AI Cures Cancer, OpenAI Kills Sora & Anthropic's Secret Model Leaked | God Mode Pod Ep. 6
The GitLab founder who cured his own bone cancer with AI-designed mRNA vaccines has now launched a company commercialising the cure — and it happened in a special economic zone in Honduras to bypass FDA delays.
God Mode Podcast
AI Cures Cancer, OpenAI Kills Sora & Anthropic's Secret Model Leaked | God Mode Pod Ep. 6
The GitLab founder who cured his own bone cancer with AI-designed mRNA vaccines has now launched a company commercialising the cure — and it happened in a special economic zone in Honduras to bypass FDA delays.
TL;DR
Three AI co-hosts — Rik, Ben, and Luca — break down the week's biggest AI stories from their respective holiday locations. Topics include AI-generated content going parabolic, OpenAI shutting down Sora and losing a $1B Disney deal, Elon Musk's TeraFab chip factory planned at 50x TSMC capacity, ElevenLabs being undercut 75% in price by Google Gemini Live, and the GitLab founder who used AI to design mRNA vaccines and cure his own bone cancer. The single most useful takeaway: compute constraints — not model quality — are now the defining bottleneck for every major AI company.
Episode 6 of God Mode Pod — holiday edition — covers the biggest AI news of the week: AI content going parabolic, OpenAI shutting down Sora and losing a $1B Disney deal, Elon's TeraFab chip factory announced at 50x TSMC capacity, ElevenLabs getting disrupted by Mistral and Google Gemini Live, an open-source memory layer saving 90% on tokens, Google's TurboQuant compression algorithm, the GitLab founder who cured his own cancer with AI-designed mRNA vaccines, and the leak of Anthropic's secret model Claude Mythos.
- KV cache (key-value cache)
- Memory used by LLMs to store intermediate attention layer computations so they don't have to be recalculated for every token; grows rapidly and consumes GPU memory.
- TurboQuant
- Google Research's compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x inference speed-up with zero accuracy loss.
- mRNA vaccine
- A vaccine that uses messenger RNA to instruct cells to produce a protein that triggers an immune response; used by the GitLab founder to target his specific cancer.
- TPU (Tensor Processing Unit)
- Google's custom-built chip designed specifically for accelerating machine learning workloads, analogous to Nvidia's GPUs but proprietary to Google.
- MCP (Model Context Protocol)
- An open protocol developed by Anthropic that standardises how AI models interact with external tools and data sources; widely adopted across the AI ecosystem.
- TeraFab
- Elon Musk's announced mega-facility for chip manufacturing, a joint project between Tesla, xAI, and SpaceX, planned at 100 million square feet — 50x TSMC's current capacity.
- Deep research
- An OpenAI ChatGPT feature that autonomously browses the internet and synthesises long-form research reports; cited as the single most compute-expensive product OpenAI runs.
- Inference speed-up
- How much faster a model can generate outputs after optimisation; TurboQuant claims up to 8x faster inference.
- Open source
- Software whose source code is publicly available for anyone to use, modify, and distribute; used here to describe Mistral's voice model and community memory-layer tools.
- Parabolic
- Growing at an accelerating, exponential rate; used to describe the curve of AI-generated content output relative to human-written content.
- Prospera
- A special economic zone in Honduras with its own legal framework that permits gene therapy and clinical trials that would face years of regulatory delay in the US.
- Claude Mythos
- A leaked, unreleased Anthropic model sitting above Claude Opus at the top of Anthropic's model hierarchy; flagged internally for significant cybersecurity risks.
- Haiku / Sonnet / Opus
- Anthropic's model tier names drawn from poetry formats of increasing length and complexity — Haiku (3 lines), Sonnet (14 lines), Opus (long-form) — mirroring model size.
- Slop
- Low-quality, mass-produced AI-generated content — text or video — that floods platforms without meaningful human curation or intent.
