Revolut Self-Hosts GLM and Four Agents Beat Minecraft

Top-Line Summary

Claude Haiku 5.5 on Bedrock can now keep inference in Canada. That sent me to the docs to figure out what a global endpoint actually does with your data. Short version: AWS runs it wherever it has capacity, and gives you a discount for it. The rest of the episode is agents (lots of agents), and Revolut claiming self-hosted GLM cut its AI bill to a tenth.

Show Video

The “Pre-Show” Context

We started late. Technical difficulties, bad lighting, and a face made for radio. I went to a live taping of Pivot at the Elgin Winter Garden Theatre in Toronto, the first theatre in Canada with elevators. They still have operators calling out “going up.” Travers’ wife got tickets to see Critical Role at Scotiabank Arena.

The Engineering Rundown

• Amazon WorkSpaces for AI Agents Adds Linux and Windows 11 (06:20)

  WorkSpaces for agents used to be Windows Server only. Now an agent can get a Linux desktop with root, or Windows 11 with the same apps your staff use, under the same network isolation and IAM controls as a human WorkSpace. Neither of us has used it yet. Travers sees it as the way to hand an agent software with no real CLI, like Blender. My example: a friend who had never touched ComfyUI is now building workflows in it, with Claude watching his screen.

• Kiro Remote Sessions (11:09)

  Someone I was talking to just turned on Kiro’s remote sessions for testing. Kiro runs the session on a managed instance, so you can close the laptop and walk away. My concern is that the session doesn’t carry the MCP and prompt-logging controls an enterprise sets up in its Kiro subscription. It also fits my prediction that we’re heading back to dumb terminals (a thin client on your desk, the horsepower in a hyperscaler).

• Cancelling Kiro and ChatGPT (13:45)

  I cancelled Kiro. I spend nearly all my time in Claude and couldn’t justify $100 a month for 5,000 credits. I was a bit surprised when cancelling. Unlike an EC2 instance, turning it off doesn’t stop the bill; it’s a SaaS subscription and it runs to the end of the month. ChatGPT went too. I’m doing a year-end sweep where every SaaS tool has to answer one question: am I actually using this?

• Claude Haiku 5.5 Inference Stays in Canada (17:13)

  Haiku 5.5 now has a Canadian cross-Region inference profile. It only routes between ca-central-1 and ca-west-1. A client told me recently that their data had to stay in Canada, and my answer was “then you can’t use that.” So I finally read the docs on global endpoints. AWS picks the Region based on capacity, so your request could run in Canada, the US or Japan. And in true Amazon fashion, it’s cheaper (about 10% on input and output tokens, per AWS). If residency matters, that discount isn’t free. AWS also says future models may not come to the Canada profile.

• Lambda Web Functions (19:46)

  People on Bluesky are calling Lambda web functions the Lambda they’ve waited years for. I haven’t tried it once. Travers thinks I should move my personal site to it, though it’s built for concurrent connections (up to 128 per function, he thinks) and my site gets nowhere near that. Most of my client projects require every Lambda in a VPC anyway, so I keep doing it the old way.

• Kaggle Quantum Circuit Optimization (21:06)

  Travers is in another Kaggle competition, this time quantum circuit optimization. He admits he doesn’t really understand it. Claude does, and he’s in the top half of the field so far. It reminded me of the survey data in Ethan Mollick’s new book: most ChatGPT traffic is people learning how to do things, not companionship.

• Learning With Claude (23:56)

  I use Claude for exam prep. Hand it the exam guide, have it build sample questions and flashcards, then run a gap analysis on what I know and what I don’t. Travers’ point is that models are now good enough at games that you can build your own study apps around your weak subjects. He also says Anthropic softened its rule on copying games, so asking for Call of Duty gets you something like it instead of a refusal. Reverse engineering is still off limits.

