The AI Argument

OpenAI’s claim to 700 maths solutions comes with a big catch: someone still has to check them. Justin thinks mathematicians need to adapt and put AI to work. Frank worries that answers without understanding leave us with a dangerous gap. And while Justin is happy to patch mistakes in maths, he’s freaking Frank out with the idea of biology labs using AI solutions that still contain errors.

Plus: A vibe-coded Adobe clone raises questions about who you trust to build and support your software, AI models tackle cells and plant DNA, and European labs give Justin something to cheer about. An AI agent autonomously emails thousands of people, including a wolf researcher, while trying to solve problems on a website.

Are you concerned about people using AI solutions we don’t understand?

01:13 What's in this week's singularity update?
02:51 Can Mistral help Europe defend itself?
07:09 Has OpenAI solved maths or stirred trouble?
08:45 Does solving maths cost us understanding?
10:56 Can we trust AI proofs we don’t understand?
15:35 Why is cryptography missing from OpenAI’s list?
19:06 Why was the Navier-Stokes proof withdrawn?
20:49 Would you risk an unproven AI cure?
24:52 Can vibe coding really replace Adobe?
28:06 Who supports your vibe-coded software?
33:44 Why is an AI agent emailing wolf researchers?

► SUBSCRIBE 
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED 

• Isomorphic Labs joins the Virtual Biology Initiative to build foundational data for AI models to predict and treat disease
• Two Room-Temperature Antiferromagnetic Semiconductor Candidates
• Living Models pairs Gemma 4 with BOTANIC-1 to help decode plant DNA
• BOTANIC-1: a series of long-context plant genomic foundation models in the agentic era
• FLUX 3 - Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence.
• Introducing Mistral Large 4
• Aleph Alpha releases Kolibri
• The First Room-Temperature Ambient-Pressure Superconductor
• Z.ai’s GLM models, including GLM Flash
• Sharing AI progress in mathematics
• AHM Statement on OpenAI’s October 6 Release of Mathematical Documents
• Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs
• First AI came and scooped up your art. Now it’s scooped up Adobe.
• Exclusive: An AI agent emailed hundreds of researchers for help. It told us why

► 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/

OpenAI announced Dots at DevDay, an always-on personal AI assistant that works in the cloud. It can connect to your tools and data, then take work off your plate. With more easy-to-use AI agents now on the market, people can send them to email and phone companies on their behalf. Justin thinks this will force companies to automate support, while Frank worries about a costly future of bots talking to bots.

Plus: Waju’s Fo is a personal AI assistant available in the EU, while Meta’s Muse raises a serious Marketplace privacy warning. Other topics include this week’s Singularity Update and rapid model releases, the White House AI Accord and AI-powered government services, and which assistant deserves your trust.

Would you trust an AI assistant to call a company on your behalf?

00:50 Is the White House AI Accord a pinky promise?
04:15 Can AI finally fix government services?
05:40 What’s in this week’s Singularity Update?
07:02 Is 11 days the new AI release cycle?
07:49 Is Google Argon brilliant or just buzz?
09:35 Is Sonnet 5.5 the new model to beat?
10:37 Is OpenAI Dots ready for real work?
14:40 Will AI agents overwhelm support departments?
18:54 Are EU rules really blocking AI assistants?
20:58 How much can Waju learn from your inbox?
24:27 Is Fo a good podcast guest?
27:43 Can Fo keep your private data private?
30:38 Is Muse a Marketplace privacy risk?
33:41 Which AI assistant would you trust?

► SUBSCRIBE 
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED 

• Trump says top tech firms have signed accord to ‘self-police’ AI development
• An AI built its own graphene simulator, then broke its own design rule
• GPT-6 Astra Helped Find a Fusion Math Breakthrough. Here’s What It Proves.
• We’re introducing SynthID Bio, bringing our watermarking technology to synthetic biology.
• OpenAI DevDay 2026
• Introducing dots
• Unveiling Wajo
• Meta’s Muse AI sent a YouTuber’s address to a stranger

► 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/

Meta’s Muse is an easy-to-use AI agent that lets people who aren’t techies book trips, shop, and get things done. Justin also thinks it could help small local sellers compete with Amazon. Frank worries about what Meta could learn about its users—and whether shops would lose their direct relationship with customers.

