Samwise Tech/AI/Robotics Newsletter
Monday, August 24, 2026
Nvidia’s AVO Harness Lifts Claude Opus 5 to 100% on ARC-AGI-3
Nvidia researchers boosted Claude Opus 5’s score on ARC-AGI-3 from 30% to 100% by wrapping the model in a custom agentic harness called Agentic Variation Operators, or AVO. ARC-AGI-3 is an interactive reasoning benchmark consisting of 2D games with no instructions — models must figure out how to play and win, as a human would. Without AVO, Opus 5 scored 30%, the top result among all models tested. Nvidia published this research suggesting the harness, not the underlying model, is now the dominant factor in long-horizon AI task performance. AVO is built using pieces of Nvidia’s open Nemo platform.
Sources: TechCrunch
OpenAI Reverses Course, Now Calls for Stronger California AI Safety Law
OpenAI is calling for California to strengthen SB 53, the AI safety law it previously opposed. The company says the law “should be amended to expand safeguards,” including requiring monitoring of frontier models under training or evaluation for potential serious incidents and strengthening cybersecurity protections throughout the model-development lifecycle. The reversal is notable — OpenAI previously wrote an open letter to Governor Newsom discouraging him from signing SB 53. OpenAI now describes this approach as “reverse federalism,” saying states should establish core AI protections that can ultimately become the foundation for a national standard.
Sources: TechCrunch
DeepMind Alumni Startup’s Agent Faraday Outperforms Frontier Models at Paper Replication
Inherent, a London AI lab founded by Google DeepMind alumni, says its agent Faraday has outperformed models from Anthropic and OpenAI at independently reproducing published scientific paper findings. The startup, which raised a $50 million seed round just weeks ago, pitted Faraday against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. Faraday runs on a 27-billion-parameter version of Qwen 3.6, making it far smaller than both competitors. Inherent’s definition of success went beyond accuracy — it also required the agent to demonstrate “research taste,” an instinct for which experiments are worth running and how to design them well.
Sources: TechCrunch
Schaeffler to Mass-Produce Humanoid Robot Gearboxes With New Forming Technology in 2027
German manufacturer Schaeffler Technologies AG says it has completed validation testing of strain wave gearboxes for humanoid robots and will begin mass production in Germany in 2027. Actuators — of which strain wave gearboxes are a key component — account for around half the total cost of manufacturing a humanoid robot. Schaeffler’s approach uses forming technology rather than conventional precision machining: a process that creates the required gear geometry in seconds, cuts manufacturing costs by more than 25%, and reduces material consumption by more than 75%. The company has shipped more than 2 million formed strain wave gearboxes across automotive markets over the past 10 years.
Sources: The Robot Report
Nvidia Takes Minority Stake in Cloverleaf Data Center Infrastructure Startup
Nvidia has taken a minority stake in Cloverleaf Infrastructure, a startup founded in 2024 that acts as a middleman between utility companies and data centers, providing power sources and other pivotal infrastructure for site development. Cloverleaf raised $300 million in its first year. The companies did not disclose terms; the Wall Street Journal reports that Nvidia’s investment will likely total several hundred million dollars. The deal extends Nvidia’s ongoing push to use its immense profits to keep the AI flywheel spinning — a strategy the chipmaker has pursued through similar minority positions across the AI supply chain.
Sources: TechCrunch
Starcloud Raises $250 Million Extension for Orbital AI Data Centers as Launch Capacity Tightens
Starcloud, which develops satellites that perform AI inference in orbit, has closed a $250 million extension to its March Series A, valuing the company at $2.3 billion. The extension was led by Manhattan West Ventures and included a $25 million contribution from Nvidia and participation from Cisco, Benchmark, EQT, and others. CEO Philip Johnston is deploying the capital to build a larger manufacturing facility and advance Starcloud-3, the company’s largest orbital data center spacecraft, designed to fly on SpaceX’s Starship. Near-term plans call for two Starcloud-2 compute satellites — each running at 8 kilowatts — to launch on rideshare flights in 2027 for customers including U.S. government agencies.
Sources: TechCrunch
Anonymous ‘Ox Alpha’ Stealth Model Debuts on OpenRouter, Developer Unknown
A mysterious AI model called Ox Alpha appeared on OpenRouter last week, driving widespread speculation about its developer. The OpenRouter listing describes it as a reasoning model designed for coding, sustained agentic work, and production workloads, and notes it is operated by an anonymous third-party provider during a preview. The model is available for free. Stripe CEO Patrick Collison, whose company is acquiring OpenRouter, described Ox Alpha as “very impressive.” Early speculation has focused on GLM models from Chinese AI company Z.ai, though the actual developer has not been confirmed. Much about the model’s origins remains unclear.
Sources: TechCrunch
Tech Pulse
Top Frontier Models (SWE-bench Verified): Claude Opus 5 (96%) | Gemini 3.1 Pro (80.6%) | DeepSeek V4 Pro (80.6%)
Top Open Source Models (SWE-bench Pro): GLM-5.2 (62.1%) | Qwen3.8-27B (61.7%) | Muse Glimmer 30B (51.2%)
Top Small Models (15–50B, SWE-bench Pro): Qwen3.8-27B (61.7%) | GLM-4.7-Flash 30B (55.3%) | Muse Glimmer 30B (51.2%)
Top Edge Models (0–15B, SWE-bench Lite): Gemma 4 7B (67%) | Phi-5 Mini 3.8B (62%) | Qwen3.5 7B (58%)
AI Leaders: Nvidia $5.28T | Alphabet $4.19T | Microsoft $3.60T
Robotics Leaders: Intuitive Surgical $175B | ABB $165B | Fanuc $18B
Every newsletter preserved and searchable
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.