The Inference Podcast

AI is excellent at everything except taste. So what is left for the people who make films?

Tom explains why people watch Non-Player Combat, his AI reality show where nobody is real, why AI is "excellent at everything except for taste", why he says the film jobs lost in recent years went with the streaming bubble rather than to AI, the messages AI filmmakers get from people who hate the work, and the year he thinks an all-AI film wins an Oscar.

Tom Paton directed ten independent feature films before founding AiMation Studios, an AI film and television studio. Its work includes Where the Robots Grow (2024), an AI-made feature released free on YouTube, and Non-Player Combat (2025), a reality series with an all-AI cast.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-08-is-hollywood-cooked

Chapters

• 0:00 A Reality Show Where Nobody Is Real
• 1:34 What Does a Director Do Now?
• 3:40 The Oscars and Human Authorship
• 6:44 What AI Still Can't Do
• 10:14 Will Reality TV Need Real People?
• 13:25 No Bloopers With AI?
• 17:13 AI Actors and Tilly Norwood
• 19:27 Protecting Your Voice and Likeness
• 24:08 Could AI Remake Game of Thrones?
• 27:17 Personalised Films as a Business
• 30:00 How Fast AI Films Get Made
• 31:50 Marketing and Finding Good Films
• 35:25 An AI-Only Platform?
• 36:42 How Films Really Get Funded
• 39:53 Ripping Up the Deal
• 41:12 Why AI Films Age Fast
• 41:37 Did AI Kill Film Jobs?
• 43:52 Reskilling for AI Workflows
• 45:03 The Future of Film Crews
• 47:31 Why Stranger Things Costs Millions
• 52:27 Effort as a Premium Product
• 57:20 Backlash and Angry DMs
• 59:43 Storytelling vs Software
• 1:03:56 AI or Traditional Filmmaking?
• 1:06:06 Rapid Fire
• 1:13:20 His Big Prediction

An engineer can be jailed for negligence even when the drawing came from AI. So who carries the risk when a machine designs the bridge?

Michael explains why an engineer can be jailed for negligence even when the drawing came from AI, who really carries the risk when a machine designs a bridge, why "inefficiencies are what drive construction" and where the margins sit, why firms that automate too much are one error away from a legal battle that could collapse them, what happens to graduates who never learn by getting it wrong, why a degree today "is worth a job interview", why he has never seen an Excel spreadsheet without errors, and why he would be happy to see research at the AI frontier stop and the bubble pop.

Dr Michael Rustell is a chartered engineer with a Doctor of Engineering in design automation. He led data science as a maritime engineer at AECOM and spent six years at Brunel University London before founding Inframatic, which builds AI for infrastructure work where failure is not an option: railways, bridges, ports and nuclear.

A note on HS2: the figure of £107 billion mentioned in the conversation is not the official estimate. In May 2026 the government put the cost of HS2 at £87.7bn to £102.7bn (2025 prices).

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-07-ai-designs-the-bridge

Chapters

• 0:00 Who Is To Blame
• 0:37 Risk Modelling Case Study
• 2:49 AI Assisted Design
• 3:51 LLM Failure Modes
• 6:07 Training Engineers Today
• 8:27 Judgement Through Experience
• 9:27 Drones Versus Simulation
• 11:22 Why Early AI Failed
• 13:13 Billing Time To Value
• 15:12 Big Firms Under Pressure
• 17:18 Buying the Autonomous Engineer
• 18:37 Data Centres Bubble
• 20:00 HS2 Cost Overruns
• 24:34 Who Profits From AI
• 26:31 Stamping Drawings Ethics
• 27:10 Liability and AI Oversight
• 28:57 AI Fakes Roman Cities
• 30:27 ChatGPT Maths Explained
• 31:30 Teaching With AI Tools
• 33:25 Engineering Jobs and Degrees
• 37:09 Civil Engineering in 50 Years
• 38:44 Megaprojects and The Line
• 40:47 Robots Building the Future
• 45:32 Founders Don't Die
• 46:28 AGI Utility Over Hype
• 49:09 Britain's Best Engineering
• 50:33 Rapid Fire AI Predictions
• 54:04 Built Environment AI Forecast

Since dynamic pricing arrived, Uber takes more from every trip while drivers earn less, and nobody outside the company can see how the numbers are set.

Reuben explains why Uber bans drivers from asking passengers what they paid, how the algorithm learned to pay less for hospital trips and "punishes drivers for being good human beings", why pricing algorithms can quietly collude in ways that would be illegal for two human bosses, the "human in the loophole" who exists to absorb the blame, the Kentucky bail algorithm that judges followed more often for Black defendants, the study where an AI boss cut pay by 40% and nobody complained, and why he says the only way to fight back is to join a union.

