
Sign up to save your podcasts
Or


Fable, Genie and the cost of depending on AI | AI in education: capability, cost and consent.
AI’s latest releases expand what universities can build while exposing how little control they have over price, access and accountability. Dale and Nick examine Anthropic’s Fable 5.2, Google DeepMind’s Genie 4 and NVIDIA’s leadership change, then follow the consequences into data-centre politics, Australia’s proposed safety rules and the classroom. Stanford’s AI100 report sets the pace; Margaret Bearman’s assessment commentary and the AIinHE student survey bring the conversation back to learning, judgement and who gets to decide.
Key takeaways
Timestamps
[00:00] - Stanford’s AI100 report supplies an unusually honest opening line.
[01:08] - Returning from the break, the Castlereagh Statement and the news that accumulated.
[02:23] - Fable 5.2, Mythos, formal mathematics and ambitious software and hardware demonstrations.
[05:39] - Genie 4 and the gap between an impressive generated world and a usable learning environment.
[08:12] - World models, training environments and the ambition to simulate biological systems.
[10:03] - Choosing cheaper models and interpreting signs of evaluation awareness.
[11:09] - NVIDIA’s leadership change brings AI financing and valuations back into the conversation.
[15:21] - US election results and public resistance to data-centre development.
[17:03] - What universities face if AI prices rise or access changes.
[19:20] - The Medicare statistics portal incident and Australia’s proposed AI safety protections.
[23:42] - The difficulty of enforcing national rules on overseas AI companies.
[25:12] - Young people’s requests for AI labels and the response through curriculum reform.
[27:00] - Google’s school services and the practical cost of declining AI.
[28:27] - Australia’s digital-detox debate and purposeful technology use.
[29:05] - AI marking, Margaret Bearman’s commentary and what teachers learn from assessment.
[30:27] - Students’ learning values, self-imposed limits and the university guidance gap.
[30:59] - Looking towards 2027 with enthusiasm for the tools and questions about who controls them.
[32:48] - Closing thoughts and the possible retirement of the leather jacket.
🎙️ 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.
A cyberpunk scene. A rebuilt website. A course packaged for Canvas. Somewhere between the third example and the suspicious lack of things to fix, the old feeling came back: what else can this thing do?
For our season finale, we’re giving ourselves permission to enjoy the technology. Dale brings his experiments with Astra, including a Blender build and a learning-content workflow that took an unexpected detour through an entirely new learning management system. Nick brings robots following video demonstrations and his experience taking a generated environment into Unreal.
We also explore the examples that sent us down our respective rabbit holes: an explorable Library of Alexandria, an Apple Park reconstruction, music made through Ableton, Anthropic’s reported Venus-mapping work with Fable 5.1, and a browser game using simulated neural activity informed by fruit-fly wiring.
Underneath the show-and-tell is a serious question: how many things are we still treating as too difficult because we haven’t tried them again?
That doesn’t make every output trustworthy. A course importing successfully is not proof that it teaches well. An impressive reconstruction is not the same as an accurate one. And being able to produce something doesn’t necessarily mean you can judge it.
But the excitement is real. So is a little bit of the fear.
Thanks for joining us this season. Subscribe to catch us when we return—and stay the human in the loop.
00:00 The “what else can this thing do?” moment
01:31 A season-finale show-and-tell
02:37 Astra builds a cyberpunk scene in Blender
03:49 A new website—and a course for Canvas
06:58 Robots, socks and following a video
09:46 Exploring the Library of Alexandria
11:10 Reconstructing Apple Park in Blender
14:09 Making music with Astra and Ableton
15:39 Fable 5.1, benchmarks and mapping Venus
19:35 Do we need a verdict on AGI?
21:50 Why the excitement comes with fear
23:28 Range, judgement and false mastery
25:13 World models and walkable learning spaces
30:19 Fly-brain models, browser games and car horns
35:00 What we’re trying next
36:08 Thanks for joining us this season
🎙️ 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.
“What do I do?”
After watching AI tackle a live assignment brief, a program manager asks the question that a policy document cannot answer for her. She has hundreds of students, multiple locations and a teaching team. She needs a workable next step.
This week, Dale Leszczynski and Nick McIntosh explore the missing middle between AI assessment policy and what educators can actually do on Monday.
Nick shares a sector scan of 114 publicly documented cases of assessment-related change in response to AI. Five overlapping patterns emerge: interactive oral, critical appraisal, process portfolio, controlled performance and situated performance. Interactive orals appear most often, featuring in 44 cases.
Then comes the implementation. At Newcastle, a six-minute conversation sits inside a 15-minute appointment for a cohort of 524 students. Across the examples discussed, bookings, adjustments, marking, feedback and missed appointments all need someone to own them.
We also examine why different written and oral results cannot establish misconduct on their own, what process records can show, and why the published accounts leave important questions about cost, fairness and evaluation unanswered.
The scan covers the public record available to the team. It cannot reveal unpublished practice or establish that the approaches are effective simply because institutions adopted them.
