Friday, August 14

The signal

1. OpenAI Ultrafast: GPT-5.6 Sol at 750 tokens/sec via Cerebras What happened: OpenAI and Cerebras announced Ultrafast mode on August 13; it runs GPT-5.6 Sol on Cerebras wafer-scale hardware at up to 750 output tokens/second (14X standard baseline of ~53 tok/sec), in limited API preview with no claimed quality reduction. Why it matters: Frontier-level intelligence at near-real-time speed changes the build math for voice agents, live security response, and high-volume agentic loops where the current bottleneck is generation latency rather than model quality. Confidence: Confirmed Sources: OpenAI (primary) · Cerebras (primary) · TechCrunch (independent)

2. darioamodei / Anthropic in talks to buy Decart AI for ~$6B What happened: Bloomberg reported August 13 that Anthropic is in preliminary talks to acquire Israeli GPU-efficiency startup Decart AI for roughly $6 billion; no deal signed, talks could collapse. Why it matters: Decart’s DOS inference-optimization stack and chip-efficiency research would reduce Anthropic’s per-token cost at a moment when it has stacked ~$71B in compute commitments via off-balance-sheet SPVs and is targeting a Q4 2026 Nasdaq listing. Confidence: Reported Sources: Bloomberg (independent) · Fortune (independent) · Benzinga (independent)

Since last brief

• claude-code auto mode is now the default permission model for Pro, Max, and Team accounts as of today (Aug 14); enterprise and API rollout follows over the coming month TechCrunch (independent)
• Qwen 3.8 Max open weights (2.4T-A95B + FP8 checkpoints) shipped on HuggingFace and ModelScope on August 12, closing the watchlist window; the companion 27B dense model is still absent latent.space (community) · llm-stats.com (independent)

Models & research

Qwen 3.8 Max open weights (2.4T-A95B)

Claim: Alibaba released downloadable weights on August 12; 2.4T MoE with 95B active parameters, 1M-token context, native image + video input; claims 86.1 on OSWorld-Verified, ahead of Claude Fable 5 (85.0) and GPT-5.6 Sol Max (83.2). Evidence: Weights confirmed on HuggingFace. OSWorld-Verified score is Alibaba’s own evaluation with no independent reproduction yet. Caveat: Benchmark is company-run. License terms not yet disclosed. The 27B dense companion model remains absent. Worth your time: Read if you’re evaluating open-weight computer-use agents; Skim otherwise.

NVIDIA Nemotron 3.5 Lightning + NeMo Switchyard

Claim: Released August 11; Nemotron 3.5 Lightning is a 30B MoE (3B active), Mamba-2 + MoE + Attention hybrid, 1M context, claimed 4X faster than comparably-sized models. Switchyard is an open-source LLM router that dispatches calls across models mid-task; NVIDIA reports 74% cost reduction when routing between Lightning and Claude Opus 4.8. Evidence: NVIDIA official release. Customer case studies: Cognition (-28%), Ramp (-58% cost, -33% runtime). All reported by NVIDIA. Caveat: Speed and cost claims are NVIDIA’s own; independent benchmarks pending. Worth your time: Skim if you’re routing mixed-complexity workloads across open and frontier models.

Tools worth touching

NVIDIA NeMo Switchyard

• Does: Open-source model router that dynamically dispatches LLM calls across open and proprietary models mid-task, compatible with OpenAI and Anthropic APIs, no application rewrites required
• Why now: Just released (Aug 11), fully open source; deployed users report 28-74% cost reduction on mixed-complexity workloads
• Try if: You run workloads where a meaningful fraction of tasks don’t need frontier model quality
• Skip if: Your workloads are uniformly hard, or you already route via an LLM proxy
• Links: NVIDIA blog · VentureBeat · SiliconANGLE

cathrynlavery/diagram-design

• Does: 29 self-contained HTML+SVG editorial diagram types built for Claude Code, Codex, and Pi; no external dependencies, no Mermaid, no shadows
• Why now: +4,475 stars today (14.6K total); strong practitioner signal that AI-native diagram tooling has a real gap
• Try if: You want an agent to generate publication-quality architecture or flow diagrams in a consistent visual style
• Skip if: Mermaid or draw.io already fits your workflow
• Links: GitHub

Moves on the board

Other

• Anthropic’s compute commitments via off-balance-sheet SPVs reportedly total ~$71B over the last two months, keeping hardware debt off its balance sheet; run-rate revenue cited at $30B (from $9B end-2025) WION · HPCwire
• Governor Newsom directed California agencies on Aug 10 to establish a first-in-nation AI Cyber Defense Program inside the CA Cybersecurity Integration Center; AI-enabled vulnerability detection and incident response CA Governor

Watchlist

• OpenAI ExploitGym / Kill Switch Act: still in House committee; no Senate companion bill (opened 2026-07-23)
• Kimi K3 llama.cpp PR #26185: community reports suggest a merge completed by Aug 4 but mainline still reportedly cannot read the architecture; status unclear (opened 2026-07-23)
• White House AI security framework: open-weight exemption confirmed Aug 4; full text still not public (opened 2026-07-30)
• EU AI Act enforcement: technical compliance dialogues ongoing; no formal investigation or named GPAI provider yet (opened 2026-07-26)
• Anthropic eval breach / PyPI transcript: METR review ongoing; transcript still unpublished; Aug 6 deadline passed (opened 2026-07-31)
• OpenAI Astra: “Critical” cyber flag; in 30-day govt review; September is earliest possible release (opened 2026-08-02)
• Meta Muse Spark 1.1 breach: post-mortem and regulatory response still pending (opened 2026-08-06)
• California Frontier AI Safety Act: passed Assembly Aug 10; awaiting governor signature (opened 2026-08-11)
• Demis Hassabis / Google DeepMind: Kavukcuoglu now SVP; Discovery Loop launched Aug 5; watching Gemini roadmap and talent pipeline impact (opened 2026-08-11)
• Anthropic / Decart AI acquisition: $6B talks ongoing per Bloomberg Aug 13; watching for deal closure or collapse (opened 2026-08-14)