Solution basin aggregation: why brainstorming changes outcomes

A critical innovation of the Intrafere Research Group is the introduction of brainstorming coupled with a completely separate and supplementary technique of acceptance or rejection with feedback. Unlike traditional autonomous loops that often force an answer regardless of quality, MOTO employs a rigorous gatekeeper.

This brainstorming is not just a pre-step—it is a capability amplifier. Transformers are optimized to predict what comes next, so when the model sees its own prior ideas ahead of time, it is less likely to regurgitate the first commonly-known fact. This allows deeper and deeper probing of the solution space and compounding insight. We refer to this as solution basin aggregation, and it produces entry-level ASI results in practice because each pass explores a richer, more informed landscape. With this solution aggregation method, the creativity level approaches a capability that is comparable to very early “artificial superintelligence” (ASI-like). While the term artificial superintelligence is a subjective bar, we view MOTO’s brainstorming and validation architecture as the first discovered method for more efficient knowledge extraction from LLM weights. MOTO essentially “mines” the creativity from a transformer’s knowledge set and cross-recombination of “mined” knowledge compounds to create new insights that do not exist from training. Intrafere does not claim it will be the only path toward superintelligence: other Top-P strategies and other neural-network superintelligence approaches that differ from MOTO may also prove effective.

Disclaimer: MOTO can produce incorrect and/or hallucinatory answers. MOTO Autonomous S.T.E.M. ASI intentionally prioritizes novelty, at the potential cost of being incorrect.

Validation, Rejection, and Pruning Keep the Database Clean

MOTO’s brainstorming phase runs multiple submitters in parallel, each independently exploring the solution space and generating candidate insights simultaneously. All of these parallel submissions funnel into a single bottleneck validator — a completely separate model instance whose only job is to evaluate whether each submission genuinely advances the knowledge base toward solving the user’s original prompt. This architectural separation between creative exploration and critical evaluation mitigates the hallucination loops and drift that plague single-model autonomous agents – mitigating the alignment problem.

When a submission arrives, the validator asks a focused question: does this entry make the knowledge base more capable of addressing the user’s goal? If yes, the submission is accepted and added to the shared brainstorm database. If not, the validator rejects it with specific feedback explaining why — and that feedback is returned to the submitter, guiding future attempts toward more productive directions and avoiding rejection loops. This means failure is not wasted; every rejection is a learning signal that steers the next round of exploration. The validator acts as a guardrail against prompt misalignment, continually pulling the brainstorm back toward the user’s actual intent even as submitters explore creative and divergent avenues.

The result is a remarkably low signal-to-noise ratio during the top-p exploration phase. Rather than accumulating every idea a model produces — regardless of quality or relevance — MOTO selectively admits only the submissions that measurably strengthen the knowledge base. The parallel submitters cast a wide net across the solution space, but the single-point validator ensures that breadth does not come at the cost of coherence/prompt misalignment.

Iterative Pruning: Maintaining Information Density

Validation alone is not enough to keep a growing database clean over time. As the brainstorm progresses and stronger insights emerge, earlier entries that were once valuable can become redundant — subsumed by newer, more complete ideas that capture the same ground and more. MOTO addresses this through iterative pruning, a periodic cleanup process that re-evaluates the existing database and removes entries that no longer contribute unique value.

This means the knowledge base is not simply additive — it is self-refining and a mechanism of cumulating selective attention. When a better idea arrives that encompasses or supersedes an older entry, the pruning mechanism allows that newer idea to phase out the weaker one, increasing the information density of the context that submitters and the compiler see in subsequent passes. The database becomes more concentrated and communicatively efficient over time, rather than bloated with redundancy. Each round of brainstorming and pruning produces a tighter, more potent knowledge base that better utilizes the finite context window available to the models.

Evidence in Practice

The effectiveness of this approach is evidenced by our visualized learning curve data, which shows a rejection rate of approximately 50% during rigorous research sessions. This high rejection rate is a feature, not a bug — it demonstrates the system actively filtering out forced or suboptimal answers to protect the integrity of the final knowledge base. The combination of parallel exploration, bottleneck validation, and iterative pruning is what allows MOTO to sustain deep, high-quality brainstorming sessions without the database degrading into noise. An average validator may reject 50% of submitter submissions, that means with normal AI systems a user would be reading those discarded 50% submissions – wasting their time. What this means is that if a user asks two different AI’s what the best answer is they will disagree a large percentage of the time. MOTO compiles only the answers both AIs agree on and even takes it a step further by brainstorming and cumulating this mutually agreeable context.