Lambda MicroVMs, a Runaway Bedrock Bill, Midjourney Does Ultrasound

Top-Line Summary

The most useful thing in this episode is a failure. Brett’s OpenBrain synthesizer ran successfully on its first night and never again, timing out silently while EventBridge dutifully retried it three times per run, and the only alarm that ever went off was the Bedrock bill going from about five dollars a month to fifty. Around that: Lambda now runs microVMs, AWS Blocks is a TypeScript framework that smells a lot like Beanstalk, a long list of services including Bedrock Agents and Kendra goes into maintenance mode on July 30th, and Midjourney announced it can do full-body ultrasound scans.

Show Video

The “Pre-Show” Context

Back after a two-week gap, one week of sickness and one week of sitting on his ass, which everyone needs occasionally. Brett’s AC unit died mid-week during a 40 degree stretch, and he only figured it out after a night of both him and his wife saying “is it warm in here?” and blaming the weather. The unit is old enough that a replacement was the expected outcome, but the techs topped up the coolant, swapped a capacitor, and the vents started blowing cold again. Twelve hours of a very warm house, crisis averted.

The Engineering Rundown

• Lambda microVMs (02:28)

  Firecracker microVMs are now available in Lambda, and the announcement leans heavily on VM-level isolation for AI workloads, which tells you exactly who AWS thinks the buyers are. Both hosts landed on the same use case immediately: an isolated sub-agent, or a tool call you hand to an agent that you want to run exactly once, idempotently, with nothing leaking out. Travers had already seen someone flip their GitLab runners over to microVMs. Brett has something in the back of his head for these, though Claude has apparently been trying to talk him out of it. Beyond that there’s not a lot to say. You can run microVMs in Lambda now. Go to it.

• Bedrock AgentCore adds WAF support (04:00)

  You can now attach a web ACL to an AgentCore gateway, alongside the pile of other things WAF already fronts. The take is short because the news is short: if your agent is public-facing and it supports a WAF, put a WAF on it. The broader observation is that everything AWS ships this year is AgentCore, and the running joke by the end of the segment is that AgentCore will eventually absorb every other service on the platform.

• AWS Blocks (04:51)

  An open-source TypeScript framework for composing application backends on AWS: reusable blocks of infrastructure that live inside your TypeScript project rather than in a separate IaC repo. Travers immediately clocked it as the same idea as SST, which surfaced around the time the CDK was maturing and Terraform had fully sunk its roots in, and which nobody has heard much from lately (probably because it’s the same crew behind OpenCode). You don’t need an AWS account to start, which took Brett a minute to absorb, and you can drop into the CDK if you want but the docs say you shouldn’t need to. Brett’s analogy is Beanstalk: you tell it “this is a Python app, make it look like this,” and the infrastructure is somebody else’s problem. Load-balanced web app pattern, static site pattern, whatever.

• AgentCore memory now supports cross-account access (07:44)

  Attach a resource policy, and other accounts can read the memory. Neither host has a use for it at the scale they’re working at, but the shape of the thing is obvious enough: a central memory store with multiple solutions talking to it, which starts making sense in large implementations or when you’re spread across continents and want the memories reachable from everywhere.

• Secrets Manager gets safer agent access (08:47)

  The pitch is that an agent can retrieve the secret it needs without the secret ever landing in the context window. Brett has suffered here. He watched Claude print what looked like a GitLab access token to context, asked it whether that’s what he’d just seen, got an apology, told it to continue, and watched it print another one. Then a third. The tooling fix has worked better than the instruction fix: he’s moved everything to 1Password service accounts and given the agent tools that fetch credentials, rather than a paragraph in CLAUDE.md saying NEVER WRITE SECRETS in capital letters and hoping. Travers has done the same thing with deterministic scripts and tuned runbooks. It still does weird shit occasionally.

