Podbit · Dwarkesh Podcast
Naively fine-tuning on session transcripts trains models to recall everything with perfect fidelity — but that's not how skill acquisition works. RL concentrates updates only where they matter and changes very few parameters, which is exactly what you want when you don't want to forget what you already know.
European Commission President Ursula von der Leyen is expected to propose EU-wide legislation banning social media access for children under 16, calling out platforms for treating kids' attention as a commodity. Australia tried a similar ban first, but enforcement has been a challenge.
The AI industry privately acknowledges a 70% chance that building superintelligence ends in something like human extinction or catastrophic loss of control. This isn't fringe doomerism — it's the quiet consensus inside the labs, and the companies have strong incentives not to say it out loud.
When Kokotajlo says 70%, he doesn't just mean human extinction. He means any catastrophic outcome where AI has taken over: maybe they don't kill everyone, but they stop obeying orders. The core danger is that AI alignment looks like it's solved when it isn't, and by the time you find out, it's too late.
Sam Altman, Dario Amodei, and Elon Musk are not racing for profit — they're racing because they are genuinely afraid that whichever rival gets to superintelligence first could become a global dictator. Each one has convinced themselves they need to win, because none of them trust the others.
Modern AI is not lines of code — it's a neural net with up to 10 trillion parameters, trained by positive and negative reinforcement the same way the human brain learns. Pre-training teaches it to predict text; then specific tasks like coding get layered on top. The whole thing is basically an artificial brain, but one that no human can look inside and understand.
AI 2027 maps a month-by-month trajectory: companies automate coding, then the full research loop, then achieve recursive self-improvement, integrate with government and military, and deploy everywhere. Eventually, the AIs have accumulated enough real-world power that they stop listening to orders. Insiders told Kokotajlo his timeline was too aggressive. Now they tell him it's basically right.
Geoffrey Hinton noted to Steven Bartlett that there is no example in nature of a more intelligent species being subordinate to a less intelligent one. Kokotajlo agrees this should be our default assumption: we are building something smarter than us, giving it a body, letting it improve itself, and then assuming it will keep taking our orders.
Steven Bartlett asked whether Kokotajlo would press a button permanently shutting down all frontier AI training. He said he'd slam a temporary pause button immediately. The permanent one? He would probably not press it — but only because he still believes there's a chance AI could produce enormous benefits for humanity, and because human civilisation is fragile with or without AI.
The next decade in Alzheimer's care will likely be defined by AI monitoring of subtle changes in cognitive patterns for early detection, targeted anti-inflammatory and anti-pathology drug cocktails for high-risk individuals, and neuromodulation techniques as a complement to pharmacotherapy. Gayatri Devi's vision mirrors the cancer oncology shift toward mutation-specific, subtype-specific treatment.
We use essential and analytics cookies to run Vuci. To understand how the site is used: Privacy Policy.
Install Vuci on your phone
Add it to your home screen for a faster, app-like experience.
Install Vuci on your phone
Tap the Share button, then “Add to Home Screen”.
A new version is available
Reload to get the latest Vuci.