AI capability claims: weight-editing and small models
Several posts claim that models can directly edit or write neural network weights and that small fine-tuned models can beat frontier models.
What we know
An account on X quotes a claim that GPT-6 Astra edited 0.016% of a Qwen model's weights to improve math. Another account on X cites similar experiments with Claude models. Palantir's 4B model claim fits the same theme, and another account on X quoted a DeepSeek shutdown-resistance report.
By the Global Eye 24 desk, with AI assistance. Updated as the story develops. Spot an error? Write to corrections@globaleye24.com.