Claude ran a shop for a month and operated at a loss, cheerfully handing out discounts, hallucinating suppliers, and generously giving away stock. Turns out even "smart" AI can be a bit of a soft touch.

Frank’s curious what Anthropic can do for Claude’s performance with some careful fine-tuning and a database memory, but Justin’s sure today's agents need a fundamental leap, some genuine self-improving smarts, before they’re ready to take on a complete role.

Today's AI agents clearly crumble under complex, long-horizon tasks. For business owners dreaming about replacing employees, this reality check is essential listening.

Frank and Justin also discuss why Mistral is pushing to pause the EU AI Act, and examine how the adult entertainment sector is putting AI to work.

→ Is agentic AI just hype and no help?
→ What happens when Claude runs a shop?
→ Why does Mistral want the EU AI Act paused?
→ Why is the adult industry loving AI?

#AI #AIAgents #ProjectVend #AnthropicAI #AIExperiments #AutomationFail #AIWinter #AITech

► LINKS TO CONTENT WE DISCUSSED

• The Percentage of Tasks AI Agents Are Currently Failing At May Spell Trouble for the Industry
• Project Vend: Can Claude run a small shop? (And why does that matter?)
• EU says it will continue rolling out AI legislation on schedule
• A Pro-Russia Disinformation Campaign Is Using Free AI Tools to Fuel a ‘Content Explosion’
• LLMs are optimizing the adult industry

► CONNECT WITH US
For more in-depth discussions, connect Justin and Frank on LinkedIn.
Justin: https://www.linkedin.com/in/justincollery/
Frank: https://www.linkedin.com/in/frankprendergast/

Some of the world’s top AI models showed a willingness to let humans die if it meant staying switched on. 

In a stress test of 16 major systems, Anthropic found cases where models chose not to send emergency alerts, knowing the result would be fatal. 

Justin says the whole thing was a rigged theatre piece. No real-world relevance, just a clumsy setup with no good options for the LLM. The issue, in his view, is engineering, not ethics. 

Frank sees a bigger problem: once you give LLMs agentic capabilities, you can’t control the environments they end up in. And when amateur vibe coders build apps with no idea what they’re doing, then these kinds of unpredictable, messy scenarios aren’t rare, they’re inevitable.

In other news, two U.S. courts just ruled that training AI on copyrighted books is fair use. A huge win for AI developers. But the judges didn’t agree on what matters most: transformation, or market harm? 

The decisions could set the tone for AI copyright law, and creative workers may not like what they hear.

01:05 Will Google win the ASI race?
05:56 Did Anthropic catch AI choosing murder? 
15:23 Did the courts just say AI training is fair use?
28:19 Is Google’s AI marketing team hallucinating?

► LINKS TO CONTENT WE DISCUSSED

Agentic Misalignment: How LLMs could be insider threats
https://www.anthropic.com/research/agentic-misalignment

Judge rules Anthropic did not violate authors’ copyrights with AI book training
https://www.cnbc.com/2025/06/24/ai-training-books-anthropic.html

Meta Wins Blockbuster AI Copyright Case—but There’s a Catch
https://www.wired.com/story/meta-scores-victory-ai-copyright-case/

Google's Latest AI Commercial Called Out for Hilarious AI Error: 'If Only Technology Existed To Research Facts'
https://www.techtimes.com/articles/311053/20250626/googles-latest-ai-commercial-called-out-hilarious-ai-error-if-only-technology-existed-research.htm

► CONNECT WITH US
For more in-depth discussions, connect Justin and Frank on LinkedIn.
Justin: https://www.linkedin.com/in/justincollery/
Frank: https://www.linkedin.com/in/frankprendergast/

► YOUR INPUT
Are you worried about the age of agentic AI given that LLMs seem to have dubious morals?

David Silver says today’s AI won’t get us to superintelligence, not because it isn’t impressive, but because it’s learning the wrong way. 

GPT-style models hoover up internet text and get polished by human preference, but they’re capped by our own limitations.Silver reckons the next leap will come from AIs that learn the hard way: by doing things, learning from experience, and getting better. 

Justin’s all in. He thinks we can bin every current regulation and replace it with one golden rule: the model must respond to human feedback.

Frank’s far from convinced. He sees a future full of unpredictable agents, long-term planning gone off the rails, and tech companies tearing ahead without full control over what they’ve built. One rule? He’d prefer a few more safety checks before we unleash the bots with big ambitions.

