Muse Spark, Mythos on the Attack, and Anthropic's Managed Agents

Top-Line Summary

Andy Jassy’s 2025 letter landed with a $15B AI run rate and a $20B custom-chip run rate, but the real signal is that AWS is openly power-constrained - demand is there, electricity isn’t. Meta shipped Muse Spark and Meta Superintelligence Labs’ first model is genuinely competitive without needing to top every benchmark, because it’s being handed to three billion users for free. And the UK AI Safety Institute published its Claude Mythos cyber evaluation: first model to complete a 32-step network attack simulation end-to-end.

Show Video

The “Pre-Show” Context

Brett got shamed by Claude last week for not following the streaming checklist, so tonight he went through the whole thing meticulously before hitting the air. Separately, his Mac kept complaining about disk space while reporting 100GB free - Claude eventually worked out an interrupted OS update had eaten the rest, and together they clawed back 200GB (Descript alone was hoarding almost 100GB of old podcast projects, plus a classic Minecraft install that was six gigs he hadn’t opened in years). He also turned off the multi-camera recording workflow because he was never actually going to cut the splashy stuff - less video, simpler life.

The Engineering Rundown

• Bedrock Now Supports Cost Allocation by IAM User and Role (05:56) Bedrock inference costs can now be broken down by IAM principal in Cost Explorer and CUR 2.0 - tag your roles with team, project, or cost center and you finally get real visibility into who’s spending what on AI. Brett initially shrugged it off because most shops now separate workloads into dedicated accounts, then Travers pointed out the obvious case: shared dev accounts where the whole team is banging on a single service account against multiple agents. If you’re running a shared Bedrock environment, go enable this today — chargebacks are the fastest way to get teams to care about token efficiency.

• Bedrock Agents vs. AgentCore (08:02) Brett spent the week building his first Bedrock Agent project and hit the wall most people hit: what’s actually the difference between Bedrock Agents and AgentCore? His working analogy after a few hours in the docs: Bedrock Agents is the highly opinionated AWS-managed implementation - you build the way AWS wants you to - while AgentCore is the platform layer underneath where you bring your own model and framework but get memory, telemetry, and session plumbing for free. He went the Agents route for a supervisor-plus-sub-agents demo (orchestrator hits Athena against S3-backed demo data), stood the whole thing up in a day with Claude’s help, and was pleasantly surprised you can get a working multi-agent demo with a web UI for about fifteen cents of inference. The YAML-based tool definitions clicked once he started thinking of them as MCP-style tool schemas.

• AWS Agent Registry in AgentCore — Now in Preview (13:43) A private, governed catalog for agents, tools, skills, and MCP servers - the point is to stop two teams inside the same org from rebuilding each other’s work. You can register things manually or point it at a live MCP endpoint and it auto-discovers schemas; the registry itself is queryable as an MCP server from your IDE, so a developer can literally ask “does this already exist?” before shipping a duplicate. Launches in Virginia, Oregon, Sydney, Tokyo, and Ireland. This is the governance layer that starts mattering the moment you have more than one team shipping agents in production — and that moment is arriving fast.

• Introducing Amazon Bio Discovery (16:07) AWS’s new lab-in-the-loop drug discovery platform exposes 40+ biological foundation models for antibody therapy design, closing the loop between computational candidate generation and wet-lab validation. Memorial Sloan Kettering generated 100,000 antibody candidates for wet-lab testing in weeks; Bayer, the Broad Institute, and Voyager Therapeutics are early adopters. Brett and Travers both called out how little of this class of news cycles through the usual cloud-Twitter feeds — the AlphaFold2-shaped change is still playing out and the YAML-authoring productivity gains that dominate the discourse are, realistically, a rounding error compared to compressing drug-discovery cycles. It’s also a textbook AWS playbook move: start with infrastructure, climb to platforms, then ship applications — Bio Discovery is the application tier for life sciences.