• AWS service availability updates, June 30th (12:33)

  A quiet post with a long list. Going into maintenance mode on July 30th and closed to new customers: Bedrock Agents (now retroactively “Bedrock Agents Classic,” which prompted a Coke Classic / New Coke riff), Amazon Kendra, Amazon Q for Business, IoT Device Defender and Detect, AWS Directory Service Simple AD, Mainframe Modernization (which appears to have become AWS Transform), and a stack of SageMaker features presumably being folded into Studio. Brett’s Bedrock Agents memory: doing a customer demo and discovering a hard cap of five agents total, one orchestrator and four sub-agents. It worked for the demo, and it dates the thing perfectly, because nobody is capping at four sub-agents in 2026. The real complaint is the compounding one. AWS has always been a double-edged sword, a thousand Lego pieces you assemble yourself, and now on top of the sprawl you have deprecations and a rebranding campaign running simultaneously. Kendra probably lost to knowledge bases. Q for Business lost to whatever Quick Suite is called this week.

• Amazon Quick Desktop, free preview over (18:17)

  Brett fired it back up after somebody asked him about it, and got bounced straight to a subscription page. The graph view of your calendar and email contacts was genuinely slick, the kind of thing you could drill into and see relationships. It is not twenty dollars a month slick. He can build his own graph with a markdown file and a skill in his second brain repo. Right-click, close, move on with the day.

• Migrating off GitHub to self-hosted Gitea (20:52)

  Travers got sick and tired of GitHub, specifically Actions, and started moving his personal projects to a Gitea server on AWS. It’s a single Go binary, SQLite underneath, nightly dumps to S3, running on a very small instance, with Fargate Spot for the runners. He’s sympathetic to GitHub being under real pressure from the flood of agent-generated traffic, right up until he remembers it’s a multi-billion dollar company sitting on one of the largest computing platforms on the planet. Host your own. The tangent that follows: Microsoft’s 4,800 layoffs hitting the Xbox side, Sony phasing out physical media, and both hosts calling time of death on consoles while predicting a Pirate Bay revival out of sheer consumer spite.

• The rest of Travers’ infra rebuild (24:36)

  His agentic brain system is moving off his local machine onto AWS. He’s teetering on Postgres Serverless and currently thinks it’s overkill, so it’s all on one node for now. A nano instance runs fck-nat so he doesn’t have to pay for a NAT gateway, and Caddy handles the load balancing in front. Also consuming his life for the past twenty days: a Kaggle competition to train a bot to play the Pokémon trading card game, which is approximately the most Travers project imaginable.

• The /advisor flag in Claude Code (26:01)

  An experimental feature Brett was reminded about in a Discord thread. You designate a model as the advisor, and Claude escalates to it when it’s unsure about something, so you can run a cheaper model by default and pay for the expensive one only when the agent asks. The person who mentioned it was running Opus 4.8 as advisor over a Sonnet default. Brett flipped it around and set Fable as the advisor with Opus 4.8 doing everything else, as a way to taste Fable without getting throttled into a four-hour timeout. He hasn’t actually caught it making an advisor call yet.

• Fable as orchestrator, Codex as sub-agent (28:52)

  Travers has rewritten his CLAUDE.md into a delegation guide: here are the classes of task you can hand to other models. Fable orchestrates, breaks the work down, and farms it out to Opus or Sonnet. Mechanical-but-still-demanding coding tasks get broken off into Codex, which he invokes through the command line as a sub-agent. The economics are the whole point. Fable credits are scarce and Fable is genuinely good at orchestrating long-running work; Codex credits are, for now, generous. His subscription’s Fable access runs out around July 12th, after which it’s pay-per-use, and he’s banking on GPT-5.6 being good enough to cut over to.

• Multi-agent verification and the hallucination problem (30:52)

  Brett watched a video of someone running a multi-agent orchestration with reviewer agents verifying the output. His question, which nobody answered: if the first agent hallucinated something plausible, why would the reviewing agent flag it rather than accept it as context? Roll the dice enough times and one of them catches something, sure, but how do you determine which of the outputs is the hallucination? Nobody has a good answer.