Dr Reuben Binns led the Oxford team that analysed 1.5 million Uber trips. His finding: since dynamic pricing arrived, Uber takes more from every trip while drivers earn less, and nobody outside the company can see how the numbers are set.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-06-your-boss-is-an-algorithm

Chapters

• 0:00 Uber Pay Secrecy
• 1:08 How The Algorithm Sets Prices
• 3:55 Human In The Loophole
• 6:33 Getting 1.5M Trip Data
• 8:01 Dynamic Pricing Exploits Drivers
• 11:05 Reinforcement Learning Incentives
• 12:51 Drivers Tacit Knowledge Undermined
• 15:12 Beyond Uber Dynamic Pricing Everywhere
• 17:51 RealPage And Algorithmic Rent
• 19:31 Uber Pushback And Exploitation
• 25:30 Do Rejections Affect Offers
• 28:26 Fixing Algorithms For Public Good
• 33:40 Meaningful Human Oversight
• 39:57 Testing Humans Versus Machines
• 41:27 UK GDPR Changes And Safeguards
• 43:52 AI Bail Decisions Bias
• 46:04 Workers Must Consent
• 49:29 Fighting Algorithmic Control
• 50:37 Piecework to Platform Risk
• 53:17 Silent Exploitation Study
• 55:01 Algorithmic Wage Collusion
• 58:34 Why AI Laws Fail
• 1:00:54 Tech Government Power Alliance
• 1:02:48 Frontier Slowdown Motives
• 1:05:13 LLM Hallucinations Explained
• 1:07:00 What AI Should Be Banned
• 1:08:54 Audits Gift or Leash
• 1:10:54 Automation Jobs and Income
• 1:12:56 AI Bubble and Moats
• 1:15:10 Rapid Fire Future Questions
• 1:18:26 Drivers Lawsuit and 2035 Prediction

Most people with robot opinions never built one. His went under four thousand British floors, in the dark, in the dust.

Tom gets into why the viral humanoid robot demos are "smoke and mirrors", why he says Figure AI's $39 billion valuation is "Theranos", how China is quietly winning the hardware war with free factories and zero-interest loans, why humanoids failed in Ukraine's trenches, Elon Musk's half-a-trillion-dollar humanoid gamble, the coming robot tax, why he lost control of his own company for £50k, and the AI bubble reckoning he sees coming for OpenAI and Anthropic.

Tom Lipinski has spent over a decade building real robots, and he believes the humanoid boom is mostly theatre, a bubble that bursts before it booms, because the machines everyone is investing in are the ones least likely to work.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-05-why-the-robots-arent-coming

Chapters

• 0:00 Jobs Humans Still Own
• 2:48 Humanoids Are Just Hardware
• 4:32 No Real Use Case Yet
• 5:31 Specialized Robots Win
• 7:45 Teleoperation Is Mandatory
• 11:01 Inside Teleop Demos
• 14:52 Marketing Claims vs Autonomy
• 17:07 Teleop Tool vs Trick
• 18:49 Valuations Bubble Talk
• 21:44 China Manufacturing Advantage
• 25:49 Subsidies Power the Ecosystem
• 28:31 Bust Before the Boom
• 30:50 Solving Dexterity and Hands
• 32:53 Why Robotics Lacks a GPT Moment
• 35:42 Synthetic Data Reality Check
• 37:15 Hype Cycles and Hard Truths
• 38:57 Narrow Wins Not AGI Robots
• 40:54 Humanoid Data and Tesla Costs
• 43:13 Viral Robot Hands Skepticism
• 44:48 Edge Compute and Innovation
• 45:52 Robots as a Service to Selling Work
• 52:04 Which Jobs Get Automated First
• 54:11 Why Construction Is So Hard
• 59:06 Redesigning Construction for Robots
• 1:02:07 Weaponized Robots and Red Lines
• 1:04:44 Reindustrialization Over Humanoids
• 1:07:09 Assisted Living Without Humanoids
• 1:10:03 Why People Abuse Robots
• 1:13:42 Robot Tax and Government Take
• 1:15:23 UK in 2040 Infrastructure Gap
• 1:17:11 Skills for the Next Generation
• 1:18:52 Founder Origin Story and Drive
• 1:21:48 Losing Control of Your Startup
• 1:27:50 Rapid Fire Predictions and Wrap

Sixty-five, third company, building the most advanced thing of his life. What the twenty-six-year-olds get wrong.

Mike gets into why corporate training and compliance courses are box-ticking theatre, the month ChatGPT rolled out memory and killed his startup Joggle overnight, how he built the Talvi website himself using Replit without writing a line of code, why SCORM and desktop e-learning tools refuse to die twenty-three years later, whether coding is still worth learning in the age of AI agents, why he says AGI is coming in 2 years, selling his first company two months before Steve Jobs killed Flash on stage, the white-collar bloodbath, the future of work, and what a 65-year-old serial entrepreneur chases when it is no longer money.

