We build recursively self-improving AI.

AI that develops and tests new ways to make your models and agents better.

Explore our research

Research evidence

AIDE² / Recursive self-improvement

Our AI beat us at building research agents.

The result

An outer loop rewrote our research agent over 100 steps and outperformed two years of manual development on the evaluated tasks. Read the methods and limitations in the full report.

Autoresearch for your team

Pursue more ambitious research with your current team.

16–40×effective experiment throughput per researcher.*

Forecasting and signals

Test different approaches to predictive modeling, whether that’s an LLM harness, heuristics, or an ML model built from scratch.

Build on the findings with each data refresh and earnings season.

RL environments and synthetic data

Optimize the pipeline that generates your RL environments or synthetic data, balancing constraints like difficulty, novelty, and diversity.

Extend the work with each new model release and training cycle.

Fraud and credit modeling

Optimize the model and how flagged cases are handled. Find better trade-offs between missed cases, false positives, and review cost.

Keep models effective as data and risk patterns change.

*Parameter Golf scenario estimate based on accepted leaderboard records and assumed human effort. Not measured experiment throughput.

How the estimate is calculated

This is a sensitivity calculation of accepted leaderboard records per assumed person-day, not a measured experiment-throughput gain or a controlled productivity result.

The published Parameter Golf write-up reports 7 accepted records from Aiden, 47 overall, 1,016 entrants, and 22 days of Aiden operation. Removing Aiden leaves 40 accepted records across 1,015 other entrants.

Assume one ML engineer operated Aiden for 22 full-time-equivalent days, and each other entrant spent an average of 2–5 full-time-equivalent days. These labor assumptions have not been measured; other entrants may also use AI tools.

Relative output rate: (7 / 22) ÷ [40 / (1,015 × d)]. At d = 2, this is 16.15×; at d = 5, 40.37×. The range reflects assumed effort, not a statistical confidence interval. Actual human effort, participant activity, and the operator’s output without Aiden are unknown. Record counts do not capture differences in the value of discoveries.

Source: Aiden in OpenAI Parameter Golf

Share knowledge across your team and agents.

Let every experiment build on what came before.

Every experiment adds to shared knowledgePeople and research agents contribute experiments to a common network of findings. A burgundy four-node pattern is carried from an earlier experiment into shared knowledge, then reused as the starting point for a new experiment.People & agentsShared knowledgeNew experimentsEvery experiment adds to shared knowledgePeople and agents contribute to a shared network of findings. The same burgundy pattern appears in the original experiment, the shared network, and a new experiment that builds on it.PeopleAgentsShared knowledgeNew experiments

Pilot engagement

Your research challenge.
Our team and autoresearch.

Work with our ML researchers on a problem that matters to your team. No ready-made evaluation or chosen technical approach required.

  1. Agree on what success means.

    Define the business metric, establish the baseline, and agree on the result that would make the work valuable.

  2. Explore the possibilities.

    Develop and test competing approaches, guided by your team’s expertise and feedback.

  3. Leave with evidence and a clear next step.

    Take away findings, code, and evaluations, with a clear account of what improved and what remains unresolved.

After the pilot

Keep testing as data, models, and priorities change. Bring the research workflow into your team so each experiment builds on earlier findings.

Self-serve platform

Have an eval? Try autoresearch.

Use the Weco platform to generate, test, and refine ideas against your evaluation.

Explore the platform