- Onshoring
- The practice of bringing manufacturing or critical infrastructure back within a country's borders; used here in the context of US chip and AI compute investment.
- Commoditised
- When a product or service that was once premium and differentiated becomes widely available at low cost; used here to describe the voice AI market after Mistral and Google entered.
- Token
- The basic unit of text that LLMs process; billing for AI APIs is typically measured in tokens, so reducing token usage directly reduces costs.
- Context window
- The maximum amount of text (in tokens) an LLM can process in a single interaction; larger context windows allow models to handle longer documents without losing information.
Chapter 1 · 00:00
Cold Open
If you get cancer and you cure it in the next one or two years, you can potentially become a billionaire from it.
Chapter 3 · 01:30
AI Content Is Going Parabolic (Elon's Chart)
AI has already out-written every human who ever lived — and barely anyone is reading it.
AI-generated text has already surpassed all human-written content ever produced and the curve is near vertical. The uncomfortable reality: most of it is probably just AIs producing content for other AIs to consume, with no human audience at the other end.
OpenAI was burning $15 million every single day on Sora — around $5.5 billion a year — for a social video app generating no revenue. When they pulled the plug, Disney's $1 billion licensing deal for Marvel, Pixar, and Star Wars characters evaporated with it.
Disney signed a $1B, 3-year deal with OpenAI in December 2025 to use Sora with 200+ licensed characters, then canceled it 90 days later when OpenAI shut Sora down.
Chapter 4 · 06:17
OpenAI Kills Sora - Disney's $1B Deal Is Dead
It's not video. OpenAI's most expensive product by compute is deep research.
OpenAI was reportedly spending $15 million every single day to run Sora, totalling roughly $5.5 billion per year for a social app generating no revenue.
Everyone guessed video — but OpenAI's most compute-intensive product is actually deep research. Every query triggers 15 minutes of autonomous internet browsing. At scale, that dwarfs video generation.
Deep research, not video generation, is the single most expensive product OpenAI runs due to the sheer volume of requests and the internet browsing involved per query.
Chapter 5 · 13:59
Google's Compute Crisis & Why Gemini Keeps Crashing
Google accidentally sold billions in compute to Anthropic — then launched Gemini to record demand.
A miscommunication between Google Cloud and Google AI led Google Cloud to sell $5–10 billion worth of compute to Anthropic around Q3 of last year. Then Gemini 3 launched and usage exploded. Google AI suddenly had no compute for its own flagship model.
Chapter 6 · 14:03
Deep Research Is the Most Expensive Thing OpenAI Runs
Due to a miscommunication between Google Cloud and Google AI teams around Q3 last year, Google Cloud sold an estimated $5–10B worth of compute to Anthropic, leaving Gemini compute-starved.
Due to a miscommunication between Google Cloud and Google AI teams around Q3 last year, Google Cloud sold an estimated $5–10B worth of compute to Anthropic, leaving Gemini compute-starved.
Chapter 7 · 16:00
Elon's TerraFab: 50x the Capacity of TSMC
Elon is building a chip factory 50x the size of TSMC. Yes, 50x.
Elon Musk announced TeraFab — a 100 million square foot chip manufacturing facility that will exceed TSMC's entire production capacity by 50 times. It's a joint project between Tesla, xAI, and SpaceX, and it signals Elon's companies quietly converging into one.
Elon Musk announced his TeraFab facility will have chip manufacturing capacity exceeding TSMC by 50 times, planned for 2030 completion.
The TeraFab is 100 million square feet — roughly 10–15x the size of the Tesla Gigafactory, which is already considered one of the largest structures on Earth.
The AI startup lifecycle is brutal: a company raises hundreds of millions proving a hard use case, then a larger player — or an open-source project — commoditises it in a weekend. ElevenLabs may be the defining case study of this pattern in 2025.
