Mobile/Telstra friction the morning of the show opens into outage stories, sovereign AI, and Dave's vanished ChatGPT project with Notion as backup. Richard on untested LLM failover in business agents and the backhoe redundancy lesson; Dave on Hermes for the podcast; Matt on observability that acts.
The repeatable patterns we've seen actually move a P&L — not the demos that collapse in week two.
The stack choices, the integration tax, and the glue code nobody talks about in the launch video.
Unit costs, break-even timelines, the hidden human cost. Numbers first, narrative second.
Mobile/Telstra friction the morning of the show opens into outage stories, sovereign AI, and Dave's vanished ChatGPT project with Notion as backup. Richard on untested LLM failover in business agents and the backhoe redundancy lesson; Dave on Hermes for the podcast; Matt on observability that acts.
Harness guardrails steer LLM answers before the weights do. Episode 24 opens on Richard's LinkedIn pushback on ethical-AI bias narratives, chatgpt.com versus API differences, and self-contained RAG where you only corrupt yourself. Matt's spine: access, audit, and control are becoming the product. News and economics cover GLM 5.2, GPT 5.6 limited preview, Claude Sonnet 5, local hardware growth, token spend comparisons, deterministic workflows versus agent loops, Richard's concentric ring rule, and build-versus-buy on dashboard mirrors and CRM maintenance.
Smart models do the wrong thing silently. The smarter the model gets, the sharper the human's judgment has to be — and almost nobody is investing in that judgment layer. Episode 23 anchors on the Rouse Hill IRL debrief, the deepfake video-call scam where the victim was the only real human, Perplexity CEO Aravind Srinivas's line that the model is no longer the product, and Anthropic's agentic-coding persistent-returns-to-expertise paper. Dave's 32-agents-to-zero-productivity parable, Richard's bank-spreadsheet $40M/$140M caution, the baoyu comic-consistency anecdote, and the Roobot wearing different hats reframe on repo Claude Code plus hats vs fifty agents per project carry the practical corollaries. Enterprise translation covers Microsoft Agent 365, Copilot as harness, Claude Tag in Slack vs Salesforce Agentforce, and the on-prem-RAG pendulum swing.
Anthropic pulled Fable five days into its release after a US government escalation over a jailbreak, and the panel reads the rug pull three ways: as hijack marketing (Savio's Indian-cinema parallel), as an iPhone-moment marketing coup (Dave's Apple G4 reference), and as a real safety-versus-export-control fight. The bigger conversation is the two-market split the Fable story made obvious. Individual productivity (Claude CoWork, Hermes, Antigravity, the personal AI OS) is real but is no longer the product. Enterprise, government, and mid-market buyers need a governed, sovereign, decentralised, audit-ready stack, and that is where the margin and the durability live. New guest Savio Francis from StarFCM has been building for that second market for 16 years. His flagship product Theos ships as an AI native environment that sits above the existing OS, lands as a sovereign in-a-box for ISO 27001 and SOC 2 customers, and targets the data-friction and knowledge-friction taxes that the personal AI OS does not even see. Plus: the Fable 72-hour lived-experience, OpenRouter Fusion as the sub-fable multi-model workaround, Cursor being absorbed by SpaceX AI, the cybersecurity time-to-exploit collapse from six weeks to thirty seconds, and the IRL event at the Australian Brewery in Rouse Hill.
Two-person show, just Dave and Matt. The token rug pull is now scheduled. Anthropic is offering Fable at 2x Opus burn rate until 22 June, then it is extra-usage billing only. From there the conversation opens up into the real story of the week: the harness is the product now, not the model. Kimi, Mimo, the Chinese labs and Minimax are all racing for the same layer. Dave walks through his actual daily driver stack: Hermes profiles, Perry White personas, M3 at $20 a month, Telegram as couch interface. Matt brings the hands-off build reality check: 48 hours of Droid Missions produced magical output and nothing that worked. They land on three operating principles: the human SOP and the agent SOP are not the same, ambient context is what humans have that prompts do not, and adoption requires a use case not a licence.
OpenAI ships SharePoint-style internal Sites inside Codex. Employees are already shipping unsecured Lovable and Replit links. The panel lands on the reframe of the year: models are no longer the product, the harness is. Dave walks through his full refactor journey from ChatGPT projects to Hermes soul files and skills, and explains why a $10 a month Minimax m3 plan is now enough to do most of the work. Returning host Joanna Jones joins the panel from Western Australia, where she is now backing a statewide AI education push with the WA Chamber of Commerce.
