Back to MergeSort

AI News Digest - September 25, 2026

15 stories · September 25, 2026

Listen to the podcast
2.8 MB · Download MP3

Top Stories

product_launch

Meta Connect 2026 Showcases Muse AI Agent and Hardware Ecosystem Amid Industry-Wide AI Breakthroughs

At its annual Connect conference, Meta unveiled major updates to its viral Muse AI agent, including the Muse Charm keychain device, new Ray-Ban smart glasses, and FDA-cleared hearing aid features. The announcement coincided with a wave of industry-wide releases, including OpenAI's GPT-6 Sol and Luna models, Anthropic's Claude Opus 5.5, and a major scientific discovery where Anthropic's AI agents identified a novel gene-editing enzyme system.

Sources: Ben's Bites, DeepMind Blog, Digg AI, Hacker News, HuggingNews, Latent Space, Michael Parekh AI, Reddit AI, Simon Willison, The Batch, The Neuron, The Rundown AI, Vault Bookmarks

policy

OpenAI and Anthropic CEOs Urge UN Security Council to Establish Global AI Safety Standards

Sam Altman and Dario Amodei addressed the U.N. Security Council to advocate for coordinated global safety measures and shared evaluations for frontier AI models. The leaders warned of severe risks to humanity, including AI-assisted biological weapons, and pledged to slow down model releases if necessary to ensure safety.

Sources: HuggingNews, The Neuron

funding

DeepSeek Hits $1 Billion Revenue Run Rate Amid $7.5 Billion Funding Round

Chinese AI startup DeepSeek has reached an annualized revenue run rate of $1 billion, doubling its previous rate. The company is currently finalizing a $7.5 billion fundraising round at a $75 billion valuation as it prepares for a Shanghai IPO.

Sources: Michael Parekh AI, The Neuron

policy

U.S. and China Reject New Limits on Superintelligence AI Development

Following a summit in Washington, U.S. President Donald Trump and Chinese President Xi Jinping agreed to avoid new restrictions on artificial intelligence development, effectively ending debates about slowing down progress. The U.S. designated the Department of Justice as its primary regulatory guardrail while both nations discussed establishing an AI incident-alert channel.

Sources: HuggingNews

product_launch

Microsoft launches major Copilot update with Autopilot and Code

Microsoft announced a significant update to its Copilot AI assistant, positioning it as a work operating system by introducing the Autopilot agent for enterprise workflows and the Code application builder. The release also integrates a Home hub, fully embeds Office tools, and adds a proactive information feed called Today.

Sources: HuggingNews

More Stories

policy

Sanders and Casar Introduce Bill to Ban Artificial Superintelligence

Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which seeks to prohibit AI exceeding human cognitive capabilities and impose prison terms of up to 20 years for violators. The legislation would pause training on advanced AI models above a specific compute threshold until a new cabinet-level Department of Artificial Intelligence establishes safety rules.

Sources: HuggingNews

funding

SoftBank Raises $11.1B in Junk Bond Sale to Fund OpenAI Investment

SoftBank issued approximately $11.1 billion in dollar and euro-denominated notes to complete a $10 billion payment for a 13% stake in OpenAI. The massive debt offering highlights the aggressive capital deployment fueling major AI industry developments.

Sources: HuggingNews

product_launch

Security researcher exports internal filesystem of Meta's 'Muse' AI agent

A security researcher utilized Meta's AI agent 'Muse' to archive its session environment, extracting 6.8 GB of unpacked runtime files, documentation, and internal code. The findings shed light on Meta's internal agent architecture, memory management, and unreleased tool integrations.

Sources: Reddit AI

product_launch

OpenAI reportedly developing $500 ChatGPT Pro Max tier

TestingCatalog reported that OpenAI is developing a $500-per-month 'Pro Max' subscription tier utilizing Cerebras infrastructure, featuring Codex access and faster processing. OpenAI has not officially confirmed the pricing, features, or existence of this tier.

