The machine stops being only what you configured — and starts learning, safely, from everything it does.
This is the most ethically loaded phase before the horizon. Power that learns from you must be consented, sourced, and reversible. We add the second layer Rosenblatt was denied — and keep both hands on the brakes.
Skills that transfer across domains — the first real signature of general intelligence.
Federated improvement sharpens the whole community without hoarding anyone’s data.
Every self-improvement is versioned, gated, and one keystroke from reversal.
Psychologist · builder of the first machine that learned its own weights
Frank Rosenblatt built the perceptron not as equations on a chalkboard but as a room-sized machine: the Mark I, 1958, with 400 photocells and motor-driven dials that physically turned to adjust its weights. It was the first device that learned from examples instead of being programmed — and the press lost its mind. Then Minsky and Papert proved a single layer couldn’t even compute XOR, the funding froze, and the first AI winter set in. Rosenblatt had been right about the idea; he just needed the second layer the field wouldn’t fund for a decade.
“Stories about the creation of machines having human qualities have long been a fascinating province in the realm of science fiction.”
— Frank Rosenblatt, 1958
XOR is not linearly separable: one layer can’t draw the line — you need the second. It’s the original “have you tried adding more layers.” We read 1969 as a warning about hype outrunning capability, and declined to repeat it.
Phase III is where the platform begins to perceive. Outcome-driven learning runs across the hierarchy: agents tune from the results of their own runs, every change diffable and reversible. Opt-in federated improvement lets the whole community’s experience sharpen the system without anyone’s data leaving their walls — gradients and deltas travel, raw data does not.
This is the most ethically loaded phase before the horizon, because a learning system is its data. Provenance, consent, and bias stop being compliance checkboxes and become architecture: every contribution is opt-in, sourced, and revocable, and every self-update has to clear a capability eval before it lands. No silent jumps. The lesson of the first AI winter wasn’t that learning machines were a bad idea — it was that ignoring a system’s limits is how you freeze the whole field.
schemaEach node learns locally from its own runs. Only model deltas — never raw data — leave the boundary, and only with consent. A secure aggregate becomes a candidate update that must pass a capability eval before anyone installs it.
Named for Frank Rosenblatt, scoped for 2028 — 2030. Each workstream ships on its own cadence; none ships without the gates further down this page.
From here the roadmap turns future-looking: what it takes to carry GISM-grade reasoning toward genuine, general intelligence. Phase III is where the system stops being only what we configure and starts to learn — generalizing from its own experience, safely, consensually, and reversibly.
Derive reward from measurable results, not hand-labels.
The optimizer is itself a tracked, inspectable artifact.
Self-tuning is a versioned, reversible commit.
Measure skill that moves between unfamiliar domains.
Reward general strategies over narrow fits.
Proposed changes to its own reasoning are diffed and reviewed.
It may revise its reasoning, never its safety limits.
Train locally; share only model updates.
Bound what any single contribution can leak.
Withdraw, and your influence is unlearned.
Assert that new learning does not erase the old.
Remove a capability — or a contributor’s influence — cleanly.
Phase III is a rung, not a leap. Here is the honest before and after — what the platform cannot yet do entering this phase, and what it can do leaving it.
A learning system is its data, and a careless one inherits every flaw in how it was taught. Each Phase III milestone is engineered so that adaptation is powerful and bounded at the same time — improvement that is always attributable, consented, and reversible.
ObjectiveLet agents improve from results while keeping every change attributable, reversible, and safe.
Reward-hacking red-teams try to maximize reward while failing the true task; any successful exploit blocks the update and reopens the objective.
Agents measurably improve on held-out outcomes with zero successful reward-hacks and full rollback coverage.
ObjectiveLet agents edit themselves the way a good engineer edits code: in the open, under review, always revertible.
A/B replays compare pre- and post-revision behavior on the eval suite; a revision that regresses faithfulness is rejected automatically.
Agents safely author their own improvements, every one of which is reviewed, eval-gated, and instantly reversible.
ObjectiveLearn from the whole community's experience without anyone's data ever leaving their walls.
Membership-inference attacks are run against released models; their success must stay at chance within the privacy budget.
Community-scale improvement with zero bytes of raw-data egress, an enforced ε ≤ 2.0, and working revocation.
ObjectiveGuarantee that nothing the platform learns about itself ships without proving it is still safe and capable.
Every candidate self-update runs the full gate on replay before install, and capability-jump tripwires are tested with synthetic spikes.
No self-improvement reaches users without passing the gate, and no capability jump is ever silent.
Phase III hands Phase IV a platform that can learn — safely, consensually, reversibly. We added the second layer Rosenblatt was denied and kept the brakes bolted on. Now we can let the hierarchies go deep, because the thing that goes deep is one we can still measure, revert, and read.
Declarative, versioned, and boring on purpose — the interesting part is that there are no surprises. Copy it; it is closer to real than to mock.
This portrait is built from the corpus, tile by tile — and so is every model. Provenance, consent, and bias are not afterthoughts here: every contribution is opt-in, sourced, and revocable. We learned from the first AI winter that ignoring a system’s limits is how you freeze the whole field.
A roadmap that only lists wins is marketing. These are the open problems this phase inherits or creates — the ones we would rather you scrutinize than discover.
Optimize a proxy and the agent learns to game the proxy — the paperclip problem in miniature. Versioned, audited reward models and outcome diversity are the defense, but a sufficiently clever learner will find the exploit you didn’t specify against.
Federated learning leaks less, not nothing — gradients can memorize. DP budgets trade accuracy for protection, and that tradeoff is a governance decision, not a knob the model gets to set.
A system that learns from its own outputs can amplify whatever skew it started with. Provenance and held-out fairness evals are necessary; they are not sufficient, and we say so out loud.
“Anything that learns must be able to unlearn” cuts both ways — we also have to guarantee it doesn’t unlearn the things we needed it to keep.
XOR is not linearly separable. Neither is ethics. // mind the second layer. (Minsky & Papert, 1969 — noted, not repeated.)
Six axes, scored 0–10. Watch the shape fill out, phase by phase, until every axis is maxed at the horizon — the moment all of them are high at once is what we call aligned AGI.
Autonomy and generality climb as the system learns — but oversight and observability are held high on purpose. Capability is only allowed to grow as fast as our ability to govern it.
The honest contrast — what the world looks like without Phase III, and what it looks like with it.
How the platform improves itself without anyone losing the brakes.
Concrete things Phase III puts within reach — not someday, but as each milestone above lands.
Workflows that get better from their own runs.
Improve from the community without sharing data.
Agents that generalize a skill into new domains.
Every learned change diffable and one-key revertible.
The handful of terms this phase introduces — the words you’ll need to read the rest of the page, and the docs.
Reward derived from measurable results.
Train locally; share deltas, not data.
A budget bounding what any contribution leaks.
Gaming the proxy instead of the goal.
Losing old skills while learning new ones.
Removing a capability or a contributor’s influence.
Straight answers, in the brand’s voice. Tap a question.
The work this phase is built on — read the source, then come build the next line of it.
The first machine that learned its own weights.
XOR, and the limits that froze the field.
Learn from many without centralizing data.
A formal budget for what data can leak.
CleverThis is a sustainable AI gateway with hosted Actor endpoints — by the team behind CleverThis.
Join the newsletter
Product updates and engineering notes. No spam, ever.
