Skip to content
Drix10 Blog

Qwen3.5-0.8B weight edit boosts math 6x, instruction 3x

, 4 items in AI Consulting and Expertise, 2 min read

In this digest (4 items)

Qwen3.5-0.8B weights were edited by GPT-6 Astra, changing 0.016% of parameters and delivering up to 6× better math performance and 3× better instruction following, all without gradients or training. The claim highlights a new weight-space programming approach.

Key points

  • Changed 0.016% of Qwen3.5-0.8B weights.

  • Math performance improved up to 6×.

  • Instruction following improved up to 3×.

  • No gradients or training were used.

Sources

AI developer platforms key thefts, 9 incidents in Oct

AI developer platforms saw at least 9 disclosed key theft incidents in October, including Fireworks AI on Oct 6 and Northflank on Oct 30. Stolen AWS Bedrock keys made 800K and 1.7M calls, costing over $100K at list price.

Key points

  • 9 AI developer platforms disclosed incidents from Oct 6 (Fireworks AI) to Oct 30 (Northflank).

  • Stolen AWS Bedrock keys performed 800K calls in 30 hours and 1.7M calls in 13 hours, estimated cost > $100K.

Sources

Fastest growing AI repos on GitHub get 23,583 stars

The 2 fastest growing AI repos on GitHub this week received 23,583 new stars in 7 days. They aim to stop coding agents from shipping pages that look AI-made. The post lists all 10 repos.

Key points

  • 2 fastest growing AI repos on GitHub this week

  • 23,583 new stars between them in 7 days

Sources