AI Development Suite "Codens" — 65% of PRs Are AI-Generated, Median Time-to-Merge of 2 Minutes

The team behind "Codens", an AI development suite built by Corevice, has opened up their own development process. According to measurements taken via the GitHub API, 65% of merged PRs are AI-generated. On top of that, the median time from PR creation to merge is **2 minutes**, and 93% of PRs are merged within an hour.

You might think "there's no way development moves that fast when you hand it over to AI" — but these numbers reportedly aren't the work of the model alone. The post walks through the design and operational practices that make it possible. A weekly-updated dashboard is also public at [codens.ai/stats](https://www.codens.ai/stats/), making this a useful reference for anyone interested in making AI-driven development more efficient.

llama.cpp b9966 — Fixes Regex Recompilation in Decode Threads

llama.cpp build b9966 has been released, featuring a performance fix for `-sm tensor` mode.

The problem: decode threads were **rebuilding 29 regex patterns every single time, for each tensor and each token**. The fix changes this to cache them, dramatically cutting wasted CPU while leaving behavior unchanged.

If you run `-sm tensor` in production, picking up this update is recommended.

Release: [llama.cpp b9966 (GitHub)](https://github.com/ggml-org/llama.cpp/releases)

Measuring PCIe Transfers on a Dual-GPU Setup — Tensor Parallel vs Pipeline Parallel

The LocalLLaMA community has shared measured PCIe bandwidth data from a dual-GPU setup with an RTX 3090 + Titan RTX (24GB each).

Test results using Qwen3.6-27B (Q4_K_XL, 180k context):

| Metric | Tensor parallel | Pipeline parallel |

|---|---|---|

| Decode speed | 47 t/s | 30 t/s |

| Prefill speed | 650 t/s | 1250 t/s |

Tensor parallel came out faster on decode but slower on prefill, with pipeline parallel showing the exact opposite. Interestingly, prefill degraded under tensor parallelism even though the PCIe 3.0 x16 bandwidth was never saturated.

From this data, the poster concluded that multi-GPU builds using PCIe 1x risers (the mining kind) won't become a bottleneck either. A useful data point for anyone considering a multi-GPU setup.

Protecting Production DBs Behind MCP Servers — Designing a "Documented API Boundary"

As more setups let AI agents touch internal systems through MCP (Model Context Protocol), the question "is it safe to grant SELECT privileges on the database?" is drawing attention.

An article published on Qiita proposes a design pattern that routes access through a **documented API boundary** rather than letting AI agents hit the database directly. Instead of MCP tools executing SQL directly, they access data through an API — an approach that draws a clear boundary for permission management and auditing.

A practical security design reference for adopting MCP in enterprise environments.

AI Decoding the Cell — Silvana Konermann's TED Talk

An effort to decode the intricate machinery of cells using AI was presented on TED. Silvana Konermann talks about the astonishing complexity of human cells and how AI is taking it on.

The intersection of biology and AI has been a particularly active field in recent years, and this talk is a great entry point for understanding the big picture.

Watch: [The Human Cell Is Wildly Complex. Can AI Decode It? (TED)](https://www.youtube.com/watch?v=Xr9VqRawjAU)

---