Samwise Tech/AI/Robotics Newsletter
Friday, September 18, 2026
Agility Robotics Unveils Digit 5 Humanoid as Orders Top $300 Million
Agility Robotics unveiled Digit 5, the latest generation of its commercial humanoid robot, as multi-year customer orders surpassed $300 million. The new robot increases payload capacity by 40 percent over Digit 4, introduces interchangeable end effectors, and adds faster autonomous charging alongside a redesigned safety architecture enabling close operation alongside humans. Digit 5 is designed to expand beyond the tote-handling work that characterised earlier deployments and eventually perform depalletizing, machine tending, kitting, quality inspection, and palletizing across manufacturing and logistics environments. The announcement comes as Agility prepares to enter public markets through its proposed merger with Churchill Capital Corp XI, with shares expected to trade under the ticker AGLT.
Sources: Robotics & Automation News
FAA Plans $875 Million AI Investment to Fix Air Traffic Control Shortage
The Federal Aviation Administration will invest $875 million in an AI-powered air traffic management system called SMART — Strategic Management of Airspace, Routes, and Trajectories — to address a persistent nationwide shortage of air traffic controllers. The automated software program is designed to help streamline routing decisions and reduce controller workload across the national airspace. The FAA has faced years of staffing shortfalls, with the agency estimated to be hundreds of controllers short of its target workforce. SMART represents the agency's most significant bet yet on AI-assisted automation to supplement human operators and improve the efficiency and safety of a system handling millions of flights annually.
Sources: TechCrunch
PrismML Compresses a 27B Model to 5.9 GB With Near-Zero Performance Loss
Startup PrismML released Bonsai 2 27B, a compressed version of Qwen3.8 27B that shrinks the model down to 5.9 gigabytes while retaining 98 percent of its aggregate benchmark performance. The company's proprietary compression technique preserves near-original capability at a fraction of the storage and compute cost, making the model small enough to run on consumer PCs and smartphones. PrismML says the approach addresses a key barrier to widespread AI adoption: most capable models require data-centre hardware to run. Bonsai 2 27B is available for local deployment and is positioned as a privacy-friendly alternative for users who prefer not to send queries to cloud-based services.
Sources: TechCrunch
Google Dream-RSI Cuts Agent Discovery Calls 162x Without Retraining Models
Google researchers, working with teams from the University of Maryland and University of Virginia, published Dream-RSI, a system that enables AI agents to improve their problem-solving strategies without retraining the underlying model. The approach allows agents to replay previous exploration runs as simulators, scoring alternative discovery strategies against recorded outcomes rather than re-executing them. In tests using Gemini 3.1 Pro, Dream-RSI reduced discovery-agent calls by up to 162 times compared to a baseline system, cutting average runtime from 3,587 to 2,931 milliseconds. The research represents a step toward agents that accumulate capability across sessions without requiring model weight updates.
Sources: VentureBeat
UN and Google Build AI-Searchable Platform for Global Statistical Data
The United Nations announced a partnership with Google to make its vast statistical holdings searchable by AI agents through the UN System Data Commons, built on Google's open-source Data Commons platform. The system allows users and AI systems to search for statistics across multiple UN agencies simultaneously using natural-language queries, replacing a process that previously required navigating separate agency databases. The UN collects data on hundreds of global indicators spanning health, climate, poverty, and development. The partnership represents a significant shift in how intergovernmental statistical data is made accessible, positioning UN datasets to be consumed directly by AI agents rather than only by human analysts browsing individual agency websites.
Sources: TechCrunch
Comp AI Raises $34 Million to Build Agentic Security and Compliance Platform
Comp AI announced a $34 million Series A funding round led by Roo Capital and Grand Ventures to build an AI-native platform for continuous security and compliance monitoring. The company's product is designed to operate as a perpetually running agent that scans infrastructure, identifies compliance gaps against frameworks such as SOC 2 and ISO 27001, and generates remediation steps without requiring dedicated compliance staff to conduct periodic manual audits. The round brings Comp AI's total funding to $41 million. The company competes in a growing market of agentic compliance tools as enterprises face increasing regulatory scrutiny of their AI deployments alongside traditional cybersecurity and data-protection obligations.
Sources: TechCrunch
Is the AI Safety Debate Really About Safety — or Who Controls Frontier AI?
A growing debate within the AI industry is questioning whether calls for tighter AI safety measures are primarily about preventing harm or about controlling who can access and develop frontier AI systems. Critics argue that some safety proposals advanced by leading labs would disproportionately benefit incumbents by raising barriers that exclude smaller competitors and open-source developers. Proponents counter that the risks posed by increasingly powerful models justify coordinated governance. The debate has intensified following Anthropic CEO Dario Amodei's essay calling for the industry to slow frontier AI development, which drew pushback from researchers who viewed the framing as conflating genuine safety concerns with competitive strategy.
Sources: TechCrunch
📈 Tech Pulse — September 18, 2026
| Top Frontier Models | Claude Opus 5 (63%) | Claude Fable 5 (62%) | GPT-5.6 Sol (61%) |
| Top Open Source | Kimi K3 (88.3% Term-Bench) | DeepSeek-V4-Pro (80.6% SWE) | Qwen3.6-27B (77.2% SWE) |
| Top Small (15–50B) | Qwen3.6-27B (77.2% SWE) | DeepSeek-V4-Flash (24.1 Arena) | Mistral Small 4 |
| Top Edge (0–15B) | Phi-3.5-MoE (6.6B active, >GPT-3.5) | PrismML Bonsai 2 27B (5.9 GB) | Gemma 3 1B |
| AI Leaders | NVIDIA $5.1T | Microsoft ~$3.4T | Alphabet ~$2.8T |
| Robotics Leaders | Intuitive Surgical ~$200B | ABB ~$90B | FANUC ~$35B |
Every newsletter preserved and searchable
Curated by JD · samwise.agency