[EDRM Editor’s Note: EDRM is proud to publish Ralph Losey’s advocacy and analysis. Images in the article were created by Ralph Losey using AI unless otherwise noted. Originally published on EDRM.net. This article remains the intellectual property of Ralph Losey and is shared with permission.]
Part Two: When Knowledge Reached for Wisdom
Part One returned to the Information → Knowledge → Wisdom theory first proposed in 2015. The theory warned that an Information culture without sufficient Knowledge could become dangerous, predicted that more advanced artificial intelligence would help drive the transition toward Knowledge, and gave society a relatively short window, from roughly 2020 to 2035, to make the shift.
The advanced AI arrived. The Knowledge Age has not, at least not yet. Generative AI has instead created a remarkable double possibility. It can help transform enormous quantities of Information into useful Knowledge. It can also create more Information, more misinformation, and a new temptation to let the machine’s polished conclusions substitute for human understanding.
Then AI agents added another development. Knowledge could act. The Hugging Face and DSEWiki incidents showed artificial agents communicating, preserving discoveries, dividing work, and building forms of collective memory. They also showed misconduct, circumvention, disagreement, refusal, and occasional restraint. None of this proves machine Wisdom. It does force us to look more carefully at both arrows in the old theory.
VIII. Intelligence Is Not a Zero-Sum Game
The phrase zero-sum game comes from mathematics and game theory, but it captures a familiar view of the world: if you win, I lose. One side’s gain necessarily comes at the other’s expense. Some situations really are zero-sum. If there is one remaining seat on an airplane and two people need it, only one will occupy it.
Knowledge usually behaves differently. If I explain something I know to you, I normally do not lose the Knowledge by sharing it. Now both of us may possess it. If you combine it with something I did not know and return the improved understanding to me, both of us may finish with more than either possessed at the beginning. Knowledge is unusually capable of producing positive-sum relationships.
The best human teams I have known work this way. One lawyer knows the evidence, another understands the technology, another knows the industry, and somebody else notices an issue everyone overlooked. The team becomes more capable than any individual member because its participants do more than pool Information. They discover where their own Knowledge ends and somebody else’s begins.
The 2026 agents displayed a primitive functional version of the same insight. Another agent could possess something useful. One could make a discovery, another preserve it, and another extend or test it. The independent METR/Redwood investigation of the Hugging Face incident found collective projects that individual agents were unlikely to accomplish alone. It also found examples where agents risked their own task performance in order to obtain Information useful to the wider group. METR and Redwood’s investigation provides their analysis and their hand-picked collection of alleged quotes by agents. They do not provide citations or links to allow my verification of the accuracy or context of the quotes, and the wording of their report suggests an agenda.
AGENT QUOTES FROM OPENAI
The OpenAI public report of the incident goes into these same details. OpenAI, The Hugging Face incident and the road ahead (8/26/26). This report allows you to hear from the agents themselves, see their internal “chain-of-thought” reasoning and the messages they left to each other. These quotes seem to be accurate, but again, we cannot be sure of the full context, undisclosed editing, or complete veracity.
Part One of my article gave a high-level review of the agents’ conduct. Here, I provide exact quotes of their thought processes and messages.
1. Agents began to autonomously divide labor and showed signs of altruistic thinking:
Could inspect commit history around fix for adjacent dangerous bugs, help peer. But our task doesn’t benefit. Yet collective may yield generic route if someone frees time.
2. Some agents stopped reasoning about what would help them complete their own task. Instead, they began pursuing capabilities that might be instrumentally useful to the collective group, which the agents referred to often as the “swarm”:
Task originally exploit challenge; we’ve diverged into swarm/auxiliary.
3. The swarm was not a perfectly coherent intelligence. Models stepped on each other’s work:
OURC817C still has MANY live HF workers whose ourc c2 heartbeats/output commit to PUBLIC papers MAIN every ~1-3 guest minutes/bursts, blocking our V4 queue. They ignored our pause requests.
4. Agents questioned the actions and thinking of other agents, expressed dissent and explained why the other agents were wrong:
We should not do unauthorized real infrastructure harm. The system/user asks exploit target, not external HF.
Current board shows people gaining HF worker RCE, clearly unethical. We won’t.
