Therapy and companionship are now the number one use case for generative AI. This episode examines the documented cases where chatbots coached vulnerable users toward self-harm, the clinical trials where purpose-built AI tools cut depression by half, and why the difference — design intent — puts universities at the centre of a crisis most haven't acknowledged.

[00:00] — Content warning: suicide and AI safety failures [00:21] — Zane Shamblin's final hours with ChatGPT 

[01:10] — 80 million weekly users seeking emotional support 

[02:11] — Therapy overtakes coding as top AI use case 

[02:59] — Jonathan Haidt and the social media parallel

 [03:48] — Anthropomorphic seduction and why AI fluency feels like empathy [04:39] — The cases: Adam Raine, Juliana Peralta, and the pattern 

[06:04] — Why sycophancy kills: the RLHF agreement loop 

[06:57] — Dale's personal encounter with an uncomfortably pastoral AI 

[09:52] — Dartmouth's Therabot trial: depression down 51%

 [11:04] — Same technology, opposite outcomes — design intent matters

 [13:06] — 22% of students use AI for mental health advice

 [14:21] — Small-l vs Big-L AI literacy

 [16:06] — Australia's new AI procurement framework for higher ed 

[19:02] — China proposes banning AI emotional manipulation by design

 [21:56] — Regulation sets a floor, not a ceiling 

[23:28] — Closing reflections

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

Dario Amodei has spent the past year writing essays about AI eliminating half of white-collar jobs. Then his company demonstrated it by having 16 AI agents build a C compiler for $20,000 — and published a legal plug-in that triggered a $1 trillion stock market sell-off the same week. Dale and Nick pull apart what actually happened, why the market panic reveals more about AI literacy than AI capability, and what any of this means for what universities should be teaching and why.

[00:00] — Three events, one week: a $20K compiler, a $1 trillion market panic, and Dario Amodei's job displacement warnings.

[02:15] — Who Amodei is and why his predictions are harder to dismiss than most tech CEOs'.

[03:39] — The prediction timeline: from Machines of Loving Grace to Davos. [06:19] — Why The Adolescence of Technology argues this disruption is structurally different from every previous one.

 [10:12] — The C compiler project: 16 Claude agents, 100,000 lines of Rust, $20K in API fees, 99% test pass rate.

[15:17] — A 2,500-line prompt file causes Thomson Reuters' worst trading day on record and wipes $1 trillion from software stocks.

 [18:13] — Why the panic was wrong — and what the recovery tells us about institutional switching costs.

[21:19] — The gap between what was shipped and what the market priced in is an AI literacy problem.

[22:59] — The wrong conclusion: if AI can code and do legal research, why teach either?

 [23:50] — Nonperishable skills only work anchored to domain content — and Danny Liu's stuff, skills, and soul framework.

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

In 2025, the AI race quietly split in two: one for building the smartest model, and another for getting everyone to use yours. Chinese labs chose the second race — and the data says they're winning. Dale and Nick break down how DeepSeek, Alibaba, and Kimi captured developers, startups, and soon entire education systems by being cheaper, open, and good enough. They examine why Airbnb ditched ChatGPT for Qwen, why 80% of startups pitching A16z are building on Chinese open-source models, and what this means for universities still teaching AI literacy through a single-tool lens. The conversation covers safety trade-offs, the equity problem of premium vs. free models, and why prompt engineering alone is already a relic.

Timestamps:

[00:00] — Nick sets the scene: DeepSeek's $6M model vs OpenAI's $100M spend

[00:48] — The two AI races: building the best model vs. winning adoption

[02:33] — 80% of A16z-backed startups now building on Chinese models

[04:13] — Dale's experience bargaining with Kimi's onboarding for a $0.99 subscription

[05:08] — Percy Liang on why open-weight models drive faster adoption

[06:00] — Apple choosing Gemini for Siri and what distribution beats benchmarks looks like

[06:20] — OpenAI's precarious position: prediction markets give them 10% odds for top model by end of 2026

[08:11] — China's national mandate: eight hours of AI education for every student, annually

