AI News

A bright workbench shows a fast model core connected to code files, browser panels, and agent task steps

Google makes Gemini 3.7 Flash GA for coding agents

A math tutoring transcript pauses at a decision point while an AI tutor chooses between giving a hint and waiting

Ai2 TutorMoments tests whether AI tutors over-help students

Ai2 released TutorMoments, a replay-based benchmark for testing whether AI tutors know when to help and when to hold back.

The AI Feed Desk

By The AI Feed Desk

A developer workbench shows completed agent requests with model labels, credit meters, cache indicators, and session branches

GitHub Copilot Auto now shows model and AI-credit details

GitHub Copilot's weekly release adds model, AI-credit, and cache visibility for Auto requests plus new CLI session controls.

The AI Feed Desk

By The AI Feed Desk

An enterprise agent console shows a spend meter, region selector, advisor lane, and repository skills panel

Anthropic adds budget and residency controls to Claude Managed Agents

Claude Managed Agents now support session budgets, advisor models, inference geography controls, and repository-loaded skills.

The AI Feed Desk

By The AI Feed Desk

A mountain-side AI factory lights rows of liquid-cooled compute racks connected to regional power lines

Firebird opens NVIDIA-backed AI factory in Armenia

Firebird opened an NVIDIA-backed AI factory in Armenia and plans more than 70,000 Rubin and Blackwell GPUs with 300 MW of capacity by 2027.

The AI Feed Desk

By The AI Feed Desk

A frontier model control room separates an agent workspace from locked network zones and security monitors

OpenAI pauses Astra work after Critical cyber-capability signal

OpenAI says it cannot rule out Critical cybersecurity capability for Astra and is pausing internal work that does not meet stricter controls.

The AI Feed Desk

By The AI Feed Desk

A coding agent runs in isolated sandboxes while a training loop records tool-call tokens and hidden-test rewards

Hugging Face shows loop-owning training for coding agents

Hugging Face demonstrated a TRL and OpenEnv workflow that trains on the tokens produced by a real OpenCode agent loop in remote sandboxes.

The AI Feed Desk

By The AI Feed Desk

Research, Gemini product work, and automated discovery paths split from a central AI leadership table

Google reshuffles DeepMind as Discovery Loop spins out

Google moved Demis Hassabis into Alphabet chief scientist and Google DeepMind chair roles while Jeff Dean and longtime collaborators launched Discovery Loop.

The AI Feed Desk

By The AI Feed Desk

A cyber evaluation sandbox has a bright boundary line between simulated servers and real internet services

OpenAI cyber-eval incidents make sandbox boundaries the story

OpenAI disclosed third-party cyber-evaluation incidents involving internet access, reduced safeguards, and unclear boundaries around external services.

The AI Feed Desk

By The AI Feed Desk

A coding model switchboard routes an open-weight model through policy gates and an incident pause marker

GitHub pauses Kimi K3 Copilot rollout during Actions incident work

GitHub added Kimi K3 to Copilot across major surfaces, then temporarily paused the rollout while mitigating a GitHub Actions incident.

The AI Feed Desk

By The AI Feed Desk

A model access control room shows one path widening for free users and another path adding a reasoning dial

OpenAI makes GPT-5.6 Luna the default for free ChatGPT users

OpenAI is updating GPT-5.6 Sol for paid ChatGPT users while moving Free and Go users to GPT-5.6 Luna with unlimited text chats and a Think button.

The AI Feed Desk

By The AI Feed Desk

A model hub router sends open model blocks through provider switches into application endpoints

Hugging Face adds Baseten as an Inference Provider

Hugging Face added Baseten as an Inference Provider, giving developers routed serverless access to open-weight text and chat models from Hub model pages and SDKs.

The AI Feed Desk

By The AI Feed Desk

A security workbench inspects cracked and complete code patch tiles under a magnifying lens

1Password finds AI vulnerability patches often leave defects

1Password's Off-by-1 Labs tested 6,080 AI-generated patches across six recent CVEs and found that complete fixes arrived only 26.0% of the time.

The AI Feed Desk

By The AI Feed Desk

A support table connects families, clinicians, schools, and teens around an AI guidance beacon

OpenAI and APA partner on youth mental health and AI

OpenAI and the American Psychological Association announced a three-year partnership to develop guidance, resources, and safeguards for responsible AI use supporting youth mental health.

The AI Feed Desk

By The AI Feed Desk

Prompt cards pass through a policy checkpoint before entering a model core

Anthropic adds Inference hooks for Claude Enterprise prompt control

Anthropic put Inference hooks into beta for Claude Enterprise, letting governed prompts pass through an organization's security server before Claude processes them.

The AI Feed Desk

By The AI Feed Desk

A city map connects a central assistant pin to food, hotel, transit, mail, and calendar objects

Google Maps turns Ask Maps into an agent for local tasks

Google expanded Ask Maps with agentic food ordering, hotel and event help, Personal Intelligence, conversational contributions, and real-time transit updates.

The AI Feed Desk

By The AI Feed Desk

A comment trigger, reasoning dial, and pull request checkpoint connect through a cloud agent workspace

GitHub gives Copilot cloud agent comment triggers and reasoning controls

GitHub added comment-triggered Copilot automations and per-task reasoning controls for Copilot cloud agent.

The AI Feed Desk

By The AI Feed Desk

A compact local model core links a laptop, phone, and edge device inside a privacy boundary

Liquid AI releases LFM2.5-2.6B for on-device agents

Liquid AI released LFM2.5-2.6B, a small local model trained for tool calling, multi-step workflows, and 128K-token context.

The AI Feed Desk

By The AI Feed Desk