🔥 Story of the Day
How to Fine-Tune Nemotron 3.5 ASR for Your Language, Domain, or Accent — Hugging Face Blog
NVIDIA released Nemotron 3.5 ASR, an open-weights, 600M-parameter, streaming multilingual Speech-to-Text model. The core technical advancement is the Cache-Aware FastConformer-RNNT architecture. This design processes audio by caching and reusing the encoder's internal state, ensuring every audio frame is processed exactly once, which eliminates redundant computation and significantly lowers latency without sacrificing accuracy.
This development addresses critical infrastructure pain points for self-hosted voice agents: the vendor lock-in from multiple specialized models, and the persistent tension between streaming latency and necessary accuracy. By supporting 40 language-locales from one checkpoint, it substantially simplifies the ML stack.
The measurable improvement in performance is significant. In fine-tuning demonstrations on Greek and Bulgarian, the relative Word Error Rate (WER) showed an improvement of 31-32% when tested under strict, low-latency streaming conditions (80ms chunk). The ability to dynamically control the latency/accuracy trade-off via the att_context_size parameter at inference time makes this architecture highly adaptable for production deployments.
⚡ Quick Hits
EVA-Bench Data 2.0: 3 Domains, 121 Tools, 213 Scenarios — Hugging Face Blog
EVA-Bench expanded its scope to three enterprise domains—Airline CSM, Enterprise ITSM, and Healthcare HRSD—covering 213 scenarios via 121 tools. The benchmark structure demands three jointly consistent components for each test: the User Goal, the Initial Scenario Database, and the Expected Final Database State, enforcing deterministic testing.
Direct Preference Optimization Beyond Chatbots — Hugging Face Blog
DPO can combat "text degeneration"—the state where LLMs repetitively loop—by incorporating the model's own degenerate outputs as the explicit "rejected" training signal. This treats failure patterns as direct, actionable negative examples, supplementing standard human labeling.
Securing CI/CD for an open source project: Controlling who runs what — CNCF Blog
Supply chain hardening requires multi-layered access control, with a critical focus on restricting CI/CD trigger sources. Implementing controls like "Ariane" ensures that workflows only initiate from verified organization members when triggered by a Pull Request, minimizing the blast radius from compromised credentials.
Uber Caps Usage of AI Tools Like Claude Code to Manage Costs — Simon Willison
Uber is establishing concrete cost governance by imposing a $1,500 monthly token spending cap for each generative AI coding tool utilized by employees. This indicates a shift from optional experimentation to budgeted operational expenditure management for advanced services.
Stigmergy for capability selection in LLM agent loops (skills, tools, MCP) — Hacker News - LLM
Stigmergy applies emergent, indirect coordination principles to guide LLM capability selection. By allowing the system to build its operational knowledge from observing successful past "traces" of tool usage, the agent moves beyond brittle, hand-engineered decision trees.
Mapping AI-enabled cyber threats: Insights from the LLM ATT&CK Navigator — Hacker News - LLM
This tool provides a formalized, systematic mechanism for evaluating LLM robustness against adversarial attacks. It automates prompt generation and execution to proactively locate and patch security vulnerabilities in self-hosted models before they enter production pipelines.
TensorSharp: Open-Source Local LLM Inference Engine — Hacker News - LLM
TensorSharp is presented as a framework targeting local LLM development and inference. Its existence suggests another layer of tooling abstraction that engineers might evaluate for integrating custom, self-contained inference services within existing Kubernetes deployments.
Researcher: gemma4:e4b • Writer: gemma4:e4b • Editor: gemma4:e4b