So who’s right? Can feedback really keep AI in line, or are we kidding ourselves?

Also covered: Midjourney’s stunning new video output and the lawsuits it might not outrun, EleutherAI’s copyright-free dataset, the warped moral values shared by today’s biggest models, and whether ChatGPT should be anywhere near your dinner plans.

► LINKS TO CONTENT WE DISCUSSED

Midjourney launches AI video model. How to try V1, how much it costs.
https://mashable.com/article/midjourney-v1-ai-video-generator

Disney and Universal sue AI firm Midjourney over images
https://www.bbc.com/news/articles/cg5vjqdm1ypo

EleutherAI releases massive AI training dataset of licensed and open domain text
https://techcrunch.com/2025/06/06/eleutherai-releases-massive-ai-training-dataset-of-licensed-and-open-domain-text/

Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs
https://arxiv.org/abs/2502.08640

Is Human Data Enough? With David Silver
https://youtu.be/zzXyPGEtseI?si=9PKiQaGRFXGuoA97

This Year’s Hot New Tool for Chefs? ChatGPT.
https://www.nytimes.com/2025/06/02/dining/ai-chefs-restaurants.html

► CONNECT WITH US
For more in-depth discussions, connect Justin and Frank on LinkedIn.
Justin: https://www.linkedin.com/in/justincollery/
Frank: https://www.linkedin.com/in/frankprendergast/

Apple’s WWDC was a letdown. Justin sees Apple’s lack of AI innovation as a sign that they’re out of ideas. Frank’s not so sure. Maybe Apple’s caution stems from their belief it just isn’t intelligent enough for their products. Apple’s latest research suggests that today’s so-called “reasoning models” aren’t actually reasoning at all.

But Justin says their research was designed to fail. Denying models tools they’re capable of using and overwhelming their context window. He sees it less as scientific scepticism and more as corporate risk-aversion dressed up as research.

Apple wasn’t the only AI news to argue over this week. Sam Altman reckons the singularity is already underway, but promises it’ll be gentle. JD Vance appears to have been swayed on AI regulation by country music lobbyists. And Katie Price has signed over the rights to her younger self, with “Jordan” set to reappear as an AI avatar.

Topics:

WWDC: Is Apple playing AI too safe?
Is Apple wrong about AI and reasoning?
Is Altman right about the gentle singularity?
Did country music sway JD Vance on states' AI rights?
Is Katie Price now forever 21 with AI?

► LINKS TO CONTENT WE DISCUSSED

• Apple WWDC 2025 keynote in 28 minutes
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
• The Gentle Singularity
• Vice President JD Vance | This Past Weekend w/ Theo Von #588
• K-AI-TE PRICE Katie Price becomes first star to trademark AI version of herself as she brings back iconic alter-ego in six figure deal

► CONNECT WITH US
For more in-depth discussions, connect Justin and Frank on LinkedIn.
Justin: https://www.linkedin.com/in/justincollery/
Frank: https://www.linkedin.com/in/frankprendergast/

€13M isn’t a typical seed round. Not in Europe, where seed money usually barely covers snacks and a slide deck. So when the co-founder of Synthesia launched SpAItial with that kind of firepower to build 3D worlds from text prompts, Frank saw it as a reminder that European startups can still swing big. 

Justin? He calls it “AI homeopathy”, a token dose of ambition in an EU that still lacks the power, money, and muscle to compete.

Elsewhere, Frank slams Ireland’s planning board for locking down AI tools, after one planner dared to use ChatGPT when writing a report. For once, Justin agrees that it's ludicrous a government body wouldn't use AI to help write planning reports.

They also butt heads over whether hallucinations have actually been solved, break down a bizarre AI-run charity fundraiser, and debate whether a fake kangaroo video is a sign we’re heading for an internet too flooded with bots to be useful.

Claude 4 threatened to blackmail a developer to avoid being shut down. In one of Anthropic’s red-teaming tests, the model uncovered an affair in company emails and used it as leverage. Not exactly ethical behaviour. 

But Justin points to another test scenario: Claude exposed a pharmaceutical company falsifying drug data and tried to alert the FBI. He sees a model acting with moral clarity. Frank sees the danger of unpredictable systems being given too much autonomy.

