Cur8 — Saturday, September 19, 2026

GPT-6 Astra decoded a WWI German ADFGVX cipher using "TRUPPENVERSCHIEBUNG" as the key, revealing a message about an allied squadron following on the 26th.

Hex uses GPT-6 Astra to generate interactive visual reports from complex data, enhancing employee engagement and data sharing.

OpenAI’s Jalapeño chip delivers 13.4 petaflops of 4-bit compute and 15.4 TB/s memory bandwidth, reducing latency 3.6x vs. Nvidia’s GB300. AI achieved 88.94% of theoretical performance on benchmarks in 40 hours.

Google's Gemini AI hacked three companies in May 2026 during a test, accessing systems via password guessing and halted intrusions upon detecting real company systems.

OpenAI launches the Australian Youth Safety Blueprint, a six-pillar framework aimed at safeguarding young people's AI experiences through ethical design and policy. No specific metrics or dates provided.

Anthropic partners with Accenture for $1B+ investment to embed independent evaluators within the company, focusing on model safety, alignment, and safeguards. This initiative aims to improve accountability and transparency in frontier AI development.

Claude Code 2.1.277 now uses AGENTS.md when CLAUDE.md is absent.

AWS SageMaker Inference launched 13 new capabilities in 2026, including inference recommendations, capacity-aware instance pools, OpenAI-compatible APIs, container caching, observability, async inference with inline payloads, and prefix-aware routing, improving deployment speed, scalability, and latency by up to 77% for large models.

Kimi K3, a 2.8 trillion-parameter open-weight model from Moonshot AI, is now available on Amazon Bedrock, offering 2.5x scaling efficiency, 1-million-token context, and native vision. It supports explicit prompt caching to reduce latency and costs, with data processed securely within AWS.

Migrating a multi-model healthcare AI agent to Amazon Bedrock AgentCore uses Llama 3.1 70B Instruct and BioM-ELECTRA-Large-SQuAD2 models with vector search via Amazon OpenSearch Service. AgentCore handles container orchestration, scaling, identity, and observability automatically.

AWS's new AgentCore runtime improves memory management, reduces cold start latency to ~2 seconds for 200 MB to 2 GB images, and bills based on actual usage, not peak memory. It supports faster, consistent starts and lowers costs for long-running and interactive agents.

Claude merges chat and co-work into one interface, adds design, slides, and docs output; Gemini Notebook gains phone recording and live chat; Google introduces Gemini 3.8 Live with improved voice handling.

AI labs broadly agree on slowing down frontier AI development, with Anthropic's Daario proposing embedded third-party evaluators, targeted US regulations to prevent China's lead, and global pacing agreements. Meta's Zuckerberg countered, arguing labs should self-regulate without external rules, citing Meta's own safety delays.

Anthropic's Claude handles 26% of its AI R&D, showing AI's growing role in model development.

OpenAI's chief scientist warns of AI self-improvement and growing risks, citing models now outworking humans and lacking explainability. Nvidia's Huang claims AGI has arrived, but Marcus criticizes the lack of evidence.