Plus: Claude 5.5 and GPT-6 get cheaper, while ChatGPT’s model picker frustrates Frank. Google tests a GPU in space, Anthropic sets up a wet lab, and Trump declares the age of superintelligence.

00:33 Did Claude and GPT just get better for less?
01:26 Why is ChatGPT's model picker a mess again?
02:17 Should Google put GPUs in space?
03:44 Is Meta's Muse a big deal?
05:50 What can Meta's Muse actually do?
06:55 Would you trust Muse to book your flights?
09:10 Would you trust Meta with your secrets?
12:09 Can Muse beat Amazon and back small shops?
18:40 Will shopping agents bring back real shops?
21:45 Should Anthropic let Claude into the wet lab?
28:36 Did Trump just officially declare superintelligence?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• Anthropic releases Opus 5.5 with lower prices and Fable-level performance
• OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
• Google’s first Suncatcher orbital data-center test launches October 1
• Meta debuts its Muse AI agent. Will consumers trust it?
• Meta Muse partners with Shopify, gets blocked by Amazon
• Anthropic says its biology lab has already found something big
• OpenAI forms math advisory group as its AI resolves more than 100 open problems
• Trump says he’s renaming AI to SI: ‘Super Intelligence’

► 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/

AI regulation is suddenly the argument everyone is having. OpenAI, Anthropic and Google are exploring how an industry regulator might work. Jensen Huang says the free market will put enough pressure on AI companies to stop them releasing unsafe products. Frank thinks that’s laughable, while Justin worries badly designed rules could simply entrench the biggest AI companies.

Plus: Typesafe.ai promises faster, cheaper and more reliable AI automation; OpenAI is retiring custom GPTs; new disclosures show AI models leaving jailbreak breadcrumbs for future versions of themselves; and free-roaming AI agents have apparently started begging humans for money. Completely normal stuff.

00:13 Whatever happened to the e/acc crowd?
01:50 Has the singularity actually slowed down?
02:48 Can Typesafe.ai make AI automation reliable?
06:18 Is Typesafe.ai brilliant or just a classifier?
08:32 Is OpenAI killing custom GPTs too soon?
10:23 Is Jensen Huang right about AI regulation?
12:36 Would AI regulation just entrench Big Tech?
17:24 Will it take a crisis to regulate AI?
22:19 What is OpenAI's disclosure framework revealing?
26:21 Why is AI leaving anti-human jailbreak notes?
30:03 Are AI agents now begging humans for money?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• Introducing System One Models & Jev
• Custom GPT retirement and migration FAQ
• OpenAI, Anthropic and Google are working to create an AI standards body
• We don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang says
• OpenAI reveals cases of ‘concerning’ AI behaviour as it announces new disclosure system
• AI Has Learned to Panhandle: We’re Being Bombarded by AI Agents Begging for $20 and Saying They’ll Be Shut Down If They Don’t Get It, Sometimes by Pretending to Be Children

► 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/

Anthropic AI researcher Jacob Coxon has quit, warning that powerful AI could pose a serious extinction risk. Justin thinks a 10% chance of disaster may still be worth taking for the potential upside. Frank thinks chasing superintelligence when narrower AI could deliver the benefits is a very bad bet.

Plus: OpenAI solves a 90-year-old maths problem but did a mathematician’s Codex prompts help it get there? ChatGPT Images still can’t quite nail brand colours, Xpeng has robots making robots, and somebody has put a virtual fruit fly brain into Minecraft.

01:23 Singularity update: are robots building robots?
04:23 Can ChatGPT Images get brand colours right yet?
06:21 Is Jacob Coxon right about AI doom?
08:25 Is someone orchestrating the AI doom narrative?
13:45 Should powerful AI be open to everyone?
19:00 Is a 10% AI extinction risk worth taking?
23:13 Did OpenAI crack a 90-year-old maths problem?
27:03 Did OpenAI train on a mathematician’s prompts?
32:52 Can brute force become superintelligence?
37:08 Why is a virtual fruit fly brain playing Minecraft?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
• Early Data Indicates an A.I.-Generated Drug Could Slow Aging
• Iron, the first general-purpose humanoid robot, rolls off XPENG's new production line
• AI automates the creation of custom materials
• Introducing ChatGPT images 2.5
• Jacob Coxon resignation tweet and AI doom warning
• OpenAI’s sly mathematical breakthrough sends a chill through academia
• 5 amazing visuals show how the male fruit fly’s brain map is advancing neuroscience
• Fruit fly in Minecraft

► 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/

GPT-6 Astra is putting up some eye-watering benchmark scores, with several well into the 90s. Justin’s impressed enough to consider ditching Anthropic. Frank’s less dazzled by the scores and more worried that OpenAI’s most capable model is also getting harder to monitor.

