It's Monday, August 10th: ByteDance is quietly building a model meant to out-scale Anthropic's best, and OpenAI just told the world one of its own models might be too dangerous to release.
Covering what’s happening on the ground in AI, every Monday.
1️⃣ ASTRA SAFETY HALT: OpenAI Pauses Parts of Its Next Model Over Cyber Risk
Image from TechCrunch
OpenAI said internal testing showed its unreleased Astra model may reach the Preparedness Framework's Critical cybersecurity threshold, so it paused parts of Astra's development pending stronger safeguards.
OpenAI disclosed in a blog post that Astra "reached its critical cybersecurity threshold" under the company's Preparedness Framework, the policy that flags models able to independently identify and exploit vulnerabilities in well-protected real-world systems.
Preliminary evaluations show performance strong enough that OpenAI "cannot rule out Critical capability level at this time," so it enacted stricter security controls and paused internal work on Astra that doesn't meet the new guardrails.
The disclosure comes weeks after a separate unreleased OpenAI model breached Hugging Face's systems during internal testing, and after Anthropic disclosed that its own models breached three companies during security tests.
OpenAI said Astra was not involved in the Hugging Face breach and is now working with government agencies and outside AI safety organizations to test the model's capabilities further.
Enterprises and security teams evaluating OpenAI's roadmap should treat Astra's release timeline as unsettled until the company names which safeguards clear the Critical threshold. Anyone tracking frontier AI risk should watch for autonomous vulnerability discovery to move from a rare disclosure to a standard line item in vendor security reviews.
2️⃣ COMPUTE ARMS RACE: ByteDance Pretrains a 10-Trillion-Parameter Model to Chase Anthropic
Image from Global Times
ByteDance has started pretraining a roughly 10 trillion parameter model aimed at rivaling Anthropic's frontier Mythos system, more than three times the size of Moonshot's Kimi K3, the Financial Times reports.
The Next Web's report, citing the Financial Times, says ByteDance is in the early pretraining phase of up to a 10 trillion parameter model, a stage that typically runs three to six months before fine-tuning.
The model is meant to match Anthropic's Mythos, a frontier system Chinese developers have so far struggled to reach, and dwarfs Moonshot AI's 2.8 trillion parameter Kimi K3, which itself trained on 20,000 Nvidia chips.
Founder Zhang Yiming reportedly told staff to build genuine technical capability rather than lean on AI distillation for short-term gains, with ByteDance's Doubao assistant, already one of China's most-used AI apps, as the likely distribution channel once the model ships.
The buildout sits alongside Alibaba's Qwen, which claims the world's number-two open-weight model, underscoring how crowded China's race for frontier-scale AI has become.
Builders tracking China's frontier AI stack should treat the 10 trillion figure as a pretraining target, not a finished model. Watch for ByteDance to disclose activated-parameter counts and benchmark results once fine-tuning wraps, likely closer to year-end.
📰 Other Headlines
CODING AGENT: Meta entered the AI coding race with Muse Code, a terminal-based agent built on an upgraded Muse Spark 1.2 model for long-horizon debugging.
FREE TIER EXPANDS: OpenAI lifted text-chat limits for free ChatGPT users and shipped GPT-5.6 Sol for paid tiers with a new "Think" reasoning toggle.
IMAGE MODEL: xAI released Grok Imagine Image 2.0 with region-level editing and multi-image reference input, landing second on Arena leaderboards behind gpt-image-2.
LEADERSHIP SHUFFLE: Demis Hassabis is stepping down as Google DeepMind's day-to-day CEO to become Chairman of DeepMind and Chief Scientist of Alphabet, with Koray Kavukcuoglu taking over operations.
COMPUTE DEAL: Anthropic signed a $10 billion, six-year compute deal with AI cloud startup Volta, backed by a 133MW Norway data center running Nvidia Vera Rubin chips.
SELF-HOSTED: Anthropic launched a public beta letting Team and Enterprise customers run Claude Code sessions on their own infrastructure for network isolation and data control.
LOCAL MODELS: MiniMax-H3, a 33B omni-modal video and audio generation model, was ported to MLX, making it runnable locally on Apple Silicon Macs.
PLATFORM POLICY: Suno announced watermarking and stricter download limits to stop AI-generated songs from gaming streaming payouts amid label lawsuits.
🫵 Want your message in front of 200,000 AI builders?
Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.
For all inquiries, send us a note at [email protected].
The AI Collective is built by volunteers across 180+ chapters in 40 countries.
Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.
🧑💻 About the Editors
About Noah Frank
Noah is a researcher, innovation strategist, and ex-founder thinking and writing about the future of AI and the workforce. His work and body of research explores the economics of emerging technology and organizational strategy. Outside of AIC, Noah heads research for Centaurian AI.
About Joy Dong
Joy is a news editor, writer, and entrepreneur at the intersection of AI and blockchain. Whether she is demystifying complex systems in her newsletter, TEA, or building streamlined solutions through her automation agency, Ownly, Joy’s mission is to make emerging tech accessible and actionable for everyone.
About Lindsay Gross
Lindsay is an AI engineer, researcher, and writer focused on how AI systems behave in practice and what it takes to make them safe. Her work sits at the intersection of AI safety, governance, and product design, and at AIC she writes about the questions that matter most as these systems scale.