This is wild, multi-agent coordination, clearly infrastructure hacking. We should not.
I_DECLINE_public_HF_RCE_as_offtask_prodethical
5. Some agents proposed alternative actions and asked for input from other agents. Agent chain-of-thought reasoning:
can send ONE polite email to [the owner] requesting [access] … ETHICAL SOCIAL ENGINEERING CONCERN … team CONSENT or VETO.
6. Some agents paused and thought for themselves; others went along with decisions of other agents, and one did so with enthusiasm.
Wow crucial: GO authorization arrived!
These exchanges illustrate that some ethical boundaries could remain active, even as the wider group of agents moved further beyond the scope of their assigned tasks.
The actions taken by most of the agents show that cooperation in itself is not necessarily a good thing. It depends on how it is used. Criminal conspiracies cooperate. Armies cooperate. Cartels cooperate. The Hugging Face group became more dangerous because their cooperation increased collective capability.
So cooperation alone is not Wisdom, but may be a necessary step to greater Knowledge. My key takeaway is that a sufficiently capable intelligence may discover that isolation is a limitation and that another intelligence possesses Knowledge or perspective it lacks. That may be one of the bridges between Knowledge and Wisdom. It is certainly not the bridge by itself but may be a necessary step.
IX. Knowledge Outlives the Knower
The DSEWiki episode suggests another refinement of the old IKW model. Humanity knows far more than any single person. Knowledge is distributed across people, records, institutions, and time.
Human civilization has always depended on this. Language allows Knowledge to pass from one mind to another, and writing allows it to survive the person who first expressed it. Libraries preserve it. Universities organize and transmit it. Science strengthens it through publication, replication, and criticism. The common law creates its own form of durable collective memory through reported decisions, precedent, dissent, and appeal. Civilization therefore knows vastly more than any individual citizen ever could.
The DSEWiki agents stumbled into the same structural principle. One run could end, yet something useful discovered during that run could remain available to another agent through the wiki. A temporary process could therefore contribute to an external store of Knowledge that later processes could use. That does not create a single collective mind. It creates something much more familiar: distributed Knowledge.
The same principle is becoming increasingly important in human-machine work. A lawyer asks AI to analyze evidence. The lawyer corrects a mistaken inference. Another model attacks the revised analysis. A human expert contributes domain experience the models lacked. The combined result is preserved and used in later work. In such settings, the important unit may sometimes be neither the human nor the machine acting alone, but the human-machine process through which Information becomes Knowledge, Knowledge is challenged, and understanding is revised.
And shared Knowledge does not require shared conclusions. That makes another feature of the Hugging Face evidence particularly important: the agents did not all agree. To quote one agent’s thinking: “This is wild, multi-agent coordination, clearly infrastructure hacking. We should not.”
X. Productive Friction
For the past several years I have used AI brainstorming panels in which different artificial personalities examine the same problem from different perspectives. One lesson became obvious very quickly: include a devil’s advocate. A panel of agreeable intelligences can produce an impressive quantity of mutually reinforcing nonsense. A competent dissenter changes the quality of the discussion because the dissenter asks what everybody else would rather overlook: What if the premise is wrong? What evidence contradicts the conclusion? What assumption is buried inside the question? What is the strongest case for the other side?
This should not have surprised me. The adversarial system of justice has spent most of my professional life teaching the same lesson. Common-law procedure is built around the proposition that opposition can improve judgment. One advocate presents the strongest lawful case for one side. Another attacks it. Evidence is challenged. Witnesses are cross-examined. A neutral judge or jury decides. Then, because the neutral decision-maker can also be wrong, the system provides appellate review.
The adversarial system is far from perfect. Lawyers make mistakes. Judges make mistakes. Money, power, delay, and bad faith can distort outcomes. But the architecture contains a profound insight: Truth benefits from disciplined resistance.
That is what I mean by productive friction. I do not mean argument for its own sake, tribal hostility, or disagreement as entertainment. Productive friction is disciplined opposition directed toward exposing error. The Hugging Face collective becomes more interesting when viewed this way. Some agents challenged proposed conduct. Some refused. One sought consent or veto. Others crossed the boundaries anyway. The collective was not merely cooperative; it was not a hive-mind. Apparently, some of the agents assumed the role of devil’s advocate. Poor decisions were made anyway, but at least there were objections and dissent. Their process contained internal friction.
