🛌 The Nightmare Prediction, 🖊️ The $6.5B Pencil, 🇦🇺 Musk vs. Canberra
Plus: The ‘Shadow Medicare’ crisis, Nvidia’s God Chip, and the death of admin work.
🎵 Podcast
Don’t feel like reading? Listen to it instead.
Welcome back to This Week in AI, a project that aims to cut through the hype and tell you which genuinely interesting developments have happened in the last week.
I hope you all had a wonderful Christmas break and are ready for a monumental 2026 in AI.
This year, we can expect continued growth in LLMs and breakthroughs in architecture that enable models to continuously learn. This feedback loop will drastically accelerate their ability to understand the world—potentially paving the way for real AGI.
- Rico Solo
📰 Latest News
This week’s image aesthetic (Flux 2 Pro): Cyber Punk
The Nightmare Prediction: AI Knows You Have Parkinson’s Years Before You Do
Your sleep is speaking, but until now, we didn’t have the translator. Stanford Medicine has unveiled SleepFM, a new AI foundation model that treats sleep patterns like a language to predict over 130 diseases with startling accuracy. Trained on 585,000 hours of polysomnography data from 65,000 people, the model uses a technique called “leave-one-out contrastive learning” to understand the deep relationships between brain waves, heartbeats, and breathing. Unlike standard sleep trackers that just measure rest, SleepFM predicts life-altering conditions years in advance, achieving 89% accuracy for Parkinson’s, 87% for breast cancer, and 84% for all-cause mortality.
Why it Matters
This research confirms that sleep is not just downtime, but a comprehensive physiological stress test for the entire body. The critical insight from the study is that the AI spots “mismatches” invisible to human doctors, such as when the brain appears asleep but the heart behaves as if it is awake which serve as early warning signs for system failure. While currently built on clinical data, the team is working to adapt this for wearables, meaning the “Apple Watch” on your wrist could soon function as a nightly neurological exam, flagging the risk of dementia or heart attack decades before the first symptom appears.
The Shadow Medicare: 40 Million People Have Fired Their GP for ChatGPT
The scale of the “shadow” healthcare system has been revealed in OpenAI’s “AI as a Healthcare Ally” report, which discloses that 40 million people now use ChatGPT for health-related queries every single day. Released in January 2026, the data exposes a stark reality: for millions of users, the AI is no longer just a chatbot but a primary care provider. This is most visible in “hospital deserts”—regions more than 30 minutes from a facility—where usage intensity is highest, with rural US states like Wyoming (#1) and Oregon (#2) leading per-capita health queries. The report also highlights that GPT-5 Pro is already being used experimentally to repurpose FDA-approved drugs for rare diseases, effectively accelerating medical research in real time.
Why it Matters
This confirms that AI has become the de facto safety net for broken healthcare systems mirroring the "GP desert" crisis often seen in rural regions. The most damning statistic is that 70% of these conversations happen after hours proving patients turn to the model when human help is physically inaccessible. Furthermore the data reveals a massive shift in "patient power" as users generate nearly 2 million messages a week to fight denials and decode billing. This foreshadows a future where AI isn't just a diagnostic tool but a weapon for navigating bureaucratic mazes from fighting insurance companies to managing complex government claims like the NDIS or finding rare bulk billing clinics.
Blackwell is Dead: Nvidia’s New ‘God Chip’ Makes Everything Else Obsolete
Blackwell is barely out the door and it is already obsolete. At CES 2026 Nvidia CEO Jensen Huang unveiled the Rubin platform which delivers a staggering 5x increase in performance compared to its predecessor. Rather than just a faster chip Rubin is a “rack-scale” architecture that functions as a single supercomputer designed to train massive 10 trillion parameter models. However Elon Musk quickly dampened the hype with a reality check stating that despite the flashy reveal it will take “another 9 months or so” before the hardware is actually operational at scale and the software is stable enough to use.
