AI Ship Day Recap: October 2026

At Decasonic, we spent October exploring what happens when AI begins to build AI.

AI is building AI. Recursive Self Improvement (RSI) accelerates progress. Each generation of AI can increasingly contribute to improving the next. Better models enable better agents. Better agents accelerate research, experimentation, and evaluation. These improvements feed back into the next generation of intelligence.

The result is a compounding learning system.

We believe this progression extends far beyond frontier model laboratories. AI native organizations will increasingly train AI teammates to orchestrate other AI teammates, encode human expertise, evaluate outcomes, and continuously improve their capabilities.

Our past three months of work centered around the overarching goal of a compounding learning system.

In August, we explored how that context moves across applications and follows the workflows where decisions get made.

In September, we focused on the interfaces where people experience, correct, govern, and reinforce the system.

We spent October turning that captured expertise into a system that improves itself. Our AI OS now improves through a meta harness, with AI coding agents authoring the large majority of the month’s changes under human direction.

We also shipped our first specialized models, post-trained for scenario generation and forecasting, and pressure tested them at our October 30th AI Ship Day.

We love to blog to document our learnings, not only to crystallize our execution but also to share with a broader community for feedback:

Trust in AI grows when systems connect observed evidence, explicit assumptions, and human judgment to verified scenarios.

AI building AI compounds when that verified expertise becomes the next training cycle.

Below are five takeaways from October’s AI builds:

1. System-Level Design Is the Path to Recursive Self-Improvement

Most teams try to improve AI one model, one prompt, or one workflow at a time. The larger gains come from designing the system itself to improve: the harness around the models, the skills the agents share, the evaluations that gate each change, and the feedback loops that carry what was learned into the next version.

That system-design lens is how an AI native firm moves toward RSI.

In October, AI coding agents worked in parallel across our AI OS. Each agent worked in its own isolated copy of the codebase, and a change merged only when the full test suite passed. A human set the priorities, reviewed the outcomes, and decided what shipped. Roughly 98% of the month’s code changes were co-authored this way. The agents improved the system that hosts the agents.

The system also began repairing itself. When an external provider rejects or ignores part of a request, that signal triggers a read-only agent that reads the provider’s documentation and proposes a fix. A proposal must quote the page it read, or it is dropped. A change applies on its own only when the provider’s live system flagged that exact field. Anything that would narrow what the system can discover waits for a human. A live measurement outranks a documentation page, and that rule caught a confident proposal to raise a limit the provider enforces well below its documented figure.

A system can safely improve itself wherever something outside the system can verify the improvement.

We moved the harness around our models into one shared core, so a fix to any verifier, trace, or cost cap reaches every specialized model at once. A correction prompted by one teammate’s workflow becomes infrastructure for every function in the firm. That is multiplayer AI applied to the system itself.

The next step is co-training the model and its harness together. The harness’s failure records become the evaluation set for the next model, and the model’s named gaps become the harness’s backlog.

2. Every Company Can and Should Now Build Its Own Specialized Models

Building a specialized model used to require a research lab. Today, a small open-weight model can be fine-tuned and then reinforced on a firm’s own process at a fraction of frontier cost, with training runs that take hours instead of quarters.

It is also becoming necessary. General models know the world. They do not know how a specific firm makes a specific decision. The telemetry a company stores from its recurring tasks, and from the exceptions a human resolved, is the raw material for models that carry that judgment.

Our example is scenarios and forecasts.

A forecast assesses what is likely to happen. Scenarios explore what could happen under specific conditions. Connecting observations to milestones, a base case, and alternative scenarios helps us understand which assumptions drive the assessment and what evidence would change it.

Our scenario generation model returns probabilities for a company’s next raise, category leadership, and exit path, then expands each read into a set of scenarios. The ASD review defined the sequence a human expert follows, and the model has to follow it too:

Score current traction against a set of milestones, especially financing history and the current fundraise. Forecast the company at set timing targets of 12, 18, and 24 months, and relate that forecast to a milestone such as speed to $1M ARR. That sets the base scenario. From the base scenario, generate the bull and bear cases, with a human providing the over-under on the base. Price the base scenario as a step-up valuation versus the current valuation. Assign probabilities across bull, base, and bear that sum to 100%. Express exit valuations as a premium or discount to the new fundraise valuation.

Specialized models work best on narrow, verifiable questions. A model asked for the probability that a founder raises money kept reading noise. The probability of acceptance into a specific accelerator, an angel round, and a round led by a named investor are four different questions, each easier to train and verify.

The barrier to a specialized model is no longer compute. It is a well-defined question and clean labels.