• NAT gateways are a racket (33:08)

  Brett uses fck-nat everywhere on personal projects now, and has started putting it into customer non-prod environments too. Roughly thirty-five dollars a month for a machine that sits there doing nothing, and if you follow best practice you run two, so it’s seventy US a month before data processing charges. Put it in an auto scaling group at min/max/desired of one, poke the instance in the eye, and you get a two-minute outage while it reboots. He’s fine with that in non-prod. The sting is legacy environments: organizations that have been on AWS for a decade end up with sprawling VPC estates, each with its own public egress, and you look up and there are a hundred NAT gateways burning cash. Centralizing ingress and egress is the right answer and also a huge project, because everything is already wired through them. You don’t just wake up Monday and shut them all down.

• Vacation project: a heavily modded Valheim server (36:16)

  Brett rented a Valheim server from a gaming host, tried to add mods, and hit an AI support chatbot that confidently told him to click a link that did not exist in his management console. Back and forth, back and forth, until an FAQ page revealed you can only mod it one specific way, which was not the way he wanted. He cancelled and spent a day of his vacation running it on AWS instead. Frank, who has north of 800 hours in the game, now refers to it as the murder server. There are so many mobs that they resorted to the dev kill all command out of pure rage.

• The synthesizer that worked once and billed forever (38:36)

  This is the one worth your time. Brett’s OpenBrain synthesizer rolls episodic memories up into what Claude named the “state of play,” a name he hates and wants to replace (the internet is calling it a dream cycle, which he hates slightly less). He logged into the AWS account on a whim and found Bedrock costs had gone from roughly five dollars a month to thirty, on pace for fifty. The synthesizer ran successfully the first night and has never worked since. It times out. EventBridge fires it, gets the failure back, and retries three times, and each timeout burns a pile of tokens before it dies. The reason it stayed invisible for a month is the good part: every morning Slack tells him which memories are unclassified, he classifies them, and the next day the line is gone, so the system looks like it’s working. The database is tracking everything. The actual work isn’t happening. The only place you’d ever see a state of play is a Kanban card, and he was never going to open a Kanban card for the Talking Cloud project. The fixes are all boring: an EventBridge DLQ so a failed job stops instead of retrying three times, a notification off the DLQ, a CloudWatch metric filter on the failure log line, cost anomaly detection, and Bedrock cost allocation tags on the Lambda execution roles so Cost Explorer can tell him whether the money went to the synthesizer, the ingestion pipeline, or the classifier. He’s also had Claude maintain an AS-BUILT.md and a RUNBOOK.md on every merge, and forced himself to actually read them, because these projects have gotten big enough that he now has moments of “fuck, how does that work again.”

• A DocumentDB cluster that DDoS’d itself (48:13)

  Same week, different lesson. A cluster reported as down was running fine. A container was failing to authenticate against it, retrying every ten seconds, and had eaten all thousand-odd available connections. It was sitting in the logs the whole time. A metric filter on the CloudWatch log group would have caught it on day one. The common thread with the synthesizer: the monitoring has to ship with the thing, not get bolted on the day you stumble across the bill.

• GPT-5.6 lands tomorrow, in three flavors (49:44)

  Soul is the top-tier Fable-class model, Luna sits in the middle, and there’s a lower tier below that. Nobody knows how it performs yet. Brett’s note to himself was “how long before it shows up in Bedrock,” which led to the realization that Fable still isn’t in Bedrock, plausibly because of whatever export-control mess got it suspended. Travers has seen a Reddit post claiming otherwise. He’s not going to find out on his personal account’s budget.

• GLM 5.2 in Pi (52:20)

  Fast. Snappy. Both hosts have been running it and both like it. It’s close to Opus 4.8 for coding, not quite there, and it lacks the multimodal path so you can’t paste a screenshot and have it fix your CSS. Factoring in token efficiency it works out to roughly half the cost of Opus. Travers made an error he didn’t catch until later, and shrugged: Claude makes errors too, and in fact Claude is the one that built the synthesizer that’s been quietly burning his Bedrock budget. The genuinely unpleasant part of the experience is watching the dollar counter tick up at the bottom of the terminal when you’ve spent two years on a flat subscription.

• Matt’s new skills, including Wayfinder (54:44)

  Brett already loves the grill-me-with-docs skill. Travers has been through the new batch and rates Wayfinder: you define a long-running task or project, it decomposes it, and it stashes the plans in your GitHub issues so the work steers itself at the project level rather than the prompt level. It works for planning things that aren’t code, too. Brett runs everything through GitLab work items and has his agents pick up cards on their own, so this one’s going straight on his list.

