🔒 Internal run of show · for Olga + Chris · live link, not in navigation
Practical AI · Episode 48 · Show Control Panel
Your Contact Form Sorts Every Lead and Drafts the Reply. You Still Hit Send.
Friday July 10, 2026 · 11am CT · first show back after July 4 · two-week catch-up (Jun 26 – Jul 9) · title LOCKED (viewer-panel pick) · thumbnail: "Built in one evening. No code."
Two weeks off, and the story changed underneath us. It stopped being "look what the chatbot can do." It's now who controls the AI you rely on, who you can trust to grade it, and what it actually changes for a normal working person. A government switched a top model off and back on. Three new coding AIs dropped. The industry admitted its own scoreboards are shaky. The news is what's happening TO you. The deep dive is what you can do about it.
SAY IT RIGHT ON AIR — the two the prep audio got wrong:
ALBERTA: Do NOT say the agents worked in "isolated sandboxes on separate branches" — that was invented, not in the case study. What's real: a two-stage scan (rules engine flags, then Claude reviews each flag and cites the exact file and line), tests written first, and every patch reviewed and approved by Alberta's engineers before it shipped. It's Claude Code, not "cloud code."
FABLE 5 FREEZE: Do NOT explain it with "compute thresholds / dual-use like weaponry" — fabricated. Real: a US export directive citing national security over a reported narrow, non-universal jailbreak; Anthropic disputed it as overly broad while complying. Say "improved safety classifier," NOT "stronger safety filter" (that's a press paraphrase). Give both sides.
SOURCING: Alberta 6.5-year figure = Alberta's estimate, not an audit. SWE-Bench = "OpenAI found." Tilly Norwood / new models = company's own claim, wait for real-world use.
FUNDING (kept separate, 2 weeks): Week 31 AI = $6.67B (one Kling $2.8B mega-round carried it). Week 32 AI = $2.37B (SambaNova $1B, the biggest overall check that week was NOT AI). Tracker now 32 weeks / $392.7B.
0 · Cold Open
Six and a half years of government security work. Done in twenty hours.
Open on the Alberta story — it's the whole show in one image. Then straight into the news.
Picture an IT worker in Alberta's government staring at a security backlog estimated at six and a half years of work. A swarm of about 50 Claude Code agents clears the whole scan in 20 hours, before the weekend is over. "That's not a chatbot trick. That's what 'AI agents' actually means when you point them at a real, boring, enormous job. And that's where these two weeks start."
1 · News Block
9 beats + quick hits + Myth of the Week
Say the news first, then riff. ~23 min (two-week catch-up). Both sides on the government switch.
Alberta (cold open + lead) — 466M lines, ~50 agents, 20 hours, "6.5 years" by hand. Both sides: Anthropic's own tool + human sign-off required.
Three coding AIs in two weeks — Sonnet 5 (Jun 30), Grok 4.5 in Cursor (Jul 8), Meta Muse Spark + paid API (Jul 9). Callback to Ep47's Fugu / death of model loyalty — frame as payoff, not fresh.
OpenAI audited SWE-Bench Pro — ~30% of tasks broken, pulled its own recommendation. A correct answer fails over one space character.
The government switch — Fable 5 frozen worldwide, back Jul 1; + California 50% off, Sonnet 5 cheaper default (promo ends Aug 31), EU rules delayed to 2027–28. Both sides.
Meta Muse Image (Jul 7) — free, inside Instagram/WhatsApp; QR codes that scan, charts that add up. The privacy wrinkle: public-profile pull, opt-out but default on.
Tilly Norwood — AI actress cast in a real feature film. The dinner-table AI story.
Gene Wilder's voice (Netflix, Jun 30) — recreated with the family's blessing. The human heart of the two weeks.
Quick hits — ChatGPT voice · Gemini Mac app (acts on files, not your screen) · Gemini API agents · FireSat 5-meter fires.
Myth of the Weeknew beat — "SpaceX pocket iPhone killer" = Musk called it "utterly false" same day. The show's value prop in 60 seconds.
The Lead Operator — your contact form is losing you leads
Turn a passive contact form into an active operator that instantly sorts every submission into prospect / candidate / spam and routes it. Built live on revenuehire.com inside PageMotor.
One sentence: the second a form is submitted it gets handled. Nobody waits for a morning review. Prospect → captured in HubSpot + draft reply for Olga. Candidate → straight to Mileza to route. Spam → killed and suppressed. Event-driven, not a once-a-day job.
The brief line (say it right): "I wrote one brief for the job. Then Fable got up and ran." This is your PageMotor thesis happening to you — framework vibe coding, live. You described a business problem to an AI on a framework with guardrails, and a 13-year business got a new employee before midnight. Tonight you're the case study. (Deck slide: "I wrote one brief. Then it got up and ran.")
The plugin line (say it right): "I'd shipped a plugin before and didn't even know it counted. This is the first one that thinks — it stores data, decides, and runs my front door." (The /insights signup was a thin Kit wrapper. Tonight's is ~470 lines with a brain. Don't say "my first plugin.")
The story spine (if the deck dies)
Receipts: 12 submissions in 6 months, 3 real, a 20-year seller dropped twice, 12 days to answer a ready-to-buy client.
The one-sentence spec → three workers (website catches + sorts, Athena researches + drafts, Knox routes).
Some plain-English prompts, zero code → LIVE DEMO → what broke (4 lessons) → the prompt-injection question → guardrails (nothing auto-sends) → your website can work the front desk, you still make the call.
⚠ DEMO CHEAT — Olga only, do NOT screen-share this card
Screen-share ONLY revenuehire.com/admin/lead-operator-demo-2/ (note the -2). Every name on it is fictional.
In a second tab open revenuehire.com and submit the form with a company starting "DEMO" (e.g. "DEMO Braxton Co"). It hits the demo board's Live ledger within ~10 sec. Page auto-refreshes, no reload.
Lane phrasings that sort correctly on camera — Prospect: "we need to hire a closer." Candidate: "I'm looking for my next sales role." Spam: "best SEO backlink service for your google ranking."
Demo submissions never touch Knox, Mileza, HubSpot, or real data. They email only Olga, subject [DEMO].
NEVER share /admin/lead-operator/ (no -2) — that is the real board with real prospect and candidate names.
Morning check: dig +short TXT revenuehire.com — if Chris added the SPF record, fire one live test and the doorbell email works on camera. If not, the story is "the ledger watcher rings the agent." Don't over-explain.
Prep (from the viewer panel)
Lead into "four things broke" — all 5 viewers said that honesty is what earns them; the segment is a 5 without it, an 8 with it.
Have one real answer ready: is the routing hard-wired in the plugin code (strong) or just prompted (weaker)? A dev will ask.
If you can, flash one real prompt-injection attempt you threw at it + what it did. Turns the security slide from "knows the words" to "did the work."
3 · Funding — closes the show
Two weeks, kept separate
Funding is the closer, not #2. Two weeks because we were off for July 4.
Week 31 — "The Boom Was One Company." AI raised $6.67B, but one Kling $2.8B mega-round carried the whole week. Strip it out and the market was quiet. Week 32 — "The Biggest Check Wasn't AI." AI raised $2.37B (SambaNova $1B); the largest overall round that week went somewhere else. The money is still voting on who controls AI — and this fortnight it whispered.
Don't trail off. Tie it back: the news is what's happening to you; picking tools by the task, checking the source before you share, and staying the human who decides are what you do about it. Then ask for the subscribe, clearly. Judge the episode by conversion, not the clock.