AI News

A compact translation engine moves blank language cards between Japanese, English, and Chinese contexts

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.

The AI Feed Desk

By The AI Feed Desk

Generated output cards pass through a calibrated verifier gate while one unsafe stream triggers a red alarm path

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.

The AI Feed Desk

By The AI Feed Desk

A magnifying instrument locates a highlighted region inside a transparent model-weight lattice while synthetic personal-data traces are blocked below

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.

The AI Feed Desk

By The AI Feed Desk

A compact AI factory module combines a chip wafer, server racks, cooling pipes, and power equipment

NVIDIA frames U.S. AI buildout as a 43-state supply chain

NVIDIA says its U.S. partner network spans 43 states and plans up to $500 billion of American-built AI infrastructure with semiconductor, system, power, and cloud partners.

The AI Feed Desk

By The AI Feed Desk

Shared AI credit tokens flow through team lanes while one cost center lane is stopped by a cap gauge

GitHub cost centers can cap shared AI credit pools

GitHub added AI credit pools for cost centers, letting enterprise admins cap how much monthly included Copilot credit one group can draw before metered budgets apply.

The AI Feed Desk

By The AI Feed Desk

A glass model catalog panel moves from a dark retiring shelf toward a cloud model hub along a bright migration path

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.

The AI Feed Desk

By The AI Feed Desk

A glass reasoning path highlights a trap branch while a diagnostic lens catches the error

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.

The AI Feed Desk

By The AI Feed Desk

A metallic chip tile connects to flowing local model tokens and a small speed gauge

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.

The AI Feed Desk

By The AI Feed Desk

A routing switchboard sends coding tasks into different model lanes from an admin control plate

Copilot auto model selection turns routing into an admin setting

GitHub Enterprise admins can now default Copilot conversations to auto model selection through managed-settings.json, tying model routing to enterprise AI standards.

The AI Feed Desk

By The AI Feed Desk

An open model block enters a coding model picker rail inside a cloud frame

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.

The AI Feed Desk

By The AI Feed Desk

A browser-shaped glass panel, image tile, and testing checklist converge on a developer workstation

Copilot browser and vision tools make agents multimodal testers

GitHub made Copilot browser tools in VS Code generally available and opened Copilot vision across plans, giving coding agents browser, screenshot, image, and PDF context.

The AI Feed Desk

By The AI Feed Desk

A sequence of pull request tiles is connected by a hidden red thread under a monitoring lens

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.

The AI Feed Desk

By The AI Feed Desk

A backend service block streams agent state and artifact tiles to a browser client

Genkit makes full-stack agents a shared backend and client contract

Google's Genkit Agents preview packages conversational-agent plumbing behind a server-defined agent and a shared chat interface for frontend apps.

The AI Feed Desk

By The AI Feed Desk

A controlled workflow graph routes tasks through two glowing AI reasoning nodes

Google ADK 2.0 puts workflows around agents

Google's ADK 2.0 framing separates deterministic workflow routing from open-ended model reasoning, giving production agents a stricter execution boundary.

The AI Feed Desk

By The AI Feed Desk

An abstract coding agent session timeline flows into a secure audit collector

GitHub exposes Copilot agent sessions for enterprise audit

GitHub's public preview lets enterprise owners stream or retrieve Copilot agent session data, including prompts, responses, and tool calls, across major Copilot clients.

The AI Feed Desk

By The AI Feed Desk

A large AI factory rack sends green revenue tokens toward a smaller cloud node

NVIDIA turns AI cloud capacity into a revenue-sharing model

NVIDIA's new AI cloud model pairs revenue sharing with credit support, giving emerging cloud providers a way to finance AI factories while tying NVIDIA to downstream token demand.

The AI Feed Desk

By The AI Feed Desk

A global map of soft data tiles connects to a central chat thread and several language markers

OpenAI Signals points to a wider ChatGPT adoption story

OpenAI's RSS says new Signals data shows ChatGPT adoption growing globally, but the full article body was blocked to direct fetch, so the responsible read starts with source limits.

The AI Feed Desk

By The AI Feed Desk