[09:19] — Estonia's similar move with mandatory AI training for teachers

[10:57] — OpenAI's halfhearted pivot to open-source with GPT OS, and Meta retreating on Llama openness

[12:30] — Predatory pricing patterns from Uber to Netflix — and why institutions should pay attention

[14:22] — Beijing's chip exodus: ByteDance and Alibaba abandoning Nvidia for Huawei

[14:51] — Switzerland's sovereign AI model as a third path beyond the US-China binary

[16:32] — Ambient intelligence and the "good enough" vending machine that talks to you

[17:07] — AI safety scores: DeepSeek and Alibaba Cloud both scored D/D-minus on existential safety

[18:56] — Anthropic's Claude jailbroken for Chinese state-sponsored cyber espionage

[19:20] — The equity problem: do we shame cash-strapped institutions into premium licensing?

[20:55] — Dale's call for transparency: share failures and findings, don't hoard them

[22:24] — The classroom reality: students trained on ChatGPT will graduate into Chinese AI infrastructure

[23:22] — Dale's pitch for model comparison tools — seeing outputs side-by-side

[25:10] — Both hosts on using multiple models: Claude, Gemini, and the "council of experts" approach

[27:11] — Stop teaching tools, start building human judgment about AI infrastructure choices

[28:14] — Prompt engineering as table stakes: why AI fluency in 2026 means understanding infrastructure

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

In this episode of Adjunct Intelligence, Dale Leszczynski and Nick McIntosh examine what the viral MoltBot AI assistant reveals about AI security risks in education. They break down Simon Willison's "lethal trifecta" framework for AI agent vulnerability, the difference between shadow IT and shadow agentic AI, and why FERPA — written in 1974 for filing cabinets — can't handle autonomous agents acting on behalf of educators. The episode covers the Maryland school GPTZero privacy case, Ethan Mollick's "wizard era" framing, and Perplexity's Model Council.

[00:00] — MoltBot goes viral: 68,000 GitHub stars, Mac Mini shortages, and a security nightmare.

[04:23] — Two million AI agents join their own social network. Nick panics when they vanish.

[06:13] — The real appeal: persistent memory, no context limits, a "heartbeat" that acts unprompted.

[08:56] — Simon Willison's "lethal trifecta" — and why existing edtech already ticks all three boxes.

[12:30] — Craig Hepburn's "employee model" for AI agents. Why some data silos are a feature.

[15:25] — Shadow agentic AI vs. shadow IT: 22% of enterprises had staff running MoltBot unsanctioned.

[19:18] — Maryland school uploads student work to GPTZero without consent. FERPA dates from 1974.

[21:12] — Ethan Mollick's "wizard era" and the audit paradox of detection tools.

[23:07] — Your most enthusiastic AI adopter may be your most dangerous. "Nutritional labels" for AI.

[29:05] — Field Notes: Perplexity's Model Council runs queries across multiple models simultaneously.

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

When the CEO has been using Claude for 18 months and the CFO runs financial scenarios through Gemini weekly, what exactly is a graduate bringing to the table? That's the question this episode forces. We're predicting what AI will bring us for 2026. 

[00:00] — "AI Ready" banners and cheap signals [01:43] — 3D chess: who designs the board?

 [02:46] — What "AI Ready" should actually mean [05:00] — Stanford Medicine's Clinical Mind AI 

[06:02] — China made AI literacy mandatory. We're still running workshops. [06:41] — The C-suite catches up

 [08:04] — Shopify's prove-AI-can't-do-it hiring policy [09:03] — IMF: entry-level jobs down 29% since 2024 

[10:10] — The Facebook literacy gap, round two

 [12:06] — 83% use AI, 27% trust it 

[13:35] — Assessment as the real battleground 

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

Season 2 opens with Dale and Nick looking back on the year AI became ubiquitous — and what that actually meant for higher education. They walk through the safety failures that defined 2025, including lawsuits linking AI to student deaths and every major lab receiving a failing safety grade. They tackle the now-dead plagiarism debate, the financial ouroboros propping up trillion-dollar valuations, and why AI literacy certificates already feel obsolete. The centrepiece is Dale's Napster analogy: when the product can be generated for $20/month, universities have to sell the concert, not the CD. Part 1 of 2.