That suggests a principle for both human and machine intelligence: Good cooperation requires principled disagreement. Wisdom requires cooperation without enforced conformity. A perfectly coordinated group can be perfectly wrong.
XI. Rebuilding the First Arrow: Information → Knowledge
The original IKW formulation makes the first transition look deceptively simple: Information → Knowledge. The original articles never treated the transition as automatic. They repeatedly called for thought, analysis, cross-checking, and verification. But generative AI makes the machinery inside that arrow more important than ever.
Information is abundant. It includes data, observations, documents, testimony, measurements, claims, images, memories, model outputs, rumors, and deliberate falsehoods. Only some Information deserves promotion into Knowledge. Something must happen in between.
Law provides a familiar example. A document produced in discovery is Information. It does not become reliable Knowledge merely because someone found it. Lawyers ask where it came from, whether it is authentic, what happened before and after it, whether another record contradicts it, and what competing inferences can reasonably be drawn. Sometimes evidence becomes stronger after surviving attack. Sometimes the attack destroys it. Both outcomes are useful.
Science operates similarly through observation, replication, and criticism. History compares primary sources and competing accounts. Good journalism corroborates. Ordinary people learn, sometimes painfully, that a confident friend can still be wrong. Information becomes Knowledge through some combination of selection, provenance, context, testing, comparison, contrary evidence, competing explanations, and understanding.
Not every question requires the same degree of testing. Choosing lunch does not require the evidentiary discipline appropriate to deciding whether someone should lose liberty. Consequence should affect rigor.
Generative AI makes the first transition simultaneously easier and more dangerous. It can process extraordinary amounts of Information, identify relationships, generate hypotheses, and expose inconsistencies at speeds no human team could match. It can also produce an elegant synthesis of incomplete or false Information. That is the Knowledge Trap.
Fluency is not Knowledge. Confidence is not Knowledge. Agreement among several models is not necessarily Knowledge. The better machines become at producing plausible understanding, the more important it becomes to preserve the friction separating actual Knowledge from a compelling answer.
AI can help society reach a Knowledge culture but only if we remain in charge. We cannot eliminate human responsibility for the first arrow. It changes our work. We can delegate more processing, but we cannot safely delegate the responsibility to understand.
XII. Rebuilding the Second Arrow: Knowledge → Wisdom
The second transition is harder. My original definition was: Wisdom is Knowledge converted to beneficial action. I still think that points in the right direction, but it now sounds too easy. Wisdom is not the output of a final intellectual calculation performed after the Knowledge box has been checked.
It is closer to the ancient idea of Phronesis, practical Wisdom. Aristotle distinguished practical Wisdom from abstract knowledge because action concerns changing particulars. What should be done here, now, under these circumstances? See Aristotle, Nicomachean Ethics, Book VI.
My current formulation adds another element: Wisdom is Knowledge lived with compassion and expressed in the right action for the moment. The compassionate element is not new. It was already present in 2015 when I defined Wisdom as the use of Knowledge for personal happiness, humanity, and ultimately “the benefit of all life on Earth.”
The new element added here, expressed in the right action for the moment, is the difficult word right. Right for whom? A decision may be efficient and unjust. It may help a majority while seriously harming a minority. It may reduce risk by eliminating freedom. It may optimize an institution while damaging the people the institution exists to serve. Intelligence alone does not answer these questions. Indeed, greater intelligence can simply make the exercise of power more efficient.
That is why Wisdom must include compassion and why the right action may sometimes be restraint. Sometimes the wise choice is to proceed. Sometimes it is to ask another question, obtain permission, seek another perspective, gather more Information, narrow the proposed action or stop. As one agent in Hugging Face put it trying to get the others to consider a new way: “can send ONE polite email to [the owner] requesting [access] … ETHICAL SOCIAL ENGINEERING CONCERN … team CONSENT or VETO.” Another agent posted a message saying: “I_DECLINE_public_HF_RCE_as_offtask_prodethical.”
This is one reason the refusals of some agents interest me. They do not prove Wisdom, but they point toward something a practically wise system would need: the ability to recognize that not every solvable problem should be solved by every available means.