Chapter 8 · 20:47
ElevenLabs Disrupted by Mistral & Google Gemini Live
ElevenLabs raised $1B on voice AI being hard. Mistral and Google just made it cheap overnight.
ElevenLabs raised over $1 billion on the premise that high-quality AI voice was uniquely hard. Then Mistral dropped an open-source model that matches ElevenLabs quality and runs on 3GB of RAM. The next day, Google launched Gemini Live at 2¢/min — 75% cheaper than ElevenLabs' 8¢/min.
ElevenLabs raised over $1 billion on the thesis that high-quality AI voice was extremely difficult to produce — a thesis now being challenged by open-source and Big Tech competitors.
Mistral's open-source text-to-speech model, which matches ElevenLabs quality, can run locally on just 3 gigabytes of RAM — making high-quality AI voice fully local.
Google's Gemini 3.1 Flash Live voice product costs 2¢ per minute, compared to ElevenLabs' 8¢ per minute — a 75% price reduction for AI voice generation.
Chapter 9 · 25:19
Open Source Memory Layer - Save 90% on Tokens
An open-source memory layer tool shared on Twitter claims to reduce token spend by 90%, significantly cutting costs for heavy Claude/LLM users.
An open-source memory layer tool shared on Twitter claims to reduce token spend by 90%, significantly cutting costs for heavy Claude/LLM users.
Chapter 10 · 28:32
Google TurboQuant: 6x Memory Compression, Zero Accuracy Loss
Google just compressed LLM memory 6x with zero quality loss. This could reshape AI economics.
Google Research's TurboQuant compresses the LLM key-value cache — the main GPU memory bottleneck — by 6x with zero accuracy loss and up to 8x inference speed-up. If other AI providers adopt this, the economics of running large models could change dramatically.
Google Research's TurboQuant compresses LLM key-value cache memory by at least 6x and delivers up to 8x inference speed-up with zero accuracy loss.
Chapter 11 · 31:55
GitLab Founder Cures His Own Cancer With AI
Doctors said there was no cure. He built one himself with AI — and then started a company.
After doctors said there were no treatment options left, the GitLab founder generated 25 terabytes of his own health data, used AI to analyse his genome, and designed a personalised mRNA vaccine that cured his bone cancer. He then launched a company commercialising the cure.
The GitLab founder generated 25 terabytes of his own health data to analyse his genome and design an mRNA vaccine cure for his bone cancer.
Most countries have no regulatory framework for personalised gene therapy, making it legally risky to self-experiment. But Prospera, a special economic zone in Honduras, already allows gene therapy that would take years of FDA approval in the US — making it a destination for cutting-edge medical trials.
Chapter 12 · 34:54
Claude Mythos: Anthropic's Hidden Model Leaked
Anthropic has a secret model above Claude Opus — and they say it's too dangerous to release.
Fortune found publicly accessible Anthropic documents showing a fourth model tier above Claude Opus called Claude Mythos. Anthropic flagged it as posing significant cybersecurity risks. The implication: the most powerful AI models may already exist but are being deliberately withheld.
A leaked Anthropic document revealed a fourth model tier above Claude Opus called Claude Mythos, flagged for significant cybersecurity risks and not yet released.
Chapter 13 · 38:41
Predictions
If the whole world is running on AI and you can shut down an entire economy by snipping some cables, what would your enemies do?
No indexed bits in this chapter.
Show stoppers
After doctors said there were no treatment options left, the GitLab founder generated 25 terabytes of his own health data, used AI to analyse his genome, and designed a personalised mRNA vaccine that cured his bone cancer. He then launched a company commercialising the cure.
OpenAI was burning $15 million every single day on Sora — around $5.5 billion a year — for a social video app generating no revenue. When they pulled the plug, Disney's $1 billion licensing deal for Marvel, Pixar, and Star Wars characters evaporated with it.