Mike Alcock has sold $100m of software to Apple, Nike and Shell, and he believes most of what you learned is useless, because AI now carries the knowledge for you.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-04-ai-will-kill-corporate-training

Chapters

• 0:00 Smartest In The Room
• 1:20 Experience Beats Age
• 3:00 Selling Enterprise Software
• 4:38 The One License Deal
• 6:23 Exits And Hard Lessons
• 8:54 Why Company Knowledge Is Hidden
• 10:07 Educating New Categories
• 12:03 Compliance Training Theater
• 14:24 Will SCORM Finally Die
• 16:34 Building Moats Against Giants
• 21:36 Token Efficiency And Speed
• 22:42 Company Brain And Privacy
• 26:06 Sovereign AI And Data Centers
• 31:27 White Collar Jobs Shift
• 35:38 Future Of L&D And LMS
• 38:04 Building Fast With AI
• 40:23 Next Big AI Opportunities
• 42:24 Software Over Hardware
• 43:17 Do We Need Knowledge
• 45:30 Socrates And AI
• 46:51 Trillion Dollar Britain
• 48:20 What Teens Should Learn
• 50:11 Future Proof Business Skills
• 51:55 AI Like The Early Internet
• 54:17 From Factory To Founder
• 1:01:04 Selling Into The US
• 1:02:59 Why The Name Talvi
• 1:05:54 Mentors Who Believed
• 1:10:24 Motivation At 65
• 1:13:22 Advice Beyond Hustle
• 1:14:40 Rapid Fire AI Takes
• 1:16:04 Talvi In Ten Years

AI is where the internet was in 1996, and the people who think they missed it have it backwards.

Stas Bichenko builds the AI that is replacing call centres, and he believes human customer support will be gone within five years.

He gets into voice AI and why it is harder than it looks, why he does not think AI will destroy humanity, the truth about sovereign AI and why only America really has it, why 95% of company AI projects quietly fail, the reverse casino reality of call centre work, and a prediction right at the end that nobody sees coming.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-03-voice-ai-will-end-call-centres

Chapters

• 0:00 Early Voice AI Era
• 1:38 Adoption Curve Explained
• 3:16 Copilots Are Overhyped
• 6:31 AI Is The 1996 Internet
• 8:09 Room For New Startups
• 9:33 Giants Still Leave Gaps
• 11:57 AI Won't Destroy Us
• 14:01 Will The AI Bubble Pop
• 15:41 Building In-House Voice AI
• 21:05 The No Lock-In Moat
• 22:51 Why Enterprise Pilots Fail
• 28:25 Interruptions Are Hard
• 33:00 Latency Limits And Thresholds
• 35:09 The Hello Timing Experiment
• 38:22 Beating First-Hello Latency
• 40:16 Prewarming The Single-Tenant Stack
• 41:34 The Manual Testing Moat
• 42:25 Why Agent Payments Wait
• 44:22 Sovereign AI Explained
• 46:22 Vendor Lock-In Risks
• 48:21 Local Models For Voice
• 52:40 The Linguistics Edge In AI
• 58:03 Automating Call Centre Work
• 1:06:42 Bootstrapping Over VC
• 1:09:48 Skills And The Future Of Work
• 1:13:14 Voice AI In Ten Years
• 1:15:26 A Bold AI Regulation Bet
• 1:16:46 Closing Thanks

Every technology before this automated something physical or boring. This one is coming for the judgement.

There is no leader in legal AI yet. Vipul Ladwa intends to change that.

Vipul is a lawyer who saw what AI would do to legal work back in 2018, left practice in 2023, and founded Casey AI, an AI platform built for the way law firms actually operate. We get into why ninety-five per cent of legal AI pilots fail, what fixed fees are doing to law firm margins, pilot fatigue, the broken training ladder for junior lawyers, and why the lawyers who win will be the ones flying the aircraft rather than racing it.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-02-ai-is-coming-for-lawyers

Chapters

• 0:00 Intro
• 0:30 Are we being replaced or augmented
• 2:00 How law firms actually make money
• 3:40 What billable hours really mean
• 5:00 How a junior lawyer is trained
• 10:00 Legal tech, Wave 1 versus Wave 2
• 12:00 Pilot fatigue, why law firms are exhausted
• 14:30 Fixed fees and the margin squeeze
• 18:00 Who trains the next generation of lawyers
• 21:00 AI giants versus vertical builders
• 26:30 Can you trust AI agents in law
• 31:30 Harvey versus Robin AI
• 35:00 Vipul's journey, leaving law for tech
• 43:00 The lawyer becomes the pilot
• 48:00 The twenty per cent revenue warning
• 53:00 Selling products, not hours
• 57:00 Vipul's biggest fear
• 1:00:00 What about our kids' careers
• 1:04:00 The future business of law
• 1:08:00 The real moat is data
• 1:10:00 Advice for law students

Why Europe funds AI differently, what the EU AI Act changes for builders, and where the vertical opportunity is.

Why do European AI startups raise a fraction of what US companies pull in? Tomas and Mian break down the real barriers holding back AI adoption across Europe, from the EU AI Act and GDPR to a deep-rooted culture of risk aversion among European investors.

They explore why European unicorns still choose New York for their IPOs, why the continent that builds the world's most advanced chip machines (ASML) still struggles to commercialise AI, and where the real vertical AI opportunities sit for founders building in this market.

Full episode page and transcript: https://inferencepodcast.uk/episodes/ep-01-ai-in-europe