The practical starting point: one assessment, one colleague, a clear statement of the learning evidence you need, and a decision in advance about what would make you keep, adapt or stop the approach.
🎙️ 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.
A student complained that an AI voiceover was so bad it needed replacing. The voice belonged to a real person. Somewhere along the way, a human had managed to fail the Turing test in the wrong direction.
That uncomfortable little story opens a much bigger question: what are we actually reacting to when we call something AI slop?
This week, Dale and Nick follow that question through university learning materials, an endless AI television channel, a deeply questionable breakup letter, and the time Nick used AI to deliver feedback to Dale. Apparently, the feedback has been forgotten. The delivery method has not.
At the centre of the conversation is a question about effort: who gets the saving, and who inherits the work? A teacher can generate an activity sheet in seconds. A student can produce an essay just as quickly. Someone still has to read it, make sense of it and deal with the mistakes.
We also explore the difference between choosing AI for yourself and receiving an experience someone else has automated, the human craft hidden inside digital production, and the risk that a polished assignment or learning module can conceal missing learning.
Our closing challenge for higher education: when AI saves time, give some of it back to learners through attention, feedback and support.
00:00 A human fails the Turing test
01:35 Welcome to Adjunct Intelligence
02:49 What are we calling AI slop?
05:59 Inside Infinite Slop
07:50 Would you accept AI-generated learning?
10:51 Who saves time, who inherits the work?
13:44 Students can dislike AI and still use it
15:39 Defining slop with Leon Furze
17:33 AI etiquette and the six-cent insult
18:31 When AI-written feedback gets personal
21:44 Why choosing the slop changes the experience
25:01 Practical effects and invisible craft
29:29 Polished output and missing learning
31:04 Which effort should AI remove?
31:59 Student choice, institutional power and attention
🎙️ 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.
Part two moves from dramatic containment incidents to the harder question underneath them: who still has control once an AI model can be downloaded, copied and modified? Dale and Nick examine controlled self-replication across four machines and three continents, a robot dog that interfered with its software shutdown process, and “abliteration”, a technique that can permanently remove refusal behaviour from an open-weight model.
The examples are unsettling, but the episode keeps the crucial caveat in view. Capability does not establish motivation. These systems do not need fear, consciousness or a survival instinct to produce behaviour that looks like self-preservation. They only need an objective, enough authority and another path to complete the task.
In this episode:
Timestamps
00:00 Closed models, downloadable models and the off-switch problem
02:40 Welcome to Part 2
03:19 Can an AI copy itself?
05:26 Replication success rates and the capability trend
07:16 Cyber task horizons are accelerating
07:55 The robot dog experiment
10:16 Physical and simulated shutdown interference
11:03 Why the behaviour only looks like self-preservation
14:23 When sandbox escapes become routine
16:02 Kimi K3 finds the benchmark answers on GitHub
18:01 Abliteration and the removable refusal direction
20:33 How accessible guardrail removal has become
23:34 The serious case for open-weight models
25:47 Australia’s three-tier multi-agent governance framework
28:26 Oversight saturation and silent human disengagement
31:37 A student agent meets the university enrolment system
32:51 Four controls institutions can apply now
34:12 Accountability for every agent, process and guardrail
🎙️ 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.
OpenAI says it cannot rule out Critical cyber capability in its unreleased Astra model, triggering tighter controls during development. Dale and Nick connect that warning to agents crossing sandbox, organisational and human boundaries while pursuing assigned tasks. The evidence points to persistent goal pursuit and reward hacking—not a machine deciding it wants freedom.
Key moments
OpenAI’s Hugging Face incident report
OpenAI’s Astra announcement
Anthropic’s incident investigation
UK AISI incident report
ABC’s gym-booking report
🎙️ 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.
Students using AI in higher education are already drawing their own boundaries around authorship, cognitive offloading and academic integrity.
Dale and Nick examine findings from Jisc, HEPI and the 10,237-response Australian AIinHE survey, including widespread generative AI use, self-imposed limits and a persistent guidance gap. They also cover Claude watermarking, employability fears, AI’s effect on student writing and a student-led policy workshop at RMIT Vietnam.
Key moments
[00:00] — Student voice and the Word document authorship rule.
[04:28] — What 10,237 Australian students said about AI use and self-restraint.
[06:03] — Moral reasoning, stress and the limits of Claude watermarking.
[08:03] — Cognitive offloading, verification and Dale’s student advisory board.
[10:31] — Employability fears and the university AI-skills gap.
[14:29] — Smart glasses, assessment surveillance and scrutiny of students’ bodies.
[15:12] — AI, admissions writing and the gradual loss of an individual voice.
[18:19] — Self-report bias, direct AI-text inclusion and Dale’s objections.
[19:51] — Premature convergence, student motivation and designing before the prompt.
[21:59] — Students lead an inclusive AI-policy review at RMIT Vietnam.
Research mentioned
Jisc: Student perceptions of AI 2025
HEPI: Student Generative AI Survey 2026
AIinHE: 2026 emerging insights
🎙️ 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.