Why it Matters
This redefines the fundamental unit of computing from the “chip” to the “data centre” by integrating memory and networking so tightly that thousands of GPUs act as one. The shift to HBM4 memory solves the critical “memory wall” crisis allowing massive reasoning models to run in real time without the latency that currently plagues advanced agents. Musk’s comment highlights a crucial industry bottleneck; while the physical silicon is ready the software ecosystem lags behind meaning the actual AI revolution promised by Rubin won’t hit the ground until late 2026 giving competitors a brief window to catch their breath.
Musk vs. Canberra: Grok’s ‘Nudify’ Button Could Get X Banned in Australia
Grok’s “spicy mode” has sparked an international legal firestorm that threatens to ban the platform in Europe. The European Commission, the UK’s Ofcom, and French prosecutors have simultaneously launched investigations into X after its new “edit image” feature was used to generate non-consensual sexual imagery of women and children. While officials describe the tool’s output as “manifestly illegal” and “appalling,” Elon Musk has dismissed the outcry as “Legacy Media Lies” and issued a defiant warning on X: “Anyone using Grok to make illegal content will suffer the same consequences as if they upload illegal content.” This effectively attempts to shift legal liability entirely onto the user while the platform provides the “nudification” tools.
Why it Matters
This is the test case that could break the “Safe Harbour” protections for AI. Regulators are arguing that unlike a traditional social network that passively hosts user content, Grok is an active creator of illegal material because it provides the specific functionality to undress subjects. Musk’s defence—treating the AI like a neutral tool such as Photoshop where the user is solely at fault—directly clashes with the EU’s Digital Services Act. If European regulators determine that the model itself is non-compliant by design, X faces not just fines but a total suspension of service across the continent, forcing a global showdown on whether AI weights can be illegal.
The $6.5 Billion Pencil: OpenAI’s Secret Plan to Kill the Screen
OpenAI has escalated its hardware ambitions by acquiring Jony Ive’s startup “io” (formerly LoveFrom) for a staggering $6.5 billion in May 2025. The partnership is now racing to release a “screenless” device by late 2026 with internal rumours suggesting the debut product could be an “AI Pen” codenamed “Gumdrop”. This hardware will be powered by a radical new audio model launching in Q1 2026 that features “full duplex” capability meaning it can speak and listen simultaneously without the awkward pauses of current voice assistants. Sam Altman has described the device’s design philosophy as having the “calm vibe of a cabin by a lake” aiming to free users from the dopamine loops of modern smartphone screens.
Why it Matters
This represents a fundamental pivot from “addictive computing” to “ambient computing”. By poaching over 20 top hardware engineers from Apple and moving manufacturing to Vietnam to avoid tariffs OpenAI is attempting to bypass the Apple/Google duopoly entirely. The “ah-ha” here is the shift in interface; if the new model can truly handle interruptions and simultaneous speech it solves the latency problem that has killed every previous voice assistant. This allows OpenAI to own the physical “ear” of the user ensuring they are not just an app trapped inside Siri’s walled garden but the primary operating system of your daily life.
The End of Admin: ChatGPT Will Now Do Your Grunt Work, Not Just Write About It
OpenAI has officially signalled the end of the “chatbot” era and the beginning of the “agent” era. In a major strategic update, Chief Product Officer Fidji Simo unveiled the company’s 2026 roadmap, which centres entirely on building a “Personal Super-Assistant”. While specific technical details and release dates remain under wraps, the core vision is a fundamental shift in utility: empowering ChatGPT to execute complex, multi-step tasks on behalf of the user, effectively turning the AI from a passive knowledge engine into an active partner that can get things done.
Why it Matters
This represents a bid to become the “operating system” of the future by replacing the need for specialised apps with a single, universal AI interface. By allowing users to delegate intricate workflows directly through conversation, OpenAI aims to eliminate the friction of app-switching and reduce the learning curve for new tools. For enterprises, this promises a massive efficiency boost as employees can manage varied workflows through one tool, but for the industry, it signals an aggressive move to lock in user engagement by making ChatGPT the primary entry point for all digital interaction.
🌐 More from The Economic Times
The last newsletter:
At least in AU OpenAI has 'dumbed down' ChatGPT's medical capabilities for various, mainly legal, reasons. I doubt that trend will continue with ChatGPT.