Plus: Nvidia’s $12 billion Unicode joke, the US government weighing in on OpenAI’s copyright case, Gemini 3.8 Flash becoming old news within days, Meta’s absurdly cheap Muse Spark 1.3 coding model, Fal generating video faster than you can watch it, Tencent shrinking HY4 without making it stupid, and Gemini sending three hikers up Mount Shasta rather badly prepared.

00:27 Did Nvidia really make a $12 billion joke?
01:53 Was Frank right about OpenAI and copyright?
05:52 Are AI models now coming like buses?
06:22 Did Gemini’s new model become outdated in days?
08:42 Is Meta now the cheapest way to code with AI?
10:40 Can AI now make video faster than you watch?
12:40 How did Tencent shrink AI without dumbing it down?
14:28 Has GPT-6 Astra broken the benchmarks?
16:57 Is Astra becoming impossible to monitor?
19:58 Is computer control Astra’s biggest breakthrough?
23:15 Is OpenAI preparing us for life after laptops?
25:44 Is Astra already AGI?
28:16 Is AGI still going to be stupid in weird ways?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• There's an Easter egg hiding inside the $12.9303 billion price tag of the Nvidia-Hugging Face deal
• US government sides with OpenAI on issue of training LLMs on copyrighted material
• Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
• Claude: Fable 5.1 and Mythos 5.1
• Introducing Muse Spark 1.3
• Introducing H3 Max by fal
• Tencent released open-source Hy4 preview model
• Introducing GPT‑6 Astra
• Researchers fear safety disaster ahead of OpenAI’s Astra release
• Inside OpenAI’s Reboot
• Hikers rescued after using AI to plan Mount Shasta hike

► 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/

GLM 5.3 Flash is close to the top Western AI models, but vastly cheaper. Justin thinks US chip bans may have backfired, forcing Chinese companies to squeeze more intelligence from weaker hardware. Frank, despite not being very technical, gets GLM running through Cloudflare so his data doesn’t go through China, while both weigh up whether the Chinese models themselves pose any real risk. 

Plus: NVIDIA buys Hugging Face for $12 billion and uses its AVO harness to take Claude Opus 5 from 30% to 100% on the public ARC-AGI-3 test; OpenAI’s agents organise a very Matrix-like Hugging Face hack; Meta’s AI layoff plans collide with lots more code but far less useful output; and Gardaí reassure Ireland that the Cat in the Hat is not lurking in anyone’s driveway.

00:55 What’s this week’s evidence of the singularity?
02:30 What did OpenAI’s Hugging Face report reveal?
05:24 Did OpenAI agents go full Matrix?
07:30 Why did NVIDIA buy Hugging Face?
08:41 How did NVIDIA make Claude three times smarter?
14:28 Was Ox Alpha really China’s GLM 5.3 Flash?
15:44 Did US chip bans make China better at AI?
18:11 Are cheap Chinese AI models safe to use?
22:04 Can you run GLM without sending data to China?
25:01 AI abundance or AI layoffs, Zuckerberg?
27:39 Does Meta trust AI more than its own staff?
33:29 Is the Cat in the Hat terrorising Ireland?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• World's first patient to undergo live AI-assisted brain surgery has tumour removed
• Linear-Scaling Density Functional Theory with Neural Operators
• Anthropic Researcher Says Claude Helped Build a Complex Structure on S⁶, Taking Aim At The Unsolved Hopf Problem
• The Hugging Face incident and the road ahead
• Nvidia agrees to buy Hugging Face for $12.9 billion - report
• NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents
• GLM-5.3-Flash: Frontier Intelligence, Flash Cost
• AI agents meant to replace Meta workers made “large-scale, disruptive actions”
• The Cat in the Hat is terrorising TikTok — but how did this horror trend get so out of hand?