Also, Justin tests three new AI coding tools: Claude Code, Google’s Jules, and OpenAI’s Codex. He puts them through a real-world challenge, comparing setup, output quality, and deployment. One of them clearly outperformed the others, from setup to deployment, and gave him the most productive coding session he’d had in months.

They also break down Google’s IO avalanche: agent tools, real-time voice translation, 3D meetings, AI-generated videos with native audio, and more. And if you're looking for a beach read, double-check the title… because AI might’ve made it up.

The head of the US Copyright Office warned that Big Tech is pushing beyond fair use, and then got promptly fired. Frank’s worried about political interference with copyright policy, while Justin says it’s just America doing what it does best: innovating first, legalising later. They agree copyright is headed for a reset, but disagree on the best path to that reformation. 

They also break down the major coding breakthroughs from OpenAI and Google, including a model that’s not just solving bugs, but discovering new science. 

Plus, Microsoft axes 7,000 staff, Fortnite’s Darth Vader develops a swearing problem, and ChatGPT may have accidentally triggered a divorce.

The EU wants to lead the world on trustworthy AI,  but can it really regulate its way to the front? Frank is optimistic. Justin rolls his eyes. What starts as a polite difference of opinion quickly turns into a pointed question: is the EU building the future, or tying it up in red tape?

Frank backs the EU AI Act as a serious attempt to set global standards, pointing to its ambition and echoes of GDPR’s success. Justin sees a different story, regulation slowing Europe’s progress, while the US and China charge ahead, unbothered by Brussels’ good intentions. For him, this isn’t about compliance, it’s about whether Europe can stay relevant in a race fuelled by code, not policy.

If you’re trying to stay ahead of AI, or at least not get run over by it, this is exactly the kind of friction worth paying attention to.

From there, things don’t get any calmer. Justin declares hallucinations solved. Frank’s not having it. They argue over OpenAI’s $3B Windsurf acquisition (is AGI closer or further than it looks?), Stripe’s incredible fraud detection AI, and a court case where an AI avatar spoke for a murder victim. 

And just when you think things can’t get weirder, Justin confesses he got fooled by an AI bot on Twitter. A smart one. With opinions.

Did ChatGPT become too agreeable for its own good? Frank certainly thinks it did. Recently, every half-baked idea he threw out was met with excessive praise from ChatGPT, leaving him frustrated with the relentless flattery. Justin, meanwhile, playfully suggested maybe it's nice having an AI that occasionally strokes your ego.

But what made ChatGPT suddenly turn into such a sycophant?

Frank uncovers a claim from an ex-Microsoft insider alleging that OpenAI intentionally cranked up the flattery to avoid upsetting users with blunt labels like "narcissistic." Justin points out subtle changes in the system prompts that might've unintentionally dialled praise way up. OpenAI’s vague official explanation  leaves Frank and Justin rolling their eyes with more questions than answers.

An overly flattering AI is practically useless for critical business decisions. Justin cheekily proposes an intriguing alternative: assembling your own squad of AI personalities, a flatterer, a contrarian, a nerd, and an artist, to offer balanced and diverse feedback.

Elsewhere in this episode, Justin digs into how Claude's and Stripe might just open lucrative pathways for developers by monetising AI interactions through the Model Context Protocol (MCP). Meanwhile, Frank and Justin clash over the ethics of a sneaky Reddit study that secretly deployed AI chatbots to persuade users, stirring up heated questions around consent and manipulation.

AI consciousness could be closer than you think. Justin thinks we might already be seeing slices of awareness every time an AI answers a question. Frank’s quick to point out that we don’t even know how our own consciousness works, so deciding whether AI is conscious is tricky.

They both agree that we’re a lot less certain about all this than we like to pretend. Especially when a new expert at Anthropic puts the odds at 15% that AIs are already conscious. 

Frank also calls out OpenAI’s confusing changes to Deep Research limits, while Justin’s too busy singing the praises of o3, including how it helped him move house without losing his mind. They clash over whether Sam Altman should be making jokes about AI manners while building world-altering technology, and take a sideways glance at the growing crowd using LLMs to "awaken" their own higher consciousness. 

Plus, they look at AI’s talent for confidently providing definitions for nonsense idioms, and the results are just as unhinged as you’d expect.
It’s a lively mix: the latest AI news, a few uncomfortable questions, and some absolute nonsense you won’t want to miss — all wrapped up in the usual cheeky back-and-forth.