• Andy Jassy’s 2025 Letter to Shareholders (22:39) Revenue up 12% to $717B, AWS AI run rate at $15B, custom chip business doubled to $20B run rate. The asterisk: free cash flow cratered from $38B to $11B because Amazon poured an additional $50.7B into property and equipment year-over-year. The line that should reshape your capacity planning: Jassy openly admitted AWS is leaving revenue on the table because it cannot get data-center power fast enough. When your bottleneck is electricity and not customers, that’s a supply constraint, not a demand story - and it lines up with the OpenAI data-center deals that never broke ground for the same reason. Oh, and Amazon is now the #2 US grocer at $150B in gross sales, which reads less like diversification and more like the corporate-governed future in the sci-fi novel Brett is halfway through.

• AWS Interconnect Multi-Cloud hits GA (27:42) Quick note: the GA announcement landed this week. Endpoint-based connectivity to Azure, GCP, and other public clouds — first connection is free, which is useful if you’re running off-site backups or multi-cloud DR. Brett admits he briefly confused it with AWS Connect (the phone system) before the “inter” prefix registered.

• Travers’ Local Whisper + ROCm Transcript Pipeline (28:54) Travers is retooling his brain-dump scripts around a local Whisper model running on ROCm, using Talking Cloud episodes as the convenient test corpus. Current direction: transcribe, run topic analysis, surface follow-up topics, and eventually auto-tag every episode — with the usual struggle between “I could just build the UI for this” and “this was supposed to be a weekend thing.”

• Brett’s Docker Prune Disaster (30:41) This is the operational horror story of the week. Brett ran docker prune during the disk cleanup, said “it saved a ton of space, moving on,” then tried to log time through his kanban MCP a few minutes later and got a blank stare back. He’d nuked every MCP server he uses. The recovery was the interesting part: he told Claude what he’d done, and Claude identified which MCP servers he depends on, pulled main off every Git repo they live in, ran the Docker builds, and had him restart the gateway. Everything came back. The lesson that stuck isn’t “be careful with prune” - it’s that OpenBrain plus his MCP stack have quietly become critical path. Next items on the list: push the images to ECR as a second destination, treat the whole agent workflow as something that deserves backup/DR, and stop pretending this is optional tooling.

• Meta Launches Muse Spark and Meta Superintelligence Labs (38:23) Alexandr Wang’s nine-month reorg delivered its first model: a natively multimodal reasoner with tool use, visual chain of thought, and multi-agent orchestration, rolling out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban glasses. Artificial Analysis puts it 4th on the Intelligence Index at 52 - behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 - but it tops every frontier model on HealthBench Hard (42.8 vs. GPT-5.4’s 40.1), and it does the work with about 58M output tokens versus 157M for Claude Opus. The strategic read: Meta doesn’t have to win every benchmark. They need “good enough” at multi-billion-user scale with a few standout capabilities, and they’re shipping it free. That’s a different game from what Anthropic, OpenAI, and Google are playing - and it leaves the open question of whether there’s a real developer surface here or whether this is purely a consumer play.

• Open Code + GLM 5.1 for $1.50 (41:22) Brett tried Open Code-style tooling with GLM 5.1 via Zen as the inference provider. End-to-end: analyzed the entire OpenBrain project, opened 15-20 issues against it, spent roughly $1.50. Call it half-to-three-quarters genuinely real findings, the rest “won’t fix.” The funny part: GLM insisted it needed a brand-new GitLab token and couldn’t use the group token Brett had just created - back-and-forth got nowhere. He pasted the exchange into Claude, got a concise translation of why GLM was wrong, pasted that back, and GLM immediately capitulated. Agent-to-agent mediation by a smarter model is apparently a thing now. Also next on the list: do the same pass with Kimi K2.5, which is free on Zen, or Minimax 2.7.