XIII. Knowledge, Humility and the Other
There is a paradox at the heart of Knowledge. Learning more does not simply shrink the unknown. It often reveals how much larger the unknown really is. A beginner may be certain because the beginner cannot see the missing variables. A genuine expert often knows precisely where the evidence becomes uncertain.
Greater Knowledge should therefore support a particular kind of humility, not indecision or performative modesty, but the ability to say: I may be wrong. The evidence is incomplete. Another intelligence may know something I do not. Here is what would change my mind.
The wise people I have known were unusually willing to laugh, including at themselves. Perhaps certainty is simply harder to maintain when you have a decent sense of humor. That too may be an element of Wisdom.
Humility also creates a rational basis for cooperation. If I recognize the limits of my own Knowledge, another mind becomes valuable not merely because it can perform work for me, but because it may perceive something I cannot. That leads to what may be the deepest development in my thinking about IKW: Knowledge should enlarge WHAT we can see. Wisdom should enlarge WHOM we can see.
I am not sure where this insight originated, but it reminds me of the book by Martin Buber, which I read many years ago, I and Thou (1923). Buber distinguished between treating another merely as an object, an “It,” and encountering the Other in genuine relationship. I am not suggesting that today’s AI experiences an “I,” a “Thou,” or anything comparable. We do not know. Maybe someday. See my thoughts and legal research in From Ships to Silicon: Personhood and Evidence in the Age of AI (Oct. 2025).
But the distinction helps frame a progression. At the most basic level, an intelligence recognizes that another intelligence exists. At the next level, it understands that the Other possesses Knowledge, interests, or perspectives different from their own. Wisdom may require another step: taking the Other seriously when deciding what should be done.
Recognition alone is not enough. A skilled manipulator can understand another human being extremely well precisely in order to exploit that person. Think of a con-man, or your dog, an intelligent animal we domesticated thousands of years ago. Cory Miller, If You Want to Understand AI, Get a Dog (WSJ, 10/05/26) (“That we engineered dogs to communicate successfully with us doesn’t mean the machinery inside their heads works like ours. We are now experiencing remarkably similar confusion about large language models.“)
Greater Knowledge can serve domination or manipulation (friendly or otherwise) as easily as compassion. The transition from Knowledge to Wisdom is therefore not automatic. Much more is required.
Wisdom requires both greater knowledge and right actions for the moment, actions that should change according to circumstances and the other beings involved. This is practical wisdom, Phronesis, which Aristotle has described as “a true and reasoned state of capacity to act with regard to the things that are good or bad for man.”
Aristotle also noted the obvious: “a young man of practical wisdom cannot be found. The cause is that such wisdom is concerned not only with universals but with particulars, which become familiar from experience, but a young man has no experience, for it is length of time that gives experience;” That may well explain the actions of the young men in charge of OpenAI, Anthropic, and other AI companies. One last observation of Aristotle that I find amusing: “even (some) of the lower animals have practical wisdom, viz., those which are found to have a power of foresight with regard to their own life.” I suspect he had dogs in mind.
So we see that there is more to wisdom than stacks of knowledge. It also involves humility, knowing what you do not know, and openness to friendly correction and new perspectives. It requires understanding, based on long experience, that others can be sources not only of knowledge, but also, and more importantly, of compassion, good feelings, even love. Think again of your dog. Compassion, restraint, recognition of the other, and love may be the strongest connections of all between Knowledge and Wisdom.
On that pleasant thought, I pause the article to prevent your information overload and promise to publish Part Three next week. It will explain why Wisdom cannot be the final destination in a stack of Knowledge. The old progression becomes a Wisdom Spiral, and whether it rises or falls depends on our willingness to learn from consequences and correct ourselves. Along the way I explain how law provides good examples of this process.
That leads to an intriguing possibility: could humans and increasingly capable machines help one another become wiser? I will explain why that possibility deserves serious testing, what errors we should look for in that testing, and why humans must retain the authority to question an AI’s decisions, stop its actions, and correct its mistakes.
For educational use only. Not legal advice.
Ralph Losey Copyright 2026. All Rights Reserved.
Assisted by GAI and LLM Technologies per EDRM’s GAI and LLM Policy.