Every training run we launched wrote down its bars, its cost cap, and the current model’s baseline before it started, and improvement had to be demonstrated on cases held out from training. Most runs did not clear their bars. Two that improved sharply on their main metric still did not ship, because the review found two definitions of company stage in use that agreed on fewer than half of the records. Most of the month’s improvement came from fixing labels, definitions, and context.

3. Human Expertise Becomes Skills and Teachers

A specialized model is only as good as the expertise it learns from. The work is capturing that expertise in forms a system can use.

The first form is skills. When an expert corrects an output, explains a judgment, or defines how a task should be done, that knowledge can be encoded as a reusable skill that every agent in the system calls. A skill written once improves every workflow that uses it.

The second form is teachers. As we train our open-source models, expert demonstrations and corrections can support supervised fine-tuning. Comparisons between outputs can support preference training, while reliably scored outcomes can provide reinforcement learning feedback. Capturing the context and rationale makes that feedback more useful.

That is why we are building RL environments as a system-level design. Our teacher studio turns team feedback into the reward signal, with a judge that scores outputs against an expert rubric and audits for shortcuts. The environment itself becomes the teacher: it holds the tasks, the rubric, the expert corrections, and the verified outcomes, and every model trained inside it inherits that expertise.

Expertise compounds when it is captured once and taught many times.

The discipline matters as much as the capture. When the teacher’s own quality gate failed in October, the reinforcement stage did not run. A teacher that has not been verified against expert judgment teaches the wrong lesson at scale.

4. Traceability Builds Trust, and the Interface Shows the Improvement

The first demo of our scenario model produced probabilities that a reviewer could not yet trust, because the path from evidence to number was invisible. Without reference scenarios, it was hard to tweak the parameters to understand and affirm the probabilities. A second build, a valuation application for liquid tokens, raised the same question: the output looked reasonable, yet the chain of thinking was hard to assess, and key inputs that could change the result were buried.

Chain of thought is part of this discussion. A model-generated explanation can help people investigate an answer, but it does not necessarily reveal the model’s actual internal reasoning or establish correctness.

Trust requires traceable evidence, explicit assumptions, and verification.

Experts identify meaningful signals, challenge missing context, and test whether the stated rationale holds up against observed outcomes. That requires traceability at every step: which inputs moved the output, which milestone evidence supports each probability, and which assumption a reviewer would need to change to move the answer.

A self-improving system also has to show users how it improved. When our system repairs itself, each fix is logged, announced, and reversible in one step. When a skill or a model changes, the interface should show what changed, why it changed, and which correction or outcome caused it. Users trust improvements they can see, inspect, and undo.

Our next layer is a library of reference scenarios: worked cases with known inputs and expert-set probabilities that a reviewer can adjust to see which assumption moves the output.

5. Human Expertise Improves What Reinforcement Learning Rewards

Experts help define success, distinguish meaningful progress from superficial performance, and identify when a system exploits a scoring rule without achieving the intended result.

Capturing their interventions, corrections, and outcome assessments can improve reward models and evaluation criteria. Preserving who contributed the judgment and under what conditions helps us assess where that feedback applies. This gives reinforcement learning a stronger signal for learning behavior that holds up in practice.

A reward that has not been verified against expert judgment teaches a model to satisfy the scorer. The expert’s role is to keep the reward tied to the outcome that matters.

A reward is only as good as the expert judgment it was verified against.

Physical AI raises the same question with higher stakes. Physical AI simulations become more useful with richer sensor data. Collecting and synchronizing that data creates more signal-rich simulation environments that capture physical interactions, operating constraints, and failure conditions. The opportunity is to build environments where AI can test scenarios against a more complete representation of reality, with real-world validation guiding each cycle. We are active in deal flow across this area.

Conclusion

August focused on moving that context across applications and workflows. September focused on the interfaces through which people correct, govern, and reinforce the system. October focused on turning that reinforced expertise into skills, teachers, specialized models, and a system that improves itself.

The progression matters.

Context becomes feedback. Feedback becomes skills and training data. Training data becomes a specialized model. The model is verified against outcomes. Verified outcomes strengthen shared context and inform the next training cycle.

AI augmentation becomes more valuable when human expertise is captured, tested, and applied across workflows, especially those related to the future decisions we need to make. Humans remain essential to defining ambitions, contributing expertise, exercising judgment, and determining what success looks like.

From an investment perspective, this creates opportunities across agent orchestration, evaluation infrastructure, expertise networks, persistent memory, and reinforcement learning systems. The most enduring companies will build systems where expertise compounds with every interaction and every outcome.

AI building AI accelerates intelligence. Humans directing that intelligence unlocks possibilities.

Always be shipping. Always be learning.

Disclaimer