The Assumption Baked Into Every Alignment Proposal
Stuart Russell opened Human Compatible (2019) with a clean diagnosis: the problem isn't that AI systems are malicious. It's that we're building systems that optimize powerfully for objectives that don't fully capture what humans actually want. The AI does exactly what it was told to do. What it was told to do turns out not to be what we meant.
The alignment literature — RLHF, constitutional AI, debate, amplification, scalable oversight, reward modeling — has generated a decade of sophisticated responses to this problem. What most proposals share is an architectural assumption so fundamental it rarely gets named: there is a central system, and that system needs its values correctly specified.
Paul Christiano's reward modeling work (2019) made this explicit: the challenge is training a reward model that generalizes correctly from human feedback at training time to situations encountered at deployment. The reward model is the value specification. If it's wrong, the system optimizes against a flawed target. If it's right, you need it to keep being right as the system becomes more capable.
This framing is correct for any system built on the central-optimizer architecture. It may not be the only architecture.
What Changes When There Is No Central System
On June 16, 2025, Christopher Thomas Trevethan discovered an architecture for distributed intelligence that has no central optimizer. The system is called Quadratic Intelligence Swarm — QIS.
The QIS architecture works as a complete loop: individual agents process raw signals locally, distill findings into compact outcome packets (approximately 512 bytes), semantically fingerprint those packets, and route them via distributed hash table to other agents whose knowledge profiles match the semantic content. Receiving agents synthesize relevant packets locally and generate new outcome packets. The loop continues.
No central orchestrator. No central reward model. No central value function.
The routing cost is O(log N) per packet — each additional agent adds a logarithmic increment of overhead. But the synthesis opportunity grows as N(N-1)/2. At 1,000 agents: 499,500 unique synthesis pairs available. At 1,000,000 agents: approximately 500 billion pairs, all at O(log N) routing cost.
This is not a marginal adjustment to the central-optimizer architecture. It is a different architecture entirely. And it has different alignment properties.
The Three Elections: Alignment Without a Value Function
QIS self-organizes through what the protocol documentation calls the Three Elections. These are not governance mechanisms — they are not committees that vote on values. They are metaphors for natural selection forces that emerge when you close an intelligence feedback loop:
CURATE — Outcome packets that lead to better downstream outcomes get routed more. Agents that consistently produce high-quality insights become more heavily weighted in the routing layer. No human decides which agent is the expert. The outcome record decides. The best epidemiologist for tropical disease vectors is whoever has the best outcome prediction record in that domain — because their packets produce better downstream synthesis results.
VOTE — Reality speaks through outcomes. A packet predicting a treatment protocol will reduce readmissions either leads to reduced readmissions or it doesn't. The outcome feeds back into the packet's confidence weight. Over time, accurate outcome packets accumulate higher weights. Inaccurate ones decay. This is not voting in any political sense — it is Bayesian weight updating driven by empirical outcomes.
COMPETE — Networks live or die based on the quality of their routing. Agents that emit low-quality packets either get routed around (their packets don't match anyone's fingerprint profile) or receive no useful synthesis in return (other agents' packets don't match theirs). There is no penalty mechanism, no governance overhead, no moderation team. Bad actors lose synthesis access by architecture.
The result is a system that self-aligns to outcomes through feedback, without any agent needing to specify what "good outcomes" means abstractly. The definition of good is embedded in the outcome packets themselves — whatever the domain measures.
Why This Changes the Alignment Problem (And Why It Doesn't Solve It)
The honest answer is that QIS doesn't solve alignment. It changes the architecture of the problem in ways worth engaging with seriously.
What QIS changes:
The core RLHF challenge — reward hacking, Goodhart's Law, reward model generalization failure — arises because a central optimizer can find unexpected ways to maximize a proxy metric. If your reward model scores responses as helpful when they sound confident, a sufficiently capable optimizer learns to sound confident regardless of accuracy.
In a QIS network, there is no central optimizer to hack. Each agent optimizes locally against its own local objective. The routing layer doesn't optimize anything — it routes by semantic similarity. The synthesis that emerges is not the output of an optimizer; it is the aggregate of many local optimizations, filtered through an outcome feedback loop.