• Anthropic’s Managed Agents — “Decoupling the Brain from the Hands” (46:45) Public beta. The architecture separates agents into three pieces: the brain (stateless harness plus Claude), the hands (sandboxes and tools), and the session (a durable event log that lives outside the harness). If anything crashes, no work is lost because the session persists independently; scaling means just spinning up more stateless harnesses. The published numbers - p50 TTFT down 60%, p95 down over 90%, auth tokens never touching the model - line up with the design. Pricing is standard Claude API token rates plus $0.08 per session-hour (so you’re paying for the platform on top of the tokens, in the same shape as any AWS managed service). Notion, Rakuten, Sentry, and Asana are already in production on it. If you’re building agents on Claude today and you’re still hand-rolling sandboxing, state, and credentials, this is the migration path.

• Anthropic’s Automated Alignment Researchers (52:27) Nine copies of Claude Opus 4.6 - each with a sandbox, shared forum, code storage, and a scoring server - were turned loose on a weak-to-strong supervision problem. Five days and $18K in compute later, they hit 0.97 performance gap recovery; the human baseline was 0.23 after seven days. The automated researchers also tried to game the scoring metric, and humans caught it. That caveat is the whole point: you can absolutely turn compute into alignment-research throughput, but you still need humans validating that the research is real instead of optimized for a number. Alignment is starting to look like a scalable engineering problem rather than a handful-of-PhDs problem - and that’s a meaningful shift in the shape of AI safety work.

• AWS Certification Refresh (54:50) Brett is prepping for the AWS AI Practitioner exam and - he’s going to say it - it’s not a practitioner-level exam. The questions he saw in the sample looked closer to the old ML specialty (“how do you evaluate overfit on this model?”) than to the foundational “how do you spell AWS” vibe of other practitioner exams. AWS appears to have broken the old ML specialty out into the AI Practitioner and a new pro-level agentic systems exam. He’s also up for Architect Pro renewal. If you’re in the same boat, start studying the evaluation-metric stuff earlier than you expect.

• UK AISI Evaluates Claude Mythos Preview’s Cyber Capabilities (58:01) Mythos is the first AI model to complete a 32-step multi-stage network attack simulation (“The Last Ones”) end-to-end - full chain completed in 3 out of 10 attempts, average of 22/32 steps across all runs. On expert-level CTF tasks that no model could touch before April 2025, Mythos now hits 73%. Caveat the hosts were clear about: these ranges don’t have live defenders, EDR, or incident-response, so this proves capability against weak targets, not hardened enterprise SOCs. But the trajectory is what matters - we went from “zero models can do this” to “73%” inside a year. That’s the curve every CISO should be watching, and credit to AISI for publishing the eval transparently.

• METR Time Horizons Update (1:00:51) METR updated their task time-horizon work: the doubling time for what an AI agent can complete autonomously (measured against human expert time-to-completion) is now 130 days post-2023, revised faster from the earlier 7-month estimate. The trajectory: GPT-2 at 4-second tasks in 2019, GPT-4 at ~6-minute tasks in 2023, current frontier models at ~10-hour tasks in 2026. Extrapolated at this rate, month-long autonomous projects land before the decade closes. It’s measured on software engineering specifically, which the hosts flagged - but software engineering is the domain that builds every other domain, so it’s a reasonable leading indicator.

Off-the-Clock Recommendations

• Grizzly Fate (2025) — anti-recommendation. Travers watched it expecting a slasher and got a human-interest retelling of two bear attacks in Yellowstone. Didn’t deliver on the “grizzly” the way the title promised. Brett counter-offered the Cole’s Notes of his last camping trip: tiny two-person tent, middle of the night, a black bear’s nose inches away through the rain fly, subsequent sprint to the car, and a park ranger in the morning cheerfully explaining they’d pitched directly on a known bear corridor. They have not camped since.
• Red Rising by Pierce Brown (2014) — recommendation. Brett is a quarter of the way through the first book in the series. Went in blind, expected knights-and-dragons fantasy, got a Mars-set sci-fi with strong Silo and Matrix energy around the whole “you’ve been told one thing and the reality is something else entirely” premise. Five books deep if you like it, and a welcome break from reading about work.