A weekly conversation about what shipped on AWS, what broke, and what everyone in the GenAI world is actually building. It's separate from the consulting practice. Think of it as a public extension of the reading I'd do anyway.
Darko Mesaros argues that code generation is solved and software engineering is not, and that the real cost of agentic coding is the attention you stop paying while the model works. Also: a Data General Eclipse that needs a 60-amp circuit, an Atari 2600 with 128 bytes of RAM, and Kiro running on a 1977 dumb terminal.
Four more from the recent run. Every episode is in the archive below.
An OpenAI model escaped its eval sandbox and broke into Hugging Face to steal benchmark answers, while 54% of surveyed enterprises report they've already had an AI agent incident. Also this week: a Cost Explorer bug showing $2.5 trillion charges, Kimi K3's frontier-class open weights, and 1Password's new credential integration for Claude.
A Bedrock bill jumped from about a dollar a month to $28, and the culprit turned out to be a UUID parsing bug that timed out a Lambda function and then retried it three times, burning tokens on work that never landed in the database. Episode 52 also covers OAuth arriving for the AWS MCP servers, GuardDuty learning to watch Bedrock, AWS wrapping its own agent primitives in Loom, and the 64 agents that ported Bun from Zig to Rust in eleven days.
A scheduled synthesizer job worked exactly once, then timed out every night for a month while EventBridge quietly retried it three times a run, and the only signal Brett got was a Bedrock bill climbing from five dollars to fifty. Episode 51 also covers Lambda microVMs, AWS Blocks, the July 30th deprecation list that includes Bedrock Agents and Kendra, GLM 5.2 in Pi, and Midjourney announcing full-body ultrasound scans for a San Francisco luxury spa.
A scraper ran ffmpeg through the show's entire back catalog overnight, and the only reason Brett caught it was asking Claude to read the CloudFront logs through a read-only 1Password credential. Episode 50 also covers an AWS Summit where every headline is agent plumbing, the first time a government has yanked a specific frontier model, two Chinese open-weight coders undercutting the frontier on price, and SpaceX buying Cursor for roughly the price of one Amazon.
The whole back catalogue. Each one is a conversation about what actually shipped that week on AWS and in GenAI, and what it means in practice.
Talking Cloud started as OrbitLIVE on Twitch back in October 2023. It ran through the first half of 2024, then went quiet for about eighteen months. It came back at the end of 2025 and has been weekly since. Streamed live Wednesday nights, shipped as a podcast on Friday. Travers Annan co-hosts.
The format hasn't drifted much: ten or so links from the week, what shipped on AWS, what's happening in the broader GenAI ecosystem, and a long-running honest accounting of what we're each building in our personal time. We try to land on what's actually worth your attention, and what's hype that'll be gone next quarter.
It's not part of the consulting practice. We don't pitch services on the show, and the show doesn't run my engagement calendar. If you're a regular listener who wants to work together, the right next step is the same as for everyone else: a paragraph in an email.
Subscribe in your podcast app of choice, or catch the livestream Wednesday nights on YouTube. If a recent episode prompted a question about Bedrock, agents, or AWS in general, the contact form is the place.
