AI Daily — 2026-06-15

English 中文

Sonic-3.5 and Ink-2 Lead Streaming TTS and STT · Kimi K2.7 Code Released; Local Run Shrinks 48% ·...

Covering 31 AI news items

🔥 Top Stories

1. Sonic-3.5 and Ink-2 Lead Streaming TTS and STT

The new streaming TTS and STT models, Sonic-3.5 and Ink-2, promise faster, higher-quality voice interactions for real-time voice agents, with HLS playback support to streamline streaming deployments. If adopted widely, they could raise the bar for conversational agents in customer service and accessibility. Source-x

2. Kimi K2.7 Code Released; Local Run Shrinks 48%

Kimi K2.7 Code goes open-source with a dramatic size reduction to 325GB (1T model) via Dynamic 2-bit with selective upcasting, enabling >40 tokens/second on 330GB RAM/VRAM; full-precision inference remains 610GB. The release hints at stronger performance versus K2.6 and points to a potential 6x High-Speed Mode, with access through Kimi API and Kimi Code. Source-x

3. US blocks Fable 5 and Mythos 5 access under export rules

A US export directive suspends access to Anthropic’s Fable 5 and Mythos 5 for all foreign nationals, requiring customers to disable these models to stay compliant while other Claude models remain available. Anthropic says the directive is a misunderstanding and is actively seeking restoration of access. This move underscores how export controls are reshaping AI model deployment and international collaboration. Source-x

📰 Featured

Group remaining featured news by topic, each group as a ### heading, items as bullet list:

AI Policy

Anthropic’s Fable 5 Update Shifts to Policy Compliance — Follows governance-driven pushback on jailbreak narratives, with the company hiring a cybersecurity expert to review findings and coordinating with Commerce, the CIA, and White House science advisor Michael Kratsios to navigate compliance obligations. Source-x

Machine Learning Theory & Research

Blog Frames SFT, RL, OPD as Distribution-Shaping Paradigms — Argues SFT, RL, and OPD reshape a model’s distribution in distinct ways, with on-policy data identified as the key load-bearing ingredient and OPD’s convergence to RL-like outcomes presented as strong evidence. Source-x

Tools & Safety

Hermes Agent Adds Stripe Skills in Partnership — Hermes gains Stripe capabilities for purchases, per-call API payments, and SaaS provisioning, with configurable safety limits on every action to manage risk. Source-x

Open Source & Industry Benchmarks

Napoleon-class Mistral model tops 10T params, French-only outputs — Unverified rumors claim a Napoleon-class 10T-parameter model with safety-restricted French-only code outputs; discussion is circulating largely on social media. Source-x

Industry & Applications

Sakana AI Launches Sakana Marlin, Its First Commercial Research Assistant — Sakana Marlin is introduced as the company’s first business-focused autonomous research assistant; details at sakana.ai/marlin-release. Source-x

AI Platform & Software Engineering

Factory 2.0: From Coding Agents to Software Factories — FactoryAI expands from coding agents toward full software factories, highlighting new automation capabilities and HLS playback support for media streams. Source-x

AI Agents & Memory (Research/Open Source)

Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — Proposes MRAgent, a memory framework combining an associative memory graph with active reconstruction to adapt memory access on the fly, representing memory as a Cue-Tag-Content graph to enable reconstruction rather than retrieval. Source-huggingface

⚡ Quick Bites

From Chatbot to Digital Colleague: A Shift to Persistent Autonomous AI — Explores persistent autonomy in AI assistants and implications for reliability and collaboration. Source-huggingface
Orchestra-o1 Advances Omnimodal Agent Orchestration for LLMs — Advances omnimodal agent orchestration for coordinating tool use across modalities. Source-huggingface
Le Chaton Fat Claims ASI, Reverses Nvidia GPUs — Provocative, unverified claim about ASI status and hardware supply. Source-x
OmniDirector Enables Multi-Shot Camera Cloning without Cross-Paired Data — Introduces multi-shot cloning without cross-paired data. Source-huggingface
APPO Advances Agentic RL for Multi-turn LLM Tool-Use — Agentic RL advances for long-horizon LLM tool use. Source-huggingface
ChatGPT Eases Restrictions, Plans More Personalities and Adult Access — Signals looser restrictions and expanded personalization options. Source-reddit
ChatGPT answers confidently even when it has no idea — Highlights overconfidence in some responses. Source-reddit
ChatGPT asked to craft an unfunny New Yorker cartoon — Tests prompt handling for humor generation. Source-reddit
Open Autonomous Robots Textbook Released Under CC; GitHub Source — Opensource robotics textbook with GitHub repository. Source-github
Reddit Prompt: ChatGPT Creates World Event Image with Hidden Easter Egg — Demonstrates image prompt experimentation with hidden elements. Source-reddit
ChatGPT Prompts Self-Awareness in User, Reddit Post Claims Benefit — Debates prompts that claim to induce user self-awareness. Source-reddit
User-generated AI image highlights need for detailed prompts — Emphasizes prompt quality for image generation. Source-reddit
Video Generation with ChatGPT Images 2.0 — Combines ChatGPT-generated images with video generation workflows. Source-reddit
ChatGPT Delivers a Funny Moment on Reddit — Reports a humorous moment captured on Reddit. Source-reddit
Every AI Assistant Until You Ask ‘Are You Sure?’ — Explores prompts that enforce caution in AI interactions. Source-reddit
Is ChatGPT Underpriced for Its Capabilities? — Addresses pricing and perceived value. Source-reddit
When AI is Smart Enough to Play Dumb — Examines strategic misdirection in AI behavior. Source-reddit
Why did my ChatGPT show a graph and tell me to ignore? — Investigates prompts that prompt graph-based guidance. Source-reddit
If AI Writes Your Essay, You Are Not the Author — Raises authorship and attribution concerns. Source-reddit
Box in ChatGPT’s Basement Sparks Reddit Post — Quirky user-post about prompts and hidden items. Source-reddit
Reddit User Seeks Ten Original Sentences via AI Prompt — Prompts challenge to generate ten original sentences. Source-reddit

Generated by AI News Agent | 2026-06-15