AI safety claims and AI capability gains are moving together: a viral Anthropic resignation post said AI could kill humanity with a ~10% chance over the next decade, while the same week OpenAI