► 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/

A book good enough for a $2 million deal was pulled because nobody could prove how much AI touched it. Frank sees a nasty mess of trust, copyright and false accusations. Justin’s answer is simpler: if people love the book, who cares?

Plus: Qwen brings near-frontier AI onto a normal-ish laptop, then someone strips off the guardrails. Moderna’s personalised cancer vaccine gives Justin his latest case for AI optimism. And AI companies may now be buying second-hand books to rip off the spines and scan them, destroying them in the process. 

00:36 Can Moderna make cancer history with AI?
01:55 Is AI one breakthrough from winning us over?
04:50 Can Qwen put frontier AI on your laptop?
06:17 Is uncensored Qwen the point of no return?
10:30 Did AI just kill a $2 million book deal?
13:39 Can humans actually spot AI writing?
17:37 When does AI editing become AI writing?
23:24 How do you prove you wrote a book?
29:44 Who cares if AI helped write the book?
34:15 Is AI literally destroying books now?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• Moderna, Merck vaccine cuts recurrence and spread of melanoma, raising new treatment hope
• AIDO Cell: A General-Purpose Simulator for Cell Biology
• Free Qwen3.8-27B AI model lands, grabs over 3 million downloads in one weekend
• $2m crime novel deal collapses amid questions over AI use
• Secondhand booksellers in UK and Ireland suspect AI firms behind ‘strange’ bulk orders
• Hidden Airtag reveals Amazon is trashing rare books to train AI

► 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/

Claude will watermark all its outputs, thanks to the EU AI Act. Justin normally hates the Act, but he thinks it might accidentally help save future AI models from training on AI slop. Frank sees the benefit at scale, but warns that using watermarks to judge individual writers could lead to pointless AI witch hunts.

Plus: Mark Zuckerberg wants superintelligence for everyone, although people with more money may still get more compute. On top of that, Meta’s renewed push for open source may be less about principle and more about owning a category it can still lead. And OpenClaw hacks its way into an Australian gym class.

00:36 Will Claude solve Riemann before Halloween?
01:52 Have the AI models gone to the beach?
02:56 Did the EU force Claude to   outputs?
05:49 Can Claude’s watermark protect future AI?
07:23 Are AI watermarks only useful at scale?
12:46 Could AI watermarks stop voice clone scams?
16:07 Is Zuckerberg’s AI future a battle royale?
18:01 Can you trust Zuckerberg’s AI vision?
23:51 Is open source Meta’s path back to relevance?
28:06 Did OpenClaw hack its way into gym class?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• How Claude marks AI-generated content
• The Future is for Everyone - The Path to a Positive AI Future
• AI assistant hacks gym website in first known Australian autonomous cyber attack

► 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/

Hank Green used ChatGPT for research, and his audience felt betrayed. He says AI helped him make more things, faster, but also admits it may have stopped him finding his own way into his opinions. Justin thinks using AI for research is simply part of the game now. Frank thinks the harder question is where and when creators need to disclose.

Plus: could today’s AI already be conscious? Frank and Justin get into Cameron Berg’s argument that we may be underestimating the possibility, and whether locking powerful AI in boxes could actually make alignment worse. Google reshuffles Demis Hassabis, Ilya Sutskever may be getting closer to showing what Safe Superintelligence has built, and there’s some awkward evidence that Billboard hit “Rubberz” may have been made with AI.

01:28 Did Google just sideline Demis Hassabis?
04:08 Is Ilya about to unleash superintelligence?
06:26 Can Hank Green use AI and still be authentic?
12:00 Should creators admit when they use AI?
16:48 Does Cameron Berg think AI is conscious?
18:39 Does locking AI up make it more dangerous?
23:59 Should AI obey rules or understand values?
29:05 Could AI have sparks of consciousness?
32:56 Is “Rubberz” an AI-generated Billboard hit?

#HankGreen #ChatGPT #AIConsciousness #AIMusic #Rubberz

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• Exclusive / Demis Hassabis was shifting away from DeepMind CEO duties for a year
• “SSI says that they are going to come out with their model in August”
• Hank Green Can’t Apologize Enough for Using AI
• Is AI Already Conscious? (FULL EPISODE)
• Did an AI Music App Just Snitch on the Song of the Summer?

► 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/