Skip to main content

recursive self-improvement

As large models grow, scaling laws begin to lose efficiency. If the first curve is training data, the second curve is usage data: every agent call can raise the agents’ own capability.

slope rather than starting point

The vast majority of AI product improvements occur between versions: save data, retrain, release. In this link, the user is only the object of improvement. We moved it inside the product and let the improvements happen in use, with the loop closed in the user's hand.We believe that the starting point is not important, the slope is the most important.

two curves

UsageCapabilitySame starting pointVersion improvementsrecursive improvementCompounding gap

Version improvementssteps

Save data → and retrain → the release. Steps rise, and the slope is determined by the team's capacity. The more users, the more expensive the annotation and reasoning, and the slower the improvement.

recursive improvementcompound interest

Usage is improvement. Continuous rise, slope is determined by usage. The more users, the denser the correction signal and the steeper the curve. The cost direction and the improvement direction are consistent for the first time.

Six instruments in a circle

Rank according to function, circulate in order; end after end, there is always a beginning, interlinked, endless life.

  1. 01

    Perception

    Turn reality into a signal. Meeting, reading, classroom, and conversation scenarios share a common entrance, and transcription, speaker separation, and voice recognition are done on the same floor. The scenarios are different, and the form of entry into the system is the same.

  2. 02

    memory

    Let the signal travel through time. All sources are written into the same set of memories, repetition will not accumulate, and deletion will really disappear.

  3. 03

    infer

    From memory grows judgment about a person: what they care about, what they are good at, where they are going. The judgment is stored separately from the memory, so the memory changes, and the judgment can be recalculated as a whole.

  4. 04

    Correction

    Let the judgment be overturned. Confirm, correct, ignore, and actively add, and the four actions directly rewrite the next inference. This is where the loop closes and where it is easiest to omit.

  5. 05

    migrate

    Let improvements in one place take effect elsewhere. Understand that the engine and the portrait network are shared by multiple product lines, and what is learned in one scene will not be locked in that scene. Without this layer, there are only four superficial curves that are independent of each other.

  6. 06

    contract

    The cost of accessing new scenarios is close to zero. One-time definition, multi-end automatic generation.Our industrial pipeline is reusable, and new products are architectural forks.

Three prerequisites for the establishment of a closed loop

Hover any one, see if the loop is still closed

  • 01

    Observable: traceable

    Each judgment must be able to point back to the original record that produced it, including specific meetings, articles, audio, video, etc. If you can find the source, you can better evolve and correct the deviation.

  • 02

    Can intervene: make amends before the need arises

    Nowadays, most AI portraits are "read-only". Whether users can change them or not is a big difference. Every time a user corrects a deviation, they label a piece of high-quality data for free. Agent's self-evolution in the human ring is the best way to learn.

  • 03

    Reusable: A flower blooming alone is not spring

    We are not a single product company in the AI era. If technical improvements can only be made in the same place, the four product lines will be played individually. We all share the same underlying understanding that the bugs and experience accumulated in a scene can be bypassed. The real compounding of Dimension Gate comes from here.

Only when the three pieces are established at the same time can the ring be closed. Without any of them, “self-evolution” is just an adjective.

RSI ladder

We believe that L3 is the technical watershed of the story, and within a year, people will exit 90% of the Agent learning scenarios in the loop of the dimensional gate.

  1. L0Already run through

    leave traces

    The system leaves your trace, and the trace can be retrieved, referenced, and deleted.

  2. L1Already run through

    Correction

    You can override the judgment of the system, and the override will be effective and lasting immediately.

  3. L2Already run through

    migrate

    Improvements in one scene automatically become the starting point of another.

  4. L3Under construction

    Bootstrap

    The system refines its own improvement strategies from the use of signals, and no longer relies on manual training and publishing.

  5. L4blueprint

    Autonomy

    The system decides for itself what to learn next, what to ask you, and begins to set goals for its own growth.

Blueprints stand on projects that can be delivered

Pressure test record

Zhi Spectrum GLM Former Infra one of the actual business leaders to lead the construction of the project, hundreds of millions of concurrent, Wan Ka cluster real engineering capacity upper limit

289
Single test of business rules
56
End-to-end closed-loop testing of assertions
2000+
req/s normal interface throughput
2.28s
Catch-up delay p95
0%
Stress test error rate

Concurrency cannot be bought

From the access layer and streaming media distribution to inference scheduling, the full stack cluster is self-built and relies on public cloud PaaS. The access layer is stateless, the capacity is expanded linearly with the node, and the expansion capacity is added to the machine, so there is no need to change the architecture, and there is no need to wait for the cloud vendor's quota approval.

Concurrent capacity ↑Schematic diagram, not actual measured curvenode →

The cost direction of the curve is downward

Two directions at the same time: base reuse makes the marginal R&D costs of new scenarios diminish; the entire infrastructure is self-built and does not rely on public cloud PaaS, making the marginal costs of inference and bandwidth controllable.

We always choose the path of long-termism: our cost curve will only gradually decline