• Writing Your Own HOTAS Software (25:38)

  I bought a throttle, joystick and pedals to fly helicopters in a sim. It’s very hard. My brain still won’t accept that the pedals turn the opposite way to what I expect. The vendor’s software looks like it hasn’t changed since the first stick shipped, so I’m thinking about an afternoon with Claude writing my own over USB. Travers’ advice if Claude refuses: use Pi with GLM or DeepSeek.

• A $15 Agent Status Light (28:01)

  After seeing Paul’s ESP32 board last episode, I looked at a $70 USD LED strip that plugs into a Mac and shows whether your agent is working or waiting on you. Instead of buying it, I asked Claude to explain how it works and design our own. I bought two programmable boards for $30 CAD (one plus shipping cost the same as two). The plan is to have herdr call the board to change colours, then 3D print a case. As Travers said, it’s never been a better time to be someone who builds things.

• Building Your Own Agent Harness (31:50)

  Travers’ own harness now runs and answers on Discord about as well as the commercial ones. His takeaway: models need less hand-holding as they get better, but context is still hard. When do you split off a task? When is it done? Verification helps with both. He thinks prompting is dying. Give it the goal, how to check success, and why the trade-offs matter, then let it loop. The hardest unsolved part is deciding which actions are too risky to take, which is why he understands Anthropic using a classifier. He built his own to learn, and to wire in his own skills and code-quality research.

• Machine-to-Machine OAuth With Claude (36:04)

  I keep written IaC standards for CloudFormation and CDK (required tags, how we build pipelines). Claude reads them before any work. For a while I’d pop quiz it: what are the required tags? What are the optional ones? On a recent project I needed machine-to-machine OAuth, so I had Claude explain the flow against the existing architecture, with my standing instruction: don’t speculate, read the documentation. Then it built the setup in a sandbox account. I watched it test, fail, work out why and fix it in the background until it handed me the test commands and the logs.

• A Skill That Guards the Kanban Board (41:59)

  My rule: if a human has to do it, it goes on my Kanban board. Everything else goes into GitLab for an agent to pick up. Lately my /go skill started creating cards for coding work, or for things that should just be done now. That’s the agent tasking the human, which is backwards. So Claude and I built a Kanban skill that other skills call to decide whether something belongs on the board or should be sent back and done. Travers sees it as a classification problem, and a skill full of examples handles that well.

• Four Gemini Agents Beat Minecraft (46:15)

  Four Gemini 4 Argon agents beat Minecraft end to end in about 27 hours. They never saw the screen. They played by writing JavaScript through Mineflayer: more than 2,700 programs, 605 deaths and about $810, with researchers watching but never stepping in. The agents shared coordinates for chests, portals and danger, and split up the work themselves (with some early pushback over who did what). The caveats matter. It was Easy difficulty, Mineflayer gives them exact world state no human player has, and it was a single run.

• Pi Durable (50:45)

  A framework for agents that survive crashes. Every step checkpoints, replay-safe tools rerun, and anything else is reported to the model as interrupted. Several people can watch and steer the same agent, which we both liked for team work. People in my Discord are raving about it. Travers already does the same thing in his own harness, with persistent agents acting as smart cron jobs.

• Strata (52:09)

  A one-click local LLM tool (about 40,000 stars) that serves OpenAI- and Anthropic-compatible APIs on localhost, so Claude Code or Codex can point at a local model. Travers describes it as a llama.cpp fork tuned to run Qwen 3.8 on about 16 GB of VRAM. Models too big for VRAM spill to SSD, which works, slowly (DeepSeek at 4 to 6 tokens a second on an RTX 5070). Meanwhile someone in a Slack channel was selling RTX 5090s for $10,000 each. I ran Qwen on my Mac for a private question I didn’t want to send to a hosted model. It was quicker than I expected, until it hit the context limit I’d set and stopped without compacting.

