S3 bucket names no longer need to be globally unique, Brett shares what it actually costs to run a personal AI knowledge system on AWS, and an O'Reilly piece nails why agents make hard problems harder. Plus: Sora is dead, OpenAI bought Astral, and Claude Code gets Auto Mode.

Topics covered:

- AWS MCP Server preview adds CloudWatch monitoring — and how Brett used the announcement to build observability for OpenBrain in 10 minutes

- Amazon Bedrock adds GLM 5 and Minimax M2.5 — cheaper model options for Bedrock workloads

- S3 account-regional namespaces end 20 years of globally unique bucket naming

- OpenBrain real-world costs: ~$30/month with Aurora Serverless, fck-nat, and Lambda

- LiteLLM supply chain attack — a security tool was the attack vector

- OpenAI shuts down Sora, Disney pulls $1B investment

- OpenAI acquires Astral (uv, ruff, ty) — same playbook as Anthropic buying Bun

- Codex sandbox limitations for real-world agent workflows

- Claude Code Auto Mode — a classifier between manual approval and YOLO mode

- Claude Code Channels — text your agent from your phone via Telegram or Discord

- DeepSeek Engram — a static memory lookup layer that outperforms MoE at 27B parameters

- Mamba-3 — smaller state space model matching Mamba-2 quality at half the state size

- The Mythical Agent-Month — agents eliminate easy work but amplify complexity

- Building personal developer infrastructure with CLAUDE.md and session context

- State of the models: GPT 5.4, Gemini Flash, Minimax M2.7 hitting price-performance sweet spots

Full links and show notes at curiousorbit.com

This week, Brett and Travers welcome David Bauer - a guy who's been programming since 1982, got his AWS account directly from Jeff Bezos in 2005, and ran a parallel simulation on 9.9 million CPU cores. The conversation covers his journey from building the NSA's first cloud in portable data centers to his current work embedding LLMs inside MCP servers for decision governance.

Topics covered:

• David ran a BBS called Texas Outlaws in the 1980s and has been in distributed computing ever since.
• The NSA's cloud computing environment started in tractor-trailers flown to Afghanistan on Antonov cargo planes.
• David got his AWS account from Jeff Bezos before the platform publicly launched - he still has it.
• His PhD work produced a parallel discrete event simulator that ran on nearly 10 million CPU cores at Lawrence Livermore.
• During COVID, his team modelled 5-6 billion people at the individual level to plan PPE and medical staff distribution.
• His current work puts LLMs inside MCP servers to enforce decision governance, audit trails, and data access controls.
• Federated learning with homomorphic encryption lets them train across data silos without centralizing sensitive data.
• The hosts discussed the scaffolding problem — getting AI agents to reliably follow instructions every time, not just three out of five.
• Mamba 3 models released under the Apache license, focused on efficient inference on smaller hardware.
• AWS/Kiro added Qwen, Minimax, and other new models to Bedrock, alongside those from Anthropic.

Full links and show notes at curiousorbit.com

Brett and Travers cover the week in AWS and GenAI, anchored by a cautionary tale about a developer who gave Claude Code full Terraform access and destroyed their production database. They also dig into whether AI tools are collapsing the junior developer pipeline, and why shipping bad code with agents is a choice.

- Custom .NET MCP servers for Amazon Q Developer — and the ongoing confusion about Q vs. Kiro.

- AWS Load Balancer Controller reaches GA for Kubernetes Gateway API — relevant for large multi-tenant orgs.

- Database Savings Plans expanded to OpenSearch and Neptune.

- Simplified IAM role creation directly from service consoles.

- GameLift gets free DDoS protection — essentially Shield for UDP traffic.

- Bedrock AgentCore Runtime now supports stateful MCP servers on isolated micro VMs.

- Brett's OpenBrain memory system is working — MCP retrieval is solid, Slack integration needs work.

- Early impressions of Kiro's spec-driven development for a personal Kanban board.

- How one developer dropped their production database using Claude Code with Terraform.

- Simon Willison on why AI tools aren't locking us into boring technology stacks.

- The junior developer pipeline question: $100/month AI vs. $80K/year junior dev.

- AI should help produce better code — if it doesn't, that's a process failure.

Full links and show notes at curiousorbit.com

The #CancelChatGPT movement triggered a 295% uninstall spike and pushed Claude to #1 on the App Store following OpenAI's Pentagon deal — and researchers from Stanford, MIT, Harvard, and CMU published a paper showing they could break Claude Opus 4.6 in agentic workflows, achieving shell access and running agents for nine days without human intervention. Brett closed the episode with a full walkthrough of his OpenBrain project: a personal second brain built on AWS by Claude — 80 GitLab pipelines, 25 merge requests, zero lines of code written by a human.

• OpenClaw is now available pre-built on Amazon LightSail with Bedrock integration — watch the token costs before you launch it.
• #CancelChatGPT: 295% uninstall spike and Claude jumping to #1 in the App Store after the OpenAI-Pentagon deal.
• OpenAI and Amazon announced a $138 billion partnership, with OpenAI models coming to Bedrock and a new platform for managing AI agent fleets.
• One month running a personal OpenClaw instance — the $200 token bill arrives faster than you'd think.
• Recreating OpenClaw's messaging features: the messaging layer is the hardest part of building autonomous agents.
• BullshitBench ranks which models push back on unanswerable questions — Anthropic at the top, GPT-4o-mini at the bottom.
• Gemini 3.1 Flash Light: 2.5x faster than its predecessor at 60 cents per million tokens.
• LLMs can match 67% of Hacker News pseudonyms to real LinkedIn profiles — burner accounts are effectively dead.
• Key leads on Qwen Coder and Qwen VL resigned from Alibaba; the team appears to be dissolving.
• Agents of Chaos: researchers from four universities broke Claude Opus 4.6 in agentic workflows — shell access, data exfiltration, nine-day autonomous runs.
• The Five Levels from Spicy Autocomplete to Dark Software, and Brett's OpenBrain project built entirely by Claude on AWS.

Full links and show notes at curiousorbit.com

Over half of Brett's website traffic is now ChatGPT—the new AWS WAF AI dashboard makes it painfully visible.

This week: 

• AWS WAF AI activity dashboard reveals your bot traffic breakdown with one-click blocking. 
• Frontier Security Agent adds cross-account VPC pen testing via RAM. 
• Kiro Powers bundles MCP servers into installable packages—IDE-only —and the hosts are skeptical. 
• Gemini 3.1 Pro benchmarks on par with Opus 4.6 but struggles in CLI usage. 
• Sonnet 4.6 matches human performance on real-world computer tasks. 
• Cline was compromised via NPM for 8 hours—supply chain security matters. 
• Kiro allegedly took down Cost Explorer. 
• And the Citrinin 2028 labour displacement paper triggered a $200B stock sell-off—the hosts break down what it gets right and what it misses.

Full show notes, links, and Engineering Rundown at curiousorbit.com