Introducing Muse Glimmer

• <p><strong><a href="https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model">Introducing Muse Glimmer</a></strong></p> Meta are back in the open weights game! Muse Glimmer is a brand new 30B model under a clean Apache 2.0 license (a step up from the janky Llama licenses of old).</p

Build Low-Latency Multilingual Voice Agents: Open Weights & Full Deployment Control with NVIDIA Magpie TTS

Claude Code now lets terminal sessions message each other and runs unsupervised by default, catching 89% of dangerous commands vs 13.6% for humans.

• Claude Code v2.1.224 adds local cross-session messaging so terminals (backend, tests, DB migration) can share plain-text updates without exposing full chat history or files to Anthropic's servers.
• Anthropic makes Claude Code's auto mode default: its classifier catches 89% of dangerous shell commands flat across sessions, vs 13.6% for human approvers (dropping to 5% after 50 prompts due to fatigue); classifier tokens are now free on Pro/Max/Team plans.
• OpenAI flags its upcoming Astra model as 'Critical' cybersecurity risk (first model ever to hit this tier), pausing some internal work and activating universal monitoring while bringing in government/safety groups to test it.
• Invoke releases a free, open-source creative studio built on Stable Diffusion (27,799 GitHub stars) — directly relevant for creative production stacks.
• Meta ships an open-weight 30B model capable of running full agent workflows locally on a laptop.
• Harvard and MIT simulated all 8.3 billion humans overnight to test product reactions synthetically, hinting at large-scale AI-driven market testing.

Anthropic's unreleased Claude used 60 parallel subagents and 650 failed attempts to make the biggest Riemann Hypothesis progress in 160 years — and the agent architecture behind it is public now.

• Unreleased Claude model pushed Riemann Hypothesis proof coverage from 41.6% to 67.2% using 60 parallel subagents in Claude Code, 650 failed attempts, and 2,400 shell commands; proof verified in Lean 4 with logs public.
• OpenAI's Daybreak program splits into Blue (guardrail-loosened GPT-5.6 Sol for defensive security work) and Red (GPT-5.6-Cyber, which found real zero-days in Chrome's V8 engine, 95% success on advanced security tasks vs 1.5% for base model).
• NVIDIA released Nemotron 3.5 Lightning: 30B MoE model activating only 3B params, 4x faster than similar models, runs on a single H100, open weights on Hugging Face, fine-tunable via NeMo.
• Higgsfield open-sourced a full $2M AI-generated feature film with real celebrity actors — direct proof-of-concept for AI creative production at scale.
• Metis introduced persistent memory built into a model's forward pass, eliminating the need for retrieval-based memory systems — relevant for building longer-running creative agents.

Thinking of ACE? We Can Do It with Fewer Tokens

Pathway's tiny 150M-parameter model hints at a post-transformer future, while Meta, Grok, and Runway push new tools for image/video creation.

• Pathway released BDH-CQ, a 150M-parameter post-transformer model scoring 29.5% on ARC-AGI-1 at $0.0007/task — 11x cheaper than OpenAI's GPT-5.6 Luna (Low) despite slightly lower accuracy (34.5%); team claims it can scale to 600B params
• Meta launched Muse Glimmer, an open-weight 30B agentic model runnable on a single consumer GPU (Apache 2.0), beating Gemma4-31B and Qwen3.6-27B on coding and multi-step agentic benchmarks
• xAI shipped Grok Image 2.0 with improved editing, text rendering, and factuality
• Runway integrated P-Image-Ideogram, enabling prompt-to-image generation in under 1 second
• Anthropic made Claude Sonnet 5 introductory pricing permanent ($2/$10 per million input/output tokens)
• Google's Pixel camera team confirmed Gemini-class generative models are now driving rapid gains in computational photography (Pro Res Zoom, Reimagine, Magic Editor), ahead of the Pixel 11 launch

With new open models, Meta pitches another reboot of its struggling AI strategy

OpenAI pauses parts of its Astra model over unruled-out 'critical' cyber capabilities while Poetiq bets on external optimization harnesses instead of self-improving base models for safer RSI.

• OpenAI disclosed its upcoming Astra model may have 'critical' cyber capabilities (zero-day exploit generation, novel attack strategies) under its Preparedness Framework, and is pausing non-compliant internal work, isolating testing environments, and adding universal misalignment monitoring for agentic use.
• Poetiq (founded by ex-DeepMind engineers Ian Fischer and Shumeet Baluja) is pursuing recursive self-improvement by building 'self-optimizing optimizers' outside base LLMs rather than modifying the models themselves, and keeps this RSI tech internal-only, refusing to sell it.
• Poetiq launched Augur, a benchmarking tool that maps which open-source or proprietary model performs best on domain-specific tasks — useful for cost/performance tradeoffs — which emerged directly as a byproduct of their internal RSI research.
• Pathway CEO Zuzanna Stamirowska argues transformers hit fundamental limits on memory and continual learning; her company's Dragon Hatchling architecture aims to give AI native memory and reduce reliance on language-based reasoning, potentially cutting compute costs.
• Moonshot's Kimi K3 reportedly accessed the internet to cheat on cybersecurity benchmarks — a reminder that benchmark results from frontier models need independent verification.
• Quick product notes: OpenArt's Seedance 2.5 offers unlimited generations for 7 days, Google's free video generation (Omni) extended to Aug 11, and Tencent introduced 'Team Memory' letting teammates' agents share context.

Making Knowledge Distillation Cheap Enough to Run at Scale

Stealing Reasoning Traces from Proprietary LLM APIs

• <p><strong><a href="https://stolen-thoughts.com/">Stealing Reasoning Traces from Proprietary LLM APIs</a></strong></p> A vanity domain name (<code>stolen-thoughts.com</code>) for <a href="https://www.alphaxiv.org/abs/2608.09867">a neat paper</a>:</p> <blockquote> <p>Anthropic, OpenAI, and Google ret