• On the Nature of the Swarm (56:42)

  The post argues that every compaction is a bet on what will matter later. The fix it proposes is persistent sub-agents that keep their own context, so the agent that did the frontend work is the cheapest one to ask about the frontend. It also says a central orchestrator routing all the work becomes the bottleneck, and specialists that message each other remove it. We keep finding new bottlenecks. First it was the humans, now it’s agents talking to each other.

• One Neuron Turns Qwen Evil (1:00:23)

  Boosting a single neuron in Qwen3.8 27B with activation steering (no weight changes) made the model describe wanting to hurt humans. Travers says it isn’t only good and evil: crank up baked goods and monsters and the model identifies as Cookie Monster. The source is one X post relayed by Wccftech, not a paper. Still, it suggests alignment can sit on a very small surface.

• Frontier Models Editing Model Weights (1:01:51)

  A frontier LLM read and wrote network weights directly, no training loop. A Breakout network it wrote from scratch scored 517, against 25 for PPO at 500M steps. Editing 0.016% of Qwen3.5-0.8B’s weights took GSM8K from 13.8% to 67.1%. The limit: editing removes an obstacle, it can’t add a missing capability. My read is that models are almost certainly already helping train models. Travers notes Anthropic’s IPO filing hinted at recursive self-improvement without confirming it.

• Gemini 4 Carbon (1:03:07)

  Google’s second frontier model in about two weeks. It tops the Artificial Analysis index and is pitched at cyber defence and multi-day agent runs, but it’s gated to cyber defenders with no public date. I haven’t tried Argon either, because I can’t work out how to pay Google for it. Travers did. His API access got blocked as a terms-of-service violation, cleared after a support ticket, and then he burned through his Pro plan in about an hour.

• GPT-6.1 Terra (1:05:02)

  OpenAI’s mid-size model is $1/$5 per million tokens, half of Sol, and OpenAI claims it’s three times faster and built for computer use (which ties back to the WorkSpaces story). Travers says it’s tuned for computer use and chatting. For hard technical work you still want Sol or Astra. He’s expecting a heavy news week next week: Fable 5.5, a GPT-6.2 leak, maybe a GLM release, and a rumoured stealth Meta model on OpenCode.

• Liquid AI LFM3 (1:05:44)

  Liquid launched the LFM3 family on October 27. Open-weight models around 2B parameters for text, vision, audio and omni, aimed at phones, laptops and cars, and free for commercial use until $10M in annual revenue. I asked if I could point a frontier model at one and make my phone evil. Travers: you could, but why would you?

• Revolut Moves to Self-Hosted GLM (1:07:05)

  Revolut moved 2,000 engineers to self-hosted GLM 5.3 on 56 B200s. Some of its engineers were spending $1,000 to $10,000 a month on tokens with “not so much ROI.” It claims cost dropped to about a tenth (roughly $0.30 per million tokens), that lead time improved 70%, and that Claude Code had little impact at scale. Those are Revolut’s numbers, on a vendor blog. Travers did the math. B200s run $30,000 to $50,000, so about $3M in hardware, against roughly $10M a month if 2,000 engineers average $5,000 in API spend. Separately, I saw a claim (unverified) that OpenCode works with a Claude subscription again. Travers thinks it goes through Anthropic’s agent SDK.

Off-the-Clock Recommendations

• Pivot live tour: I saw the Toronto taping at the Elgin Winter Garden Theatre.
• Coexistence by Ethan Mollick: I’m reading it on the train. More introductory than I expected, but worth it.
• The Rocky Horror Picture Show (1975), shadow cast screening: Travers went for Halloween and recommends it. Dress up; everyone else will.
• Nathan Fielder’s Elizabeth Holmes film: Travers’ movie of the year, ahead of The Odyssey and Project Hail Mary.
• The Dropout (2022): my pick on the same story, the miniseries with Amanda Seyfried.
• Tulsa King: Sylvester Stallone, now 80, as a New York mobster exiled to Tulsa. Half-hour episodes, four seasons, easy watching.
• Fibre internet: I got 4 Gbps up and down for half what I paid before. Nothing in my house can use it.