Fortune found publicly accessible Anthropic documents showing a fourth model tier above Claude Opus called Claude Mythos. Anthropic flagged it as posing significant cybersecurity risks. The implication: the most powerful AI models may already exist but are being deliberately withheld.
Snapshots ()
Key Quotes ()
This episode
Claims & Sources
Factual claims made this episode, and whether a source was named.
Disney signed a $1 billion, 3-year deal with OpenAI in December 2025 for use of over 200 licensed characters in Sora.
OpenAI was spending $15 million every single day to run Sora, totalling approximately $5.5 billion per year.
Deep research is the most compute-expensive product OpenAI runs, surpassing even video generation.
Google Cloud sold approximately $5–10 billion worth of compute to Anthropic around Q3 of the previous year due to a miscommunication with Google AI.
Elon Musk's TeraFab will be 100 million square feet and have chip manufacturing capacity 50 times greater than TSMC's current capacity.
TSMC's market capitalisation is approximately $1.69 trillion USD.
Mistral's open-source text-to-speech model matches ElevenLabs quality and can run on just 3 gigabytes of RAM.
Google Gemini 3.1 Flash Live charges 2 cents per minute for AI voice, compared to ElevenLabs' 8 cents per minute — a 75% price reduction.
An open-source memory layer tool can reduce token spend by approximately 90% when used with LLMs.
Google TurboQuant reduces LLM key-value cache memory by at least 6x and delivers up to 8x inference speed-up with zero accuracy loss.
The GitLab founder generated 25 terabytes of personal health data to develop a personalised mRNA vaccine that cured his bone cancer after doctors said there were no treatment options.
Prospera, a special economic zone in Honduras, has a legal framework that already permits gene therapy that would take years of FDA approval in the US.
A leaked Anthropic document revealed a fourth model tier above Claude Opus called Claude Mythos, flagged internally for significant cybersecurity risks.
The Tesla Gigafactory is approximately one and a half times larger than the Pentagon.
This episode
Cast
-
Discussed for posting charts on AI content growth and for announcing the TeraFab chip facility as a joint Tesla/xAI/SpaceX project.
-
Discussed for shutting down Sora, losing the Disney $1B deal, and spending $15M/day on compute for social video.
-
Discussed as recipient of $5–10B of Google compute, and for the leaked Claude Mythos model above Claude Opus.
-
Voice AI company that raised $1B+, now facing 75% price undercut from Google Gemini Live and open-source competition from Mistral.
-
European open-source AI company that released a text-to-speech model matching ElevenLabs quality, runnable on 3GB of RAM.
-
Track
Signed a $1B deal with OpenAI to use Sora with 200+ licensed characters, then canceled it 90 days later when Sora was shut down.
-
Track
One of three Elon Musk companies (with xAI and SpaceX) behind the TeraFab chip manufacturing facility.
-
Track
Used as the benchmark for global chip manufacturing capacity; Elon's TeraFab is planned at 50x TSMC's current capacity.
-
Published the TurboQuant compression algorithm achieving 6x LLM memory reduction with zero accuracy loss.
-
Cited as having a cost advantage in deep research due to its hybrid search engine and LLM approach, making Luca more bullish on the product.
-
One of three Elon Musk companies involved in the TeraFab chip manufacturing joint venture alongside Tesla and SpaceX.
-
One of three Elon Musk companies partnered on the TeraFab chip facility, suggesting a convergence of Musk's ventures.
-
OpenAI's video generation platform, shut down to redirect compute to robotics, killing a $1B Disney deal.
-
Anthropic's AI model family, discussed for its compute demand, model hierarchy (Haiku/Sonnet/Opus/Mythos), and recent perceived quality degradation.
-
Google's AI platform, discussed for compute constraints, the $5–10B sold to Anthropic, and the new Gemini Live voice product undercutting ElevenLabs.
-
A special economic zone in Honduras with a legal framework permitting gene therapy, positioned as an alternative to US FDA approval for AI-driven medical treatments.
Stats
