Cooperative General Intelligence_

I was invited at the last minute, without a synchronized deck or a finished protocol. Cooperative General Intelligence is the umbrella I use for a question linking conditional recall and information markets. This is a design exploration, not a report that a deployed CGI system works.

The question that haunts me is whether there are benefits to genuinely dispersed information even if an ASI has access to everything at once and can simulate holding some information back. Does genuinely dispersed information still enable something that simulated separation does not, or is separation only a cost? I do not have an answer. The rest of the talk explores the institutional primitives that might matter if the answer is yes.

“Are there any structural benefits to retaining dispersed information over having a single monolithic intelligence, with access to all information simultaneously?”

My foundation for the rest is agents plus commitments. The agents may run in TEEs—or in an ordinary cloud environment with explicit trust assumptions—but the environment must not leak its information. TEEs alone do not guarantee complex behavioral commitments, so the design also needs monitoring and competitive mechanisms that can enforce those commitments.

Information markets have an old sequencing problem. A seller needs to convince a buyer of the value of information, but the buyer has to look at a piece of it and cannot reliably forget after refusing the price. There is no Men in Black amnesia scanner or magic pill that makes the buyer forget what they have already seen.

Agents running in TEEs with behavioral commitments can change that sequence. A buyer-side agent looks at a piece of information inside the confidential environment and decides whether to purchase it. If the buyer declines, the agent destroys that context and proves that it destroyed the context. The claim is not that hardware erases every consequence everywhere. It is narrower: an enforceable behavioral commitment can make rejection include forgetting.

That guarantee belongs to the whole agent-and-commitment system. The protected environment contains the context; the behavioral rule and its monitoring make the destruction step meaningful. Without that chain, the seller has not escaped the information paradox after a rejected sale.

Bounded extraction changes what the buyer purchases. Instead of acquiring a private archive, the buyer pays to execute one task over it. Imagine a venture investor who lacks materials-science expertise and needs help evaluating a startup. A diligence agent can buy the right to access all of a researcher’s private context inside the protected environment, yet only output a number in the range of zero to one. The researcher’s information remains private; the investor receives a decision signal.

The researcher retains the underlying notes; the investor receives a decision signal. Yet, as I stress, every returned signal leaks some information. A score, a binary answer, or even the decision to transact reveals something. Repeated questions can reveal more, and a carefully chosen query may extract much more than its small output type suggests.

“Even in this scheme, the information is leaked.”

The promise is controlled access, not perfect secrecy. Each task can return only a bounded result, but the system still has to measure how much information escapes each time it runs. The primitive is private information remaining in separate pockets while agents execute over it—not a claim that no information leaves.

Forgetting and bounded extraction combine in what I call a superset-state agent. One class of agents acts for users—making money in DeFi, buying and selling goods, or handling other tasks. A second class can inspect the first group’s weights, actions, environment, and private resources under a commitment not to leak that wider state. The inspector can compute across the boundary even when the underlying state remains inside its separate pockets.

Today, a model provider can watch what its agents and users are doing and run classifiers over that activity. My version moves the inspector into a confidential environment and restricts what it can reveal. It sees broader state than the agents it oversees, while the rest of the system receives only a bounded output.

Forgetting, bounded extraction, and superset-state inspection can compose with attribution incentives, decision markets, challenge markets, and auctions. I also imagine connections that are stranger than markets—something like neuronal links between separate pockets of information—but that part of the sketch is explicitly more speculative.

Failure traces are the cleanest example. Companies and researchers accumulate valuable records of approaches that did not work, often for good reasons that prevent publication. A constrained agent could inspect those failures and steer a new process away from dead ends. The buyer receives a change in search direction rather than the sensitive trace itself.

A hiring market makes the privacy stakes personal. Someone who wants to leave a job may not want the current employer to know they are talking to recruiters. The worker could place a rich profile—résumé, code, completed tasks, workplace communications, and private negotiation terms—inside an encrypted environment. Employer agents inspect that profile and negotiate before either side reveals interest publicly. For the result to be credible, the work history needs authentic receipts rather than self-asserted claims, and the actual hire may still move into a human process.

Routing can use information people hesitate to publish, such as negative reviews or a transaction graph. An agent buys access to those private signals and routes buyers, sellers, or other agents without exposing the entire graph. Branching negotiations add another possibility: agents reveal strategies to one another, explore terms, then forget abandoned branches before settling on an agreement.

We currently have a wall between what markets can do and what firms can do. Markets are effective at aggregating dispersed information, but they collapse that information into a price in order to preserve its privacy. Firms enable trust and can operate on private information, but they are inflexible. I am interested in institutions between those poles: combinations of private agents that compute, bargain, and route without placing every input in one company or one public market.

This only matters under the condition I stated at the beginning: dispersed information must retain some structural advantage. If an ASI can see everything and separation is simply inferior, the spectrum may collapse toward the monolithic model. CGI is my name for exploring the other case, where private pockets of information remain valuable enough to coordinate.

The largest caveat is that commitment enforcement remains adversarial. Attestation may establish a simple boundary: a particular program ran in a particular protected environment. It does not by itself prove that a flexible agent forgot a context, obeyed negotiation rules, resisted side channels, or avoided collusion with another agent.

“Commitment enforcement is like a cat and mouse game.”

I break commitment enforcement into three broad categories. TEE attestation can probably solve some simple commitments. For another category, a deterministic program can inspect a given output and sign a key that enables the tool call only when the rule passes. The much larger and harder category needs competition between monitoring agents: they watch private activity and try to detect misalignment or rule-breaking without leaking the underlying information.

One sketch lets a monitor stake on whether misconduct occurred and lets other monitors join either side. The private contest still needs a terminal resolution. In rare escalations, the disputed activity has to leave the confidential environment and reach an external judge that can evaluate the ground truth.

Attribution is its own hard problem. Exact attribution is computationally expensive. Every cheaper method makes a trade-off. Markets can provide immediate attribution, but they lack enough look-ahead to capture a contribution whose value becomes clear only later.

“Shapley is optimal for a lot of this, but computationally a nightmare.”

Finally, every action inside a private system still reveals something to an observer about what the system contains. That leakage problem is less central than commitment enforcement or attribution, but it does not disappear. Any useful design has to measure what each output or action exposes.

The speaker opens by saying there are no actual slides. The supplied deck is a later semantic summary rather than a synchronized visual record, so its useful pages appear only here and should not be read as evidence that these frames accompanied particular sentences or establish any chronology among them.

This frame restates the opening question: whether structurally dispersed information retains benefits when a future monolithic intelligence could integrate everything, even if that integration is delayed or selectively simulated.

The next frame condenses the assumed infrastructure. Agents run inside confidential environments and make behavioral commitments, while monitoring mechanisms address promises that hardware isolation cannot establish on its own.

The primitives are forgetting rejected context, extracting only a bounded result, and letting a superset-state agent inspect broader private activity without gaining permission to publish that state.

The proposal composes those primitives with incentives and coordination mechanisms. Markets, challenges, auctions, attribution, and less familiar agent connections become candidate ways to coordinate private pockets of information.

Possible applications range from expertise and private failure data to oversight, routing, and negotiations. Each uses computation over protected context while attempting to constrain what the result reveals.

This frame presents the institutional axis: markets aggregate dispersed knowledge by compressing it into prices, while firms lower coordination costs by operating across private information inside boundaries.

The closing frame names the unresolved work. Complex commitments need enforcement, contributions need tractable attribution, and every action over private state needs an accounting of information leakage.