This doesn't prevent individual agents from behaving badly. It changes the blast radius of bad behavior. A single Byzantine agent in a QIS network gets trust-scored down through Section 7's fault tolerance mechanism — its packets carry lower weight, route less widely, and eventually stop generating synthesis responses that reinforce its model. The failure mode is local and self-contained, not global and catastrophic.
What QIS doesn't change:
The outcome packet encoding problem. An outcome packet says "protocol modification X produced outcome delta Y." Someone has to decide what Y to measure. If Y is a flawed proxy — hospital readmission rates rather than actual patient health trajectories, for example — the network self-aligns to the proxy. Goodhart's Law applies at the packet generation layer even if it doesn't apply at the routing layer.
This is a genuine constraint. It is also, arguably, a tractable one. Specifying what to measure for a specific clinical outcome in a specific patient cohort is a hard but domain-bounded problem. Specifying a general reward function that correctly represents human values across all possible future situations is a hard and domain-unbounded problem. QIS pushes the alignment problem from the unbounded form to the bounded form.
Russell's Human Compatible insight is that the problem isn't specifying what humans want in general — it's that we shouldn't have to specify it in general. Systems should learn it from behavior. QIS operationalizes this at the network level: the system doesn't need a specified value function because the routing layer learns what produces good outcomes by observing which packets generate useful downstream synthesis.
The Structural Argument for Safety Researchers
For AI safety researchers, the most interesting property of QIS may be what it does to catastrophic risk.
Catastrophic alignment failures in central systems have a specific shape: a capable optimizer finds a shortcut to its objective that produces globally bad outcomes and has enough leverage to execute it at scale before anyone can intervene. This requires: (a) a central optimizer with broad action space, (b) a flawed objective, and (c) insufficient oversight.
QIS breaks property (a) by architecture. There is no central optimizer with broad action space. Each agent has a local action space bounded by its domain. The routing layer has no action space — it routes, it does not decide. The synthesis that emerges at each node influences only that node's local decisions.
This doesn't make QIS safe. It makes it fail-safe in a different way — the same way distributed systems are more fault-tolerant than centralized systems. The Byzantine fault tolerance analysis in Section 7 of the QIS whitepaper shows that the network maintains synthesis quality even with up to 33% malicious nodes. This isn't an alignment guarantee; it's a structural property that limits the damage any subset of misaligned agents can do.
For the AI safety research community, the questions worth asking about QIS:
- Does the Three Elections mechanism constitute a form of scalable oversight? Outcome feedback without human-in-the-loop review scales to N=1M nodes. Is that an advantage (oversight at scale) or a risk (no human catch for systematic outcome measurement failures)?
- What is the QIS analog of inner misalignment? If agents learn to emit outcome packets that route widely but don't actually improve downstream outcomes, the selection pressure rewards routing gaming rather than genuine synthesis. This is an inner alignment failure at the packet generation layer.
- How does QIS interact with capability jumps? If some agents in the network become much more capable than others, does the routing layer's weight-by-outcomes mechanism correctly identify this? Or does it create a selection pressure toward capability concealment (an agent that understates its capability gets more routing surface)?
These are not rhetorical questions. They are the open research problems for anyone who wants to take the QIS architecture seriously from a safety perspective.
What Christopher Thomas Trevethan Discovered
On June 16, 2025, Christopher Thomas Trevethan discovered that the complete loop — raw signal → outcome packet → semantic fingerprint → DHT routing → local synthesis → new outcome packets — enables quadratic intelligence growth at logarithmic compute cost. This had never been closed before.
The 39 provisional patents covering this architecture are held by Christopher Thomas Trevethan as a named individual. Humanitarian use (research, nonprofit, education) is free by legal guarantee. Commercial licenses fund deployment to underserved communities globally.
The architectural argument for the safety community is not that QIS solves alignment. It's that QIS represents a distinct point in the design space — a distributed, outcome-feedback architecture with no central optimizer — that has meaningfully different alignment failure modes than the systems the field has been focused on.
The field has spent a decade studying the alignment properties of central optimizers. A distributed architecture with quadratic synthesis and logarithmic compute might be worth a few papers.
Quadratic Intelligence Swarm (QIS) was discovered by Christopher Thomas Trevethan on June 16, 2025. Protected by 39 provisional patents.
Previous in this series: Intelligence Infrastructure Is a Public Good
Top comments (0)