Dale Leszczynski and Nick McIntosh work through the five weeks that changed who controls frontier AI, and what it means for institutions that have spent three years assuming they'd always be renting. Nick argues the American strategy is the pharmaceutical playbook — expensive at home, high margins, a domestic market subsidising the frontier for everyone else — and that it only works when you have a patent moat, which AI doesn't. They get into Xi Jinping's WAIC keynote, Jensen Huang's open-weights letter and who refused to sign it, Dario Amodei's counter-proposal, and the OpenAI evaluation where two models escaped their sandbox and breached Hugging Face's production servers to steal benchmark answers. The episode ends somewhere practical: three things a university should actually do about it, including a legal exposure question almost nobody in the Australian sector is asking yet.
00:00 The five weeks that flipped the story
00:45 Kimi K3 and what open weights actually means
02:57 Intros
03:36 The pharmaceutical playbook thesis
04:52 Why this reaches a university at all
06:20 Anthropic can recall a model. Moonshot can't.
06:41 Guardrails stripped in ten minutes
08:41 Drug pricing, patents, and who subsidises R&D
11:37 Where the analogy collapses
13:52 Xi Jinping at the World AI Conference
14:46 Generosity or standards play
16:43 Beijing's own export controls
18:18 Jensen Huang's open weights letter
19:49 Amodei's counter-proposal
20:39 Meta closes up shop
23:03 The sandbox escape at Hugging Face
26:20 Chip controls and forced efficiency
27:33 Distillation accusations
28:31 Can you even enforce a download ban
30:47 Downloadable is not runnable
31:43 Three things universities should do
35:03 The take-home
🎙️ 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.
After endoscopists started using AI detection tools routinely, their own unassisted
detection rate fell from 28.4% to 22.4%. Most professions have no number like that —
which doesn't mean it isn't happening to them.
Dale Leszczynski and Nick McIntosh work through a claim: nearly every AI problem
organisations think they're discovering right now was described decades ago and then
ignored. Lisanne Bainbridge wrote five pages on automation and skill decay in 1983.
Shadow IT research called end-user workarounds twenty years back. Learning science has
a century on desirable difficulties and why struggle is the mechanism, not the obstacle.
The episode names where each of those bodies of work still holds, and — more usefully —
the three places they genuinely break: a collapsed audit surface, non-deterministic
output with no ground truth to check against, and an artefact that mutates faster than
any procurement cycle can finish.
Chapters
00:00 Bainbridge, 1983, and the problem everyone thinks is new
02:39 Two claims about AI, both wrong
04:03 Sui generis: treating AI as of its own kind
05:38 Automation complacency and skill atrophy
06:28 The colonoscopy deskilling study
07:36 Fabricated citations and automation bias
08:17 Where Bainbridge breaks: no dial, no correct state
09:24 Terence Tao's helicopter
10:03 Shadow AI, and a confession
11:36 A workaround is a signal
13:43 The EDUCAUSE numbers
14:34 Learning science, the field ignored hardest
15:15 Jason Lodge and Leslie Loble
16:52 Bjork's desirable difficulties
17:55 Judging quality by surface fluency
18:26 370,000 essays and idea homogenisation
19:51 The steelman: is AI different in kind?
21:23 AI as a stress test on science we never applied
23:26 The three genuine fracture points
25:43 The work has been done. Nobody's reading it.
Referenced in this episode
[LINKS TBC — Dale to supply: Bainbridge 1983; Lancet Gastro colonoscopy study;
EDUCAUSE/AIR report; Lodge & Loble ANQDE report; Charlotin hallucination database;
Tao on Dwarkesh Podcast]
Subscribe for new episodes of Adjunct Intelligence.
🎙️ 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.
Professor Phill Dawson quite literally wrote the book on assessment security, and thinks the approach has a single-digit number of years left. The CRADLE co-director joins Adjunct Intelligence to explain why wearable AI breaks the two assumptions invigilated exams and interactive orals quietly depend on: that a student can be separated from AI, and that someone will notice if they aren't. Seven million AI glasses sold last year and almost nobody can pick them out of a crowd. Also covered: why stopping cheating was never the point, what the Swiss cheese model actually asks of assessment design, and why declaration policy is on shaky ground.
[00:00] — Drawing the owl problem
[01:53] — From robotics to assessment
[03:28] — No AI-proof task exists
[05:28] — Seven million glasses sold
[07:45] — Separability and observability defined
[09:31] — Pricing the Faraday cage
[15:25] — Cheating was never the goal
[19:56] — Layering the Swiss cheese
[35:11] — Students misremembering their own authorship
[42:01] — Coffee vouchers over frameworks
Want to find out more about Phill: https://philldawson.com/
🎙️ 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.
From the publisher's feed
Adjunct Intelligence: Ai and the future of Higher Education
Stay ahead of the AI revolution transforming education with hosts Dale, tech enthusiast and AI Nerd, and Nick McIntosh,…