Godot Bans AI-Generated Code, Reports of Claude Performance Degradation, Ollama Speeds Up Gemma 4 by ~90% — AI News Digest

Godot Bans AI-Generated Code in Principle — Reviewer Burnout Reaches Its Limit

Godot, the open-source game engine, has announced that AI-generated code is now prohibited in principle for contributions. While AI-generated pull requests from new contributors have surged, the number of reviewers examining them has remained unchanged, and the burden has reportedly reached its limit. Reviewers have voiced sentiments like "I don't want to talk to a machine," and using AI-generated text in human-to-human communication is also covered by the ban.

This move could ripple across the entire OSS community. With the spread of AI coding tools, growing maintainer burden has become an issue everywhere, and Godot's decision can be seen as a symbolic step. How to strike a balance between quality control and contributions will likely be the key to running these communities going forward.

Two AI Companies Report Findings on Claude Fable 5's Performance Degradation

Two US-based AI companies have each reported the results of their investigations into whether Claude Fable 5's performance changed around its deprecation. After running performance evaluations spanning the model's deprecation, performance differences were reportedly observed on some tasks.

The problem of LLM performance fluctuating over time or across version updates is known as "model drift," and this case appears to be one such example. For companies using LLMs in production environments, it has once again underscored the importance of continuous quality monitoring.

Ollama 0.31.1 — Gemma 4 Runs About 90% Faster on Apple Silicon

Version 0.31.1, the latest release of Ollama, the local LLM runtime tool, dramatically improves performance on Apple Silicon. Gemma 4 model execution speed is up roughly 90% compared to previous releases, and the gains require no configuration changes.

The speedup is reportedly driven by improvements to MTP (Multi-Token Prediction) technology. When running LLMs locally, waiting time affects the experience even more than model intelligence, so faster inference is especially important for long-running agent workloads. This is a major tailwind for developers leveraging local AI on the Mac.

audio.cpp — A GGML-Native Open-Source Audio AI Framework

A C++/GGML-based audio AI framework called audio.cpp has been announced on Reddit's LocalLLaMA community. It covers music generation, sound effect generation, and even sound source separation within a single native framework, with models such as ACE-Step, Stable Audio, HeartMuLa, RoFormer, and HTDemucs implemented in C++.

Notably, generating 600 seconds of music with ACE-Step Turbo completes in about 60 seconds (roughly 10x faster than real time). It is also about 30% faster than the Python implementation, demonstrating the advantage of a native C++ implementation. 21 of the 28 models have been ported so far, and further progress is expected.

BPE Tokenization Breaks LLM Safety Alignment — Demonstrating a Structural Vulnerability

A paper published on arXiv, "Breaking Safety at the Token Boundary," demonstrates the mechanism by which BPE tokenization creates structural gaps in LLM safety alignment. Because safety-critical words get split into subwords, safety filters can be bypassed through character-level perturbations.

Testing across five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, and Mistral-7B) flipped 80–100% of refusal triggers, with up to 65% of those producing actually harmful outputs. The study also found that not a single fragmented prompt appeared in the three public alignment datasets surveyed (30,000 examples in total). It is research that surfaces a critical challenge for the robustness of alignment.

Can In-Context Retrieval Actually Work at the Million-Token Scale?

Also published on arXiv, "Can Language Models Actually Retrieve In-Context?" provides the first systematic examination of the scalability of in-context retrieval by language models. The authors developed BlockSearch, a 0.6B-parameter retrieval model that achieves length generalization up to 10x the scale it was trained on, though performance collapses under more extreme extrapolation.

The cause is the "attention dilution effect" — as the corpus grows, irrelevant documents dominate the softmax denominator, lowering the normalized weight assigned to the correct document. The paper mitigates this problem with length-aware attention adjustments and document-level sparse attention, achieving performance comparable to Dense Retrieval at the million-token scale. It is a significant contribution that identifies a new challenge brought about by ever-longer contexts.

---