• Midjourney pivots to full-body ultrasound (55:59)

  The strangest pivot either host has seen. The image generation company announced technology for full-body ultrasound scans at resolution they’re pitching as comparable to an MRI. You get lowered into a cylinder of water, a ring of ultrasonic sensors travels up and down pulsing and collecting billions of data points, and each slice gets run through a diffusion model that reconstructs what the internals should look like. It takes about sixty seconds and reportedly a single very large GPU. It is going to launch in a luxury spa in San Francisco, which tells you precisely who it’s for. Brett wants this democratized, because accessible high-resolution full-body imaging for health monitoring would be genuinely great. Both hosts stalled on the same question: where does that data live? An MRI or CAT scan is on the order of a terabyte per patient, and getting that into S3 over a normal link is not trivial. That’s before you get to the part where it’s the most personal information a person has.

• AI coding costs are spiraling: context is the problem, not usage (1:01:09)

  The article’s thesis is that the token spend killing organizations isn’t the coding itself, it’s the rework loop. Bad context in, wrong answer out, “no, not like that, change this, change that,” and around you go. Improve the context and you cut the loop. Brett’s had multiple conversations recently with people burning through their limits in two days and asking what they’re supposed to do now. His objection to the framing: the article says the problem is consuming tokens without generating value, but who judges value? What he finds valuable, you might not. Ultimately it’s whoever signs the cheque, which means a CFO looking at a line item. The analogy that lands is streaming: they trained everyone to binge a season in a night, then took it away, and now you watch one episode of Silo a week and pay more than cable cost. Token pricing is heading somewhere similar. Travers frames it as the zero-to-one hyperscale playbook, the Uber subsidized-ride-until-you’re-hooked move. Brett’s counter is that the ceiling exists, because if the $140 plan became an $800 plan most people would simply go back to the old way of working, and the open-weight models put a floor under how hard anyone can squeeze.

• Is Google dead? (1:07:24)

  Brett is writing Google off, which he acknowledges is a crazy sentence about the company that invented the technology. They showed up with a pile of models, everyone got excited, and then it went quiet again. Travers still rates Gemini and the Flash models for multimodal and visual work, and thinks they’ve slipped everywhere else. The Mozilla comparison writes itself.

• GPT Live and the future of voice (1:08:31)

  Somebody in Discord shared the GPT Live promo video, described as giving total Golden Girls vibes: three older women having a conversation with GPT on a phone. What makes it remarkable is the overlap. The model starts responding while the human is still finishing her sentence, the way people actually talk, which is exactly what the Modulate AI guys were chasing when they came on the show. Brett is enthusiastic about the video and skeptical about the category, because he was burnt by voice once already. He was certain everyone would have an Echo on their desk, and all he does with his now is ask for the weather and put on Spotify, and in exchange Amazon serves him ads on the screen. That said, he’s talking to Claude all day instead of typing (his wife keeps asking who he’s on the phone with) and loves the interface. On the phone it falls apart: too much latency, no acknowledgement, you’re standing there asking “did you hear me?” The remote terminal, on the other hand, is genuinely useful for working while out and about.

Off-the-Clock Recommendations

• Obsession (2025) — Travers watched it on holiday. Reportedly made for around $750,000 and has done roughly $400 million at the box office, entirely on grassroots word of mouth.
• Hokum — a classic haunted hotel horror flick, set in Ireland. Travers liked it better than Obsession, and it’s near the top of his year.
• Gene Hackman’s 1970s output — Brett has been working through the spy and military thrillers. The man has well over a hundred credits, and some of them are bad in a fun way.
• Enshrouded — Valheim-adjacent survival crafting with spells and RPG systems. Version 1.0 is expected this year.
• Infection Free Zone — top-down zombie survival resource management, and it builds the map from real street data. Put your headquarters in your actual house and try to survive the apocalypse on your actual street.
• Valheim — the heavily modded version, if you want to be murdered repeatedly. Version 1.0 leaves early access around September or October.