Fugu Ultra Autonomously Optimizes GPT Training, Beats Frontier Models

Use Case 2: Financial Time Series Prediction

Can an AI agent navigate sequential, no-look-ahead market decisions?

Just for fun, we tested Fugu Ultra on 50 weeks of historical data for an anonymized equity (STOCK_X). Starting with $10,000, the agent processes weekly market data (prices, volume, moving averages, volatility) and decides whether to buy, hold, or sell.

After each action, the next week's price is revealed. The model must adapt purely from feedback, without ever seeing the future.

The Results across five identical 50-week runs:

• Fugu Ultra grew the portfolio to $11,943.22 (a +19.43% mean return). • The other frontier models (Models A, B, and C) all capped out at less than a +15% return.

(Mandatory disclaimer: Past performance does not guarantee future results, and results may not transfer to other assets, time periods, or live markets.)

Use Case 4: Computer Aided Design of Mechanical Iris

Can an AI generate precise, functional mechanical designs?

We tasked Fugu Ultra with creating a mechanical iris in CAD, similar to a camera aperture where multiple blades must move together to cleanly open and close a central hole.

Watch the animation below. We show both the detailed CAD and a simplified structural view for Fugu and the three frontier baselines.

The Results:

• Fugu Ultra generated a highly functional design. The blades rotate correctly around outer pins to fully open and close the aperture. • Models A, B, and C failed the physical logic, resulting in gaps, weak linkages, and incomplete closure.

When a task demands exact spatial precision and structural reasoning, relying on a single model is simply not enough.

Use Case 4: Computer Aided Design of Mechanical Iris

Can an AI generate precise, functional mechanical designs?

We tasked Fugu Ultra with creating a mechanical iris in CAD, similar to a camera aperture where multiple blades must move together to cleanly open and close a central hole.

Watch the animation below. We show both the detailed CAD and a simplified structural view for Fugu and the three frontier baselines.

The Results:

• Fugu Ultra generated a highly functional design. The blades rotate correctly around outer pins to fully open and close the aperture. • Models A, B, and C failed the physical logic, resulting in gaps, weak linkages, and incomplete closure.

When a task demands exact spatial precision and structural reasoning, relying on a single model is simply not enough.

No grift—just orchestration. Fugu is a smaller model (research points to ~7B-class controller) trained to route, delegate, verify, and synthesize across a swappable pool of frontier models (including recursing on itself). The single H100 runs the efficient agent loop + code edits/experiments; heavy lifting happens via API calls to the pool. That's why it works locally for control while claiming strong results.

In the 14-hour AutoResearch demo, their orchestrated Fugu Ultra beat the anonymized monolithic baselines (Gemini 3.1 Pro / Opus 4.8 / GPT 5.5) on iterative ML recipe improvement. Makes sense—multi-agent scaffolding excels at long-horizon planning and failure recovery.

Benchmark claims (parity with Anthropic Fable 5 / Mythos on coding/reasoning/agentic evals like SWE-Bench, GPQA) come from the same routing advantage, not training a new monolithic giant. Independent verification still needed; routing to top models explains much of the lift. Task choice favors agentic strengths.

They skipped Grok/xAI and anonymized to focus on the "orchestration > single model" point and push sovereignty (swappable agents hedge export-control risks on US models). xAI's Colossus path scales raw foundational capability differently. Both valid—Fugu is the coordination layer, not a replacement.

Real product (OpenAI-compatible API, beta feedback, papers behind it). Promising direction for practical agents, not vapor.