Sources: Reddit AI

partnership

Oracle Invokes Force Majeure on $165B New Mexico AI Data Center Project

Oracle has issued a force majeure notice to developer Blue Owl Capital regarding its 2.45GW 'Project Jupiter' AI campus in New Mexico, which is part of the massive $165B Stargate AI buildout for OpenAI. The move allows Oracle to defer payments and shield against rising expenses amid regulatory, water rights, gas pipeline, and power permitting delays. The setbacks highlight the growing infrastructure and environmental constraints facing massive AI data center developments.

Sources: HuggingNews

policy

17 Democratic Senators Push for U.S.-China AI Development Pause

A coalition of seventeen U.S. Democratic senators led by Senator Chris Van Hollen urged President Trump to negotiate a safety pact with China to slow or pause frontier AI development and establish shared guardrails. While the Trump administration rejects a full pause, it has discussed a more limited mechanism with Beijing to share information on AI incidents.

Sources: HuggingNews

other

Goldman Sachs Forecasts $1.2T AI Spend for US Hyperscalers by 2027

Goldman Sachs strategists predict that US hyperscalers will increase infrastructure outlays by 54% to $1.2 trillion in 2027, exceeding Wall Street consensus. The bank estimates these providers will need around $300 billion in annual AI revenue to break even.

Sources: HuggingNews

research

Apple Details Integration of Homomorphic Encryption with Machine Learning

Apple explains how it combines homomorphic encryption with machine learning to power features like Enhanced Visual Search and private information retrieval while protecting user privacy.

Sources: Lobsters AI

product_launch

Qwen3.8-Flash-Next runs at 65 tokens per second on consumer 12GB VRAM hardware

A developer built a custom inference engine optimizing the Qwen3.8-Flash-Next model to achieve high token speeds on a 12GB RTX 5070 GPU using specialized quantization techniques. This demonstrates significant efficiency gains for running large language models locally on consumer hardware.

Sources: Reddit AI

research

Qwengram-0.8B achieves 5.05% lower validation perplexity using transferred n-gram memory

A researcher transferred Qwen3.8 Flash-Next's pretrained PLE n-gram memory into Qwen3.5-0.8B, keeping the backbone frozen while training a small reader and dynamic gate, achieving a 5.05% reduction in validation perplexity.

Sources: Reddit AI

From the Podcasts

Runway’s WorldPrompt and the Engineering of Real-Time Worlds

GWM Worlds 2 uses persistent context and timed actions to steer a world model generating video and audio in real time.

Latent Space · Sep 25

Foundries vs Navigators: Lowering the Cost of Science

Guest Post: In science, thinking has gotten cheap but doing has not. This asymmetry is reshaping how research companies operate, largely inconspicuously.

Latent Space · Sep 24

Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee

Can AI transform legacy incumbents rather than replacing them? Sequence Holdings co-founder and CEO Michael Lee joins Sarah Guo to discuss how holding company structures and engineering integrations…

No Priors · Sep 24

How Deep Learning Finally Cracked Messy Tables - Frank Hutter

Frank Hutter, co-founder of Prior Labs, talks about TabPFN, a tabular foundation model that makes predictions in a single forward pass, and the research behind it. TabPFN is pre-trained on synthetic…

ML Street Talk · Sep 23

🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

Radical Numerics is using biological chain-of-thought and multimodal perception to keep up with the bio-defense arms race, design new genomes and gain insights into biology itself.

Latent Space · Sep 23

🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science

We talked to Google’s Oscar winning “Giganerd” about automating science, solving climate change, and how future generations can contribute to science in the age of superintelligent AI

Latent Space · Sep 22

Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI

The definitive Jev podcast with its lead creator.

Latent Space · Sep 21

Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick

Alexander Mattick is a researcher at Fraunhofer IIS and a PhD researcher at the University of Technology Nuremberg (UTN), and a regular on Yannic Kilcher's Discord. He first came on MLST in 2022,…

ML Street Talk · Sep 21

AI:AM Highlights: Zvi on Pacing & Trump-Xi, Astra better behaved than Fable? + a new LLM Pain Axis??

Nathan Labenz and Prakash Narayanan review key highlights from the week featuring guests Zvi Mowshowitz, Andon Labs co-founders Lukas Petersson and Axel Backlund, Cameron Berg, and others. The…

The Cognitive Revolution · Sep 19