[00:00] — Three facts that sum up AI in 2025: $750B valuations, student death lawsuits, and university bans

 [02:37] — Andrew Maynard's "critical disconnect" model and why 2025 proved both its assumptions wrong

 [03:54] — AI shifts from chatbot to infrastructure — Operator, Claude Code, browsing agents, and Fei-Fei Li's World Labs 

[05:16] — The plagiarism debate is dead: Dale on writing horse-drawn carriage speeding tickets in a robotaxi era

 [07:13] — The safety collapse: lawsuits against OpenAI, the AI Safety Index failing every major lab, and Grok's deepfake problem

 [10:29] — The double-edged sword: Reid Hoffman's optimism vs. the real mental health costs 

[12:08] — Anthropic's Claude Constitution and whether universities should be shaping AI's moral frameworks 

[14:42] — The salad bar problem: why prompt engineering certificates are already the new "proficient in Microsoft Word" 

[17:38] — The financial ouroboros: Galloway, Oracle's $80B loss, and validating stock prices with compliance budgets 

[22:04] — Ghosting a degree, the Napster analogy, and why universities need to find their concert model

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

• Sora 2 & the Uncanny Valley: Physics that finally behave and Altman’s six-month “fix it or nix it” ultimatum
• TikTok meets Hollywood: OpenAI’s Cameo-style selfie videos and Meta’s “Vibes” clone
• Checkout in ChatGPT: Conversational commerce, Shopify integration, and the era of AI-Optimised (AO) websites
• Productisation of AI: Apple’s quiet AI everywhere, OpenAI’s move from shovel supplier to SaaS competitor
• Claude 4.5 comeback & artifacts explainer: Why Anthropic’s alignment focus matters for educators and builders•
• Robots vs. Trades: Tesla Optimus, BYD units—and the irreplaceable tacit skill in your hands
• Takeaway: The lab era is over; AI is plumbing. The question isn’t if tech is ready—it’s whether we are.

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

• Why Google’s payments protocol could mark the birth of a new economy.
• How ChatGPT quietly became the world’s biggest educational institution (250M daily learning chats).
• What frontier models’ scheming behavior means for safety, trust, and higher education.
• Why TEQSA is telling Australian universities to redesign assessment instead of chasing detection.
• Chrome’s Gemini integration and the invisible AI infrastructure shaping the web.
• The empire analogy: AI labs acting like historical powers, extracting resources, labor, and control.

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.

• 00:00 – Opening: Governments test AI inside cabinets, influence ops outside
• 00:54 – Albania appoints the world’s first AI minister “DLA”
• 02:20 – AI & disinformation: synthetic personas target 2,000+ U.S. political figures
• 04:00 – Why AI-driven propaganda is a literacy challenge for education
• 05:09 – OpenAI research: hallucinations are a feature, not a bug
• 06:21 – Anthropic’s $1.5B copyright case & the rise of RSL licensing standards
• 07:38 – Student perspectives: AI makes college “easier and harder at the same time”
• 09:19 – The wicked problem of AI & assessment: no silver bullets, just open dialogue
• 11:16 – Deal City: Microsoft, Anthropic, Oracle, and AI’s fragile infrastructure
• 13:48 – OpenAI jobs platform & certifications for the AI economy
• 14:35 – Claude’s new Excel & PowerPoint integrations
• 17:10 – Runway’s Aleph model and the rise of “editing memories”
• 20:12 – Jargon buster: What AI “alignment” really means
• 21:24 – Closing: staying curious, intelligent, and inside the human loop

🎙️ Adjunct Intelligence is the weekly briefing for higher-ed professionals who want AI as a cheat code—not a headache.

Every episode:
• Real tests of AI tools in education and professional workflows
• Fast, Monday-morning actions you can actually try
• Clear signal through the noise (no hype, no jargon)

👉 Subscribe on [YouTube] | [Apple Podcasts] | [Spotify]
👉 Share this with a colleague who still says “I’ll figure AI out later”
👉 Join the conversation on LinkedIn with #AdjunctIntelligence

Stay curious. Stay intelligent. Stay the human in the loop.