apply role-based access control and context permissions to protect sensitive data.

There are approximately 70 such objectives in the desired curriculum in total. Somehow, they want these sorts of items in an undergraduate curriculum. My take is this is simply not attainable (at any sort of scale, anyway) through linear curriculum models and linear curriculum adoption processes. In the terminology of yesterday’s 2R research call, ethical guardrails for multi-agent workflows is a ‘mystery”, not a “problem”. Reason alone won’t take us there (at least not on the level of some sort of step-by-step conventional curriculum model for undergrads. With a text like The Complete Idiot’s Guide to Unleashing AI Super-Agents Without Triggering Planetary Apocalypse.)

My idea of a “2R Bootcamp” is an offshoot of what to me is the obvious necessity of a bootcamp approach to train young humans on all this AI stuff. Effectively, to complete with AI accelerationism (woefully misguided, IMO), I’m considering the potential for humanities accelerationism to match tempo with the ML world. It’s not just that we need ethical humans. We need humans who can wrap their heads around Kant, Habermas, etc. as young undergraduates, in a big hurry, get it right, and then apply those values to a choatic workforce/battlefield of agentic AI. I say “battlefield” because AI threats and threat agents are already crawling all over this space. Niall Ferguson is setting up a university for “Navy Seals of the intellect”. He’s right about the special ops metaphor. I’m also feeling very special ops at the moment. The required curriculum is clearly esoteric on many levels. It will take something like initiation into a wisdom tradition to allow ordinary students to penetrate this world, let alone guide it ethically.

79

I’m out of the loop, is this for one of your courses?

I’m biased, of course, but all of these objectives are more achievable with Promise Protocol :slight_smile:

80

This was a curriculum design committee I got called into. Most of the participants were senior AI engineers from firms in my area. I was the token educator.

The idea is to develop 2-year or 4-year curricula to make students attractive hires to these big AI platformers. The irony, of course, is these firms have been laying off and have effectively squeezed out the entry level. To me the idea of a linear 2-year or 4-year curriculum arranged by conventional learning objectives leading to a likely hire seems out of phase with how the world currently is. So I’m considering options …

My most direct approach will be to pitch to my current school administration to set up some sort of pilot program around non-linear, recursive educational methods to leverage experiential and cognitive feedback loops (including AI-enhanced cognitive loops) in order to get a “bootcamp” training effect. I basically want to take random community college students and turn them into high tech gate crashers by training them 10x faster than ordinary institutional methods allow.

If that does not work, I can set up an independent model somehow and run the program anyway. Or shop the idea around to some other school that is interested.

Another alternative would be for me to get involved with R&D projects (like your Promise Protocol), recruit potential talent through my 1000+ LinkedIn network, and run the training program as a subunit of the R&D effort. Institutional schools can catch up later, if ever.

81

Below is a bit of a Copilot session on MCP vs Agency Protocol.

Supposing Agency Protocol manages MCP, then the human-in-the-loop question becomes, how to humans configure Agency Protocol? Also, which humans do that configuring? And what systems of ethics is influencing those configuration choices?

+++++++

How They Relate

  • Complementary Roles:

    • MCP handles technical interoperability (how models and agents connect to tools and share context).

    • Agency Protocol handles ethical and operational integrity (how agents prove they acted as promised and maintain trust).

  • In practice:

    • An agent might use MCP to call a tool or another agent.

    • Agency Protocol ensures that this interaction is logged, auditable, and aligned with agreed policies or commitments.

  • Think of MCP as the plumbing for connectivity and Agency Protocol as the governance layer for trust and accountability.

Why It Matters

  • As multi-agent ecosystems grow, interoperability alone isn’t enough—trust and compliance become critical.

  • Without Agency Protocol, MCP-based systems risk being opaque and hard to audit.

  • Together, they enable scalable, secure, and trustworthy agent networks.