LLMs can directly read and edit neural network weights, enabling rapid model steering
A new technique called weight-space programming allows models to inspect and modify other models' weights, reducing reliance on manual optimization loops. DeepSeek also open-sources a 'super-agent' harness prototype with quantified performance gains.
1 source · cross-referenced
- A technique called weight-space programming enables LLMs to directly read and edit neural network weights, reducing the need for manual optimization loops.
- DeepSeek open-sources a 'super-agent' harness prototype, reporting a reduction in p95 completion time from 29.5 to 15.7 minutes compared to a single session using subagents.
- The GDPx benchmark evaluates 206 agentic tasks across 11 domains, with Opus 5.5 leading at 12% and Gemini 4 Carbon close behind at 10.9%.
- The Epoch Capabilities Index places DeepSeek V4.1F-1017 at 159, equal to GPT 5.5, and 10 points above V4-Pro.
A new technique called weight-space programming enables large language models to directly read and edit the weights of other neural networks, bypassing traditional optimization pipelines. Researchers demonstrated that a sufficiently capable model can inspect compiled model weights—analogous to reading executable binaries—and perform tasks such as fixing bugs in a single step or improving performance over 50 steps. This capability was theoretically anticipated but has now been empirically demonstrated, with independent reproductions already reported.
The approach is framed as a form of mechanistic interpretability at scale, though current implementations are limited to simpler tasks like games and reinforcement learning policies. Despite these constraints, the technique’s emergence marks a step toward recursive self-improvement (RSI) with reduced human intervention, as it cuts out the loop of writing PyTorch/JAX code, running compute, and analyzing loss curves.
DeepSeek open-sourced a prototype agent harness, dubbed 'super-agent,' designed to coordinate long-lived router sessions with isolated worker agents. The architecture triggers tasks via GitHub events and reports a reduction in p95 completion time from 29.5 to 15.7 minutes compared to a single session using subagents in internal tests. The design emphasizes cross-task continuity versus throughput, and the team acknowledges open questions around long-term memory, state separation, and security.
A standardized agentic benchmark, GDPx, evaluates 206 tasks across 11 economically valuable domains, including chip design, software engineering, finance, and biology. The benchmark’s inaugural results show Opus 5.5 leading with 12%, followed by Gemini 4 Carbon at 10.9% at roughly 4x lower cost. Within-family comparisons indicate Sol 6.1 and Astra score similarly, while Carbon outperforms Argon across benchmarks.
The Epoch Capabilities Index places DeepSeek V4.1F-1017 at 159, equal to GPT 5.5, and 10 points above the V4-Pro variant. The index also shows Astra sitting roughly 10 points higher than V4.1F-1017, reflecting ongoing competitive dynamics in frontier model performance.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74