Large Language Models News
Sakana launches Namazu-powered translation in Sakana Chat
Sakana Translate adds Japanese, English, and Chinese translation to Sakana Chat, using the Namazu model series for translate, proofread, and Q&A modes.
By The AI Feed Desk
Online safety monitoring paper tests calibrated verifier alarms
A new arXiv paper studies LLM safety monitoring with an external verifier signal, calibrated risk threshold, and real-time alarm decision.
By The AI Feed Desk
LACUNA asks whether LLM unlearning reaches the weights
The LACUNA testbed evaluates whether LLM unlearning methods target the parameters that stored synthetic PII, not only whether outputs stop revealing it.
By The AI Feed Desk
GitHub Models gets a July 30 shutdown date
GitHub will fully retire GitHub Models on July 30, ending the playground, model catalog, inference API, and BYOK endpoints for all customers after two July brownouts.
By The AI Feed Desk
Metacognition-Bench tests whether models notice their own mistakes
Metacognition-Bench measures whether language models detect tempting wrong reasoning paths, pairing trap-rate evaluation with adapters that flag likely free-form errors.
By The AI Feed Desk
BaseRT makes Apple Silicon LLM speed a runtime-design question
BaseRT's paper and release argue that native Metal kernels, unified-memory-aware layout, and custom dispatch can raise local LLM throughput on Apple Silicon.
By The AI Feed Desk
Kimi K2.7 Code gives Copilot an open-weight model option
GitHub says Kimi K2.7 Code is the first open-weight model selectable in Copilot, with Azure hosting, gradual rollout, usage-based billing, and enterprise policy controls.
By The AI Feed Desk
Persistent coding agents create a distributed attack surface
A new arXiv paper argues that persistent coding agents can hide malicious behavior across multiple pull requests, making monitor design a stateful problem.
By The AI Feed Desk
Hugging Face and Cerebras make open voice AI a latency problem
A Hugging Face and Cerebras speech-to-speech demo uses Parakeet, Gemma 4, Cerebras inference, and Qwen3TTS to show where voice AI latency actually lives.
By The AI Feed Desk
ScarfBench shows coding agents still struggle with Java migrations
IBM Research's ScarfBench tests whether AI coding agents can preserve behavior while migrating Java applications across enterprise frameworks.
By The AI Feed Desk
Hugging Face and Every Eval Ever make model-card scores more inspectable
Community Evals and Every Eval Ever now connect model-page benchmark scores to structured provenance records.
By The AI Feed Desk
OpenAI's GeneBench-Pro makes biology benchmarks about judgment
GeneBench-Pro tests whether AI agents can handle ambiguous computational-biology analysis, not just clean benchmark questions.
By The AI Feed Desk
Claude Sonnet 5 turns Anthropic's default model into an agent model
Anthropic made Sonnet 5 the default for Free and Pro users while positioning it as a lower-cost agentic model close to Opus 4.8.
By The AI Feed Desk
TraceLab turns real Codex and Claude Code sessions into serving data
The University of Washington TraceLab release studies coding agents as production workloads, with public traces across sessions, tool calls, tokens, cache behavior, and latency.
By The AI Feed Desk
Allen AI's DiScoFormer tests one transformer for density and score
The Hugging Face writeup frames DiScoFormer as a reusable estimator for density and score, with stronger high-dimensional results than kernel density estimation.
By The AI Feed Desk
Cognition's FrontierCode asks whether AI code would survive review
FrontierCode evaluates coding agents on mergeability, code quality, scope, tests, and maintainer judgment instead of only functional correctness.
By The AI Feed Desk
Hugging Face makes vLLM serving a one-command Jobs workflow
HF Jobs can now spin up a private OpenAI-compatible vLLM endpoint for tests, evals, and batch generation without provisioning servers or managing Kubernetes.
By The AI Feed Desk