South Africa Retracts National AI Policy Over AI-Generated Fake Sources
It has come to light that a national AI policy document released by the South African government contained fabricated sources generated by AI, forcing the policy to be withdrawn. Reported by Reuters, the case is drawing attention as a concrete example of AI hallucination (plausible-sounding falsehoods) having real-world consequences at the policy level.
While AI tools help streamline document creation, failing to fact-check their outputs can cause serious problems even in public settings such as government documents. The incident has once again highlighted the need to build fact-checking mechanisms into any AI workflow.
Preliminary Findings on AI Automation Based on Worker Evaluations Released
A paper titled "Preliminary Findings on AI Automation from Worker Evaluations" has been published on arXiv. It presents preliminary research findings investigating what impact AI automation is actually having in the workplace, based on evaluations by the workers themselves.
While productivity gains from AI adoption are reported, worker dissatisfaction and concerns are also documented, making this an interesting attempt to empirically capture how working conditions are changing as automation advances. See the full paper for the detailed analysis.
UK Local Authorities Speed Up Urban Planning Decisions with Google AI
Several local authorities in England are piloting Google's AI tools for planning permission decisions. According to the Financial Times, having AI process the enormous volume of planning applications is expected to dramatically accelerate the approval process.
The use of AI in administrative procedures is advancing in many countries, but questions remain over how far AI can be incorporated into discretionary decisions such as planning judgments, and how transparency and accountability can be ensured. The effort is still at the trial stage, and full-scale deployment will require institutional scrutiny.
TurboQuant Implementation Results Diverge from the Paper, Sparking Heated Debate on KV Cache Quantization
In Reddit's LocalLLaMA community, a user who independently implemented the paper "TurboQuant" (arXiv:2504.19874) reported being unable to reproduce the results claimed in the paper, sparking debate.
According to the implementer's report, the MSE variant works as expected, but with the PROD variant, the implementation achieved only 95.8% correlation against the 99%+ correlation claimed in the paper. Moreover, even at this correlation level, attention quality degrades significantly, with top-1 accuracy dropping to around 67%.
The observation that "correlation ≠ ranking preservation" suggests that correlation coefficients alone may be insufficient as an evaluation metric for KV cache quantization. It has also been reported that implementation details such as variance scaling and bit packing significantly affect the results, meaning challenges remain for making quantization methods practical.
An Effort to Evaluate Public Repositories That Are "Friendly" to AI Coding Agents
A project titled "Which public repos are friendliest to an AI coding agent?" was featured on Hacker News. It is a systematic attempt to evaluate how effectively AI coding agents (such as Claude Code and Copilot) can contribute to open source repositories.
Project structure, documentation, and test coverage are key factors in making it easy for AI agents to understand and modify a codebase. This initiative is drawing attention as a step toward clarifying best practices for AI-assisted development.