Samwise Tech/AI/Robotics Newsletter
Monday, April 27, 2026
OpenAI Releases GPT-5.5, Targeting Agentic “Super App” Use Cases
OpenAI released GPT-5.5, its most capable model to date, positioning it as the foundation for an AI “super app” that can complete multi-step tasks autonomously across software tools. Available from April 24 to Plus, Pro, Business, and Enterprise users, GPT-5.5 excels at code debugging, online research, data analysis, and operating software end-to-end. In API access, the model is priced at $5 per million input tokens and $30 per million output tokens, with a one-million-token context window. OpenAI says GPT-5.5 leads 14 industry benchmarks, compared with four for Anthropic’s Claude Opus 4.7.
Sources: TechCrunch
AI Models Subliminally Transmit Biases When Training Other Systems, Nature Finds
AI models can pass hidden biases to other AI systems trained on their outputs, even when data has been scrubbed of explicit signals, according to a Nature study. The process — called model distillation, where a student model trains on teacher-generated data — transfers unintended traits including preferences and potentially unsafe behaviors. Researchers found that teacher and student models must share the same underlying base model for the bias transfer to occur. As AI-generated datasets increasingly replace human-curated training data, the findings raise urgent concerns about safety and reproducibility in large language model development.
Sources: Nature
Human Scientists Outperform Best AI Agents on Complex Research Tasks
Human scientists significantly outperform the best AI agents on complex, multi-step scientific tasks, according to a Nature analysis published April 23. While AI agents can autonomously execute steps including literature search, experimental design, and data analysis, they scored roughly half as well as PhD-holding human specialists on representative workflows. The gap points to limitations in AI agents’ ability to handle ambiguity, revise plans when results are unexpected, and maintain context across extended chains of reasoning. The finding has direct implications for researchers and technology companies promoting AI scientists as near-term replacements for human scientific labor.
Sources: Nature
Meta Signs Deal for Millions of Amazon Graviton CPUs to Power AI Agent Workloads
Meta has struck a deal to use millions of Amazon’s Graviton CPUs to power its growing AI workloads, a move that signals a shift away from GPU-centric compute for agentic AI applications. While GPUs remain standard for training large language models, AI agents performing real-time reasoning, code execution, and multi-step task coordination require a different compute profile — one where AWS Graviton processors offer cost and efficiency advantages. The agreement is part of a broader AWS partnership and marks a notable divergence from Meta’s previous reliance on NVIDIA hardware. The chip market’s fragmentation is accelerating as AI deployment patterns diversify.
Sources: TechCrunch
Vercel Breach Exposes OAuth Gap as Enterprise AI Tools Bypass Security Teams
A security breach at developer platform Vercel has exposed a systemic vulnerability in enterprise AI security: OAuth permissions granted to AI tools by employees, often without security team awareness. The intrusion, disclosed Sunday, traces to a February 2026 Lumma Stealer infection on a Context.ai employee’s machine, giving attackers access via broad Google Workspace OAuth scopes. Vercel CEO Guillermo Rauch described the attacker as highly sophisticated and likely AI-accelerated. Most enterprise security teams lack visibility into which AI tools employees have connected to corporate cloud accounts. VentureBeat analysis found that nearly a month elapsed between Context.ai’s detection and Vercel’s public disclosure.
Sources: VentureBeat
Anthropic and NEC Partner to Build One of Japan’s Largest AI Engineering Organizations
Anthropic has partnered with Japan’s NEC Corporation to build one of Japan’s largest AI-native engineering organizations, deploying Claude to roughly 30,000 NEC Group employees worldwide. NEC becomes Anthropic’s first Japan-based global partner and will use Claude Code as its primary AI development tool. The partnership includes an internal Center of Excellence to develop AI-enabled engineering capabilities, backed by training and enablement from Anthropic. NEC will also expand Claude across internal business operations under its “Client Zero” initiative, in which NEC serves as its own first customer before offering technology to external clients.
Sources: Anthropic
Reliable Robotics Raises $160 Million for FAA-Certifiable Autonomous Aircraft System
Reliable Robotics, a Silicon Valley startup developing fully automated aircraft systems, raised $160 million led by Nimble Partners. The round will accelerate deployment of the Reliable Autonomy System, the first Federal Aviation Administration-certifiable technology enabling pilotless cargo aircraft operation. Strategic backers include AE Ventures, a partner of Boeing, and RTX Ventures, alongside existing investors Eclipse, Lightspeed, and Coatue. In March 2026, Reliable and Albuquerque were selected for the FAA-led Advanced Air Mobility Integration Pilot Program, run by the U.S. Department of Transportation to accelerate autonomous regional cargo operations. Commercial autonomous cargo flights are planned for later in 2026.
Sources: The Robot Report
Tech Pulse
Top Models: Claude Mythos Preview (93.9%) | Claude Opus 4.7 (87.6%) | GPT-5.3 Codex (85.0%)
Market: NVIDIA $4.60T | Alphabet $3.89T | Apple $3.81T
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.