From the AI news gathered on September 15, here are six stories: a new turn in the U.S. political debate over AI safety, a major release for the local LLM community, bringing AI inference to the JVM, a hands-on account of speech recognition hardened in production, and a story close to home from Japan.
Trump Goes So Far as to Call AI Safety Concerns a "Hoax" — Rejects Calls for Stronger Safeguards
According to the BBC, Trump has responded to calls for stronger security around AI by describing the concerns themselves as a "hoax," once again rejecting demands for tougher safeguards.
This is a continuing story that has unfolded over the past several days. In early September he brushed off requests from Amodei and others to slow development, and more recently he criticized Anthropic's Dario Amodei by name. What's new this time is the shift in wording: he has moved a step beyond opposing "calls to slow down" to asserting that "the concerns themselves are a hoax."
The industry, meanwhile, keeps stacking up voluntary safety measures. Amodei has committed to stationing third-party evaluators in-house, and an evaluation standard called "AEF-1," which OpenAI and xAI have also signed, has emerged. As the temperature gap between politics and industry widens, self-regulatory frameworks could become the main battleground for substantive safety measures.
Dynamic Quantization Method "Voodoo Quant" Released Under MIT License — Optimization Open to the Community
On Reddit's LocalLLaMA community, the complete toolset for the dynamic quantization method "Voodoo Dynamic Quant" — which made a splash about two months ago for delivering SOTA-level quality at the most aggressive quantization levels on small Qwen-family GGUF models — has been released under the MIT license.
Dynamic quantization assigns a different quantization level to each tensor (a bundle of weights), as used in the GGUF format. Where the conventional mainstream approach of static analysis picks levels based on summary statistics, Voodoo Quant runs every quantization level of the model simultaneously and learns just one scalar gate per tensor and per level. Gradient descent directly optimizes "which combination of levels minimizes loss for a target file size," using KL divergence against a BF16 reference model as the loss function. According to the author, this is the first method that selects per-tensor quantization levels via backpropagation.
In comparisons against Unsloth's "Dynamic 3.0," UD 3.0 reportedly wins at high-to-mid quantization levels, while Voodoo Quant comes out ahead on the more aggressive low-bit end. The big difference is that UD 3.0's method is closed, whereas VQ has released both the tools and the methodology. For anyone running local LLMs on low-VRAM hardware, this release is a tailwind.
Show HN: jinfer, an Inference Engine for the JVM — AI in "a Single jar," No Python Required
On Hacker News, an AI inference engine for the JVM called "jinfer" was announced. It supports chat, vision, audio, embeddings, reranking, and TTS, and needs no Python runtime, no ONNX, no Docker containers, and no sidecar processes — the tagline is "AI in a jar."
The foundations were also rebuilt for the JVM. It consists of "toknroll," a pure-Java tokenizer with zero dependencies; a GGUF/safetensors module that reads and writes llama.cpp and Hugging Face model formats; "jam," a quantized matmul kernel claimed to rival llama.cpp on CPU; and "jota," a Tensor API targeting Java, C, CUDA, HIP, Metal, OpenCL, and Mojo. Along with Spring AI and LangChain4j integrations, GraalVM Native Image support reportedly lets you build self-contained binaries that start in milliseconds.
For now it is an initial, CPU-only release, and the author is asking for feedback on the API design. For organizations with enterprise Java estates, having an option that keeps LLM inference inside their existing JVM operational stack is no small thing.
GVA Compresses the KV Cache by Grouping Values — Keys Are Reconstructed On Demand
On Hugging Face Daily Papers, "Grouped Value Attention" (GVA) was proposed to improve the memory efficiency of the KV cache. The approach stores no content keys: it keeps only grouped values and reconstructs keys on demand through learned linear mappings.
Shrinking the KV cache, which tends to hit a VRAM wall during long-form generation, is a topic directly tied to both local operation and long-context processing. The detailed evaluation results are best checked in the paper itself, but as a design that reaches into the very core of the attention mechanism, it deserves attention.
A Greek–English ASR That Cleared 9 Quality Gates — A Field Record: "No Single Model Could Get There"
Also on Hugging Face Daily Papers, the authors themselves shared a record of taking "Sophea ASR," a Greek–English bilingual speech recognition model, to production. They set nine quality gates — Greek WER, English WER, language identification, hallucination on silence, and more — but no single model could pass them all.
Across two architectures and 23 training runs, the best result was 7 of the 9 gates. Blocked by trade-offs — improving Greek noise robustness would break English language identification, for example — they finally reached 9/9 with a three-model ROVER ensemble, which also improved WER on overlapping speech from 53% to 38%.
Even more interesting is the measurement side. After calibrating the audio filter with in-domain anchor utterances, discarded Greek speech dropped from 98.7% to 10.6%. The write-up also documents "five cases where the measurement tool produced plausible wrong answers" and "seven features built but never shipped." Question your measurement pipeline before you question your model — a record packed with lessons that only come from production work.
The "AI Signage" Stir at 7-Eleven Stores — Headquarters: "Used at Some Stores on Their Own, Not Provided by Headquarters"
From Japan, ITmedia AI+ reports that it interviewed 7-Eleven headquarters about AI-generated signs spotted in stores and discussed online as "creepy" and "internal documents pasted as-is." Headquarters stated that it has "confirmed that some stores are using AI to create signage," and answered that the signs are "not provided by headquarters."
In an era when generative AI is available to everyone, this is a familiar Japanese story of AI output reaching customer-facing spaces through the judgment of individual frontline staff. It stands as one of the earliest cases testing the balance between chain-brand quality control and store-level discretion.