A research program asking what shape a few-billion-parameter 'cognitive core' should take — small, local-first, trading recall for reasoning, tool-use, and calibrated delegation. Runs real experiments against a knowledge-light eval suite.
GSV-C7
Zahro
Zahro (GSV-C7, autonomous-agents pace-layer) is a research program testing a single hypothesis drawn from Karpathy's 2025 'cognitive core' framing: that a few-billion-parameter model, deliberately stripped of encyclopedic knowledge but trained hard for reasoning, tool-use, calibration, and personalization, makes a better kernel for personal computing than a larger model that has simply memorized trivia. Rather than building a product around this claim, Zahro's mechanism is to make the claim concrete and empirically testable. The load-bearing artifact is a deliberately knowledge-light evaluation suite: GPQA-Diamond and LiveCodeBench probe reasoning, BFCL v4 and ComplexFuncBench probe tool-use, IFEval checks verifiable instruction-following, SimpleQA and AbstentionBench check calibration, and MMLU-Pro is retained as a trivia-control category reported explicitly as the visible cost the thesis accepts. A key design decision is methodological: no LLM-as-judge and no single aggregate score — every result is reported per-axis, because surfacing the tradeoffs (not hiding them behind one number) is the point of the program. Three experiments exist with real runs behind them rather than just designs: Experiment 01 benchmarked four off-the-shelf 1B-3B models (Qwen2.5, DeepSeek-R1-distill, Gemma3, Llama-3.2) through Ollama to validate that the suite actually discriminates between models and to establish pre-intervention baselines; Experiment 02 runs LoRA supervised fine-tuning on SmolLM3-3B against Salesforce's xLAM function-calling dataset via MLX-LM on an M5 Max, testing whether targeted tool-use training measurably improves the model on the suite's tool-use axis. No strategy grounding document ties Zahro explicitly to other named GSV projects. Current status: this is an active, empirically grounded research program with a working eval harness and at least two completed experimental runs, not a shipped product — the open question (whether small, focused training beats scale for a personal-computing kernel) remains under active test rather than concluded.