Researchers demonstrate "Weight-Space Programming," where the frontier LLM GPT-6 Astra natively reads and writes neural network weights without gradient descent. Astra identified which of four Atari games a PPO-trained policy network was trained on solely from raw weights, and distinguished diffusion models generating circles, triangles, or squares by inspecting output-layer geometry. On 7.2M-parameter language models with input/output layers removed, Astra correctly identified training languages (Python, romaji, English) and precisely located induction heads matching ground truth. For weight writing, Astra fixed a Kinetix car policy stuck in a local minimum by editing 3.5% of weights to implement a backup-then-accelerate strategy, solving all 1,024 evaluation episodes. In zero-shot network synthesis for MinAtar games, Astra-generated weights outperformed PPO trained for 100M steps (Breakout: 517.4 vs 15.8; Seaquest: 53.1 vs 5.6; Asterix: 53.2 vs 25.8) and approached 500M-step PPO. At 1.5M parameters (Go policy), Astra wrote a 336-line Python program constructing weights layer-by-layer; with 50 forward passes (128 rollouts each), it beat equal-sized PPO trained on 20M games. On Qwen3.5-0.8B, Astra diagnosed reasoning loops caused by a specific direction in MLP layers 15 and 19, subtracting it from 131,072 weights (0.016% of parameters). With 50 forward passes (batch 64), MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%, while token-limit hits dropped from 93% to 33%. On Gemma 3 270M (8.5% loop rate), edits yielded no significant gains, confirming weight editing removes obstacles but cannot add missing capabilities.