Samwise Tech/AI/Robotics Newsletter
Sunday, August 23, 2026
No Frontier AI Lab Scores Above 3/5 on Rogue Model Containment
A new audit by Guidelight AI Standards scores five frontier AI labs on containment readiness for rogue models — and no lab scores above 3 out of 5. Chief scientist Steven Adler, a former OpenAI safety researcher, conducted the study using only publicly available information; OpenAI received the highest score while Anthropic and Meta scored lowest. The findings land as regulatory pressure mounts: California’s SB 53 requires containment frameworks, New York’s RAISE Act takes effect in January, and the AI Kill Switch Act has been introduced in Congress. A recent Hugging Face data breach adds urgency to the findings.
Sources: TechCrunch
Training AI on Copyrighted Books: Courts Say It’s Complicated
Is training AI models on copyrighted books legal? Last year, Judge William Alsup ordered Anthropic to pay a $1.5 billion settlement in a case brought by authors — but he ruled that AI training itself was lawful, penalizing the company only for sourcing books from illegal shadow libraries. Copyright law hasn’t been updated since 1976, forcing courts to interpret 50-year-old rules around fair use. In Thomson Reuters v. Ross Intelligence, the court found it was not fair use to train on Reuters content to build a directly competing legal platform. Courts have also ruled that purely AI-generated content is not copyrightable.
Sources: TechCrunch
DeepMind Spinout Inherent’s Faraday Agent Beats Frontier Models at Research Replication
London AI startup Inherent, founded by Google DeepMind alumni, says its Faraday agent has outperformed frontier models from Anthropic and OpenAI at replicating the findings of published scientific papers. Faraday runs on Qwen 3.6, a 27-billion parameter model, yet beat Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on the task. Cofounder and chief scientist Edward Hughes says the goal is building AI with ‘research taste’ — an instinct for what experiments are worth running. Inherent, which raised a $50 million seed round in May, uses reinforcement learning to train its agents rather than spelling out explicit rules.
Sources: TechCrunch
Nvidia’s Linear Math Trick Speeds AI Model Handoffs by up to 25x
Nvidia researchers have developed a cross-model KV cache transfer technique that cuts the cost of swapping AI models mid-task in agentic workflows. Normally, each model switch forces the receiving model to recompute the entire conversation — a major compute and latency bottleneck. Nvidia’s method uses simple per-head ridge regression, not deep learning, to map a source model’s KV cache into a target model’s format. On compatible model pairs across the Qwen3, Llama 3.1, and Ministral families, the technique runs 2.7 to 25 times faster than recomputing from scratch while retaining up to 98% of the target model’s standalone accuracy.
Sources: VentureBeat
Starcloud Raises $250M for Orbital AI Data Centers as Rocket Capacity Tightens
Orbital AI compute startup Starcloud has raised a $250 million extension to its Series A, valuing the company at $2.3 billion. The only known operator of a Nvidia H100 GPU in orbit, Starcloud is racing to lock in launch capacity before SpaceX phases out Falcon 9 in 2028. CEO Philip Johnston has filed for FCC permission to operate 88,000 spacecraft; the first two Starcloud-2 satellites (8 kW each) target rideshare launches in 2027 for U.S. government customers. Nvidia invested $25 million and is co-developing the Vera Rubin Space-1, a purpose-built space GPU, informed by data from Starcloud’s first orbital mission.
Sources: TechCrunch
Flock Safety CEO Calls for ‘Compromise’ as Surveillance Backlash Grows
Flock Safety CEO Garrett Langley is calling for a national ‘compromise’ between privacy and public safety as his company faces bipartisan backlash. The Washington Post identified 46 cases where police used Flock’s license plate readers for unauthorized purposes, including stalking domestic partners. Flock cut its default data retention from 30 days to seven and now requires a case code before accessing records — but both changes can be overridden via ‘Evidence Mode.’ The ACLU called the reforms a possible ‘step in the right direction,’ while noting they fall short of its recommended 48-hour retention limit.
Sources: TechCrunch
RTK Positioning and AI Are Bringing Autonomous Mowers to Half of American Lawns
A convergence of RTK positioning, 4G connectivity, AI, vSLAM, and LiDAR navigation is making autonomous robot mowers viable for mainstream households, argues Sunseeker Robotics CEO Terry Ma in The Robot Report. RTK technology — proven in precision agriculture — now delivers centimeter-level accuracy at consumer-grade costs, enabling safe navigation of complex suburban lawns. Sunseeker, which raised a $74.3 million Series B in early 2026, operates in more than 40 countries and 5,000-plus retail stores. Ma projects that one in two U.S. households could own a robotic mower within eight years, citing 6.5% annual growth in the U.S. landscaping industry from 2020 to 2025.
Sources: The Robot Report
Tech Pulse
Top Frontier Models: Claude Mythos 5 (80.3%) | Claude Fable 5 (80.3%) | Claude Opus 5 (79.2%)
Top Open Source Models: Qwen3.8 Max (67.7%) | GLM-5.2 (62.1%) | Qwen3.8-27B (61.7%)
Top Small Models (15–50B): Qwen3.8-27B (61.7%) | Muse Glimmer (51.2%) | GLM-4.7 (94.2% HumanEval)
Top Edge Models (0–15B): Phi-4 14B (84.8% MMLU) | Phi-4-mini 3.8B | Qwen3-8B (63.33% AIME24)
AI Leaders: NVIDIA $5.06T | Alphabet $4.55T | Microsoft $3.69T
Robotics Leaders: Intuitive Surgical $175.2B | FANUC ~$40B | Unitree ~$9B (private est.)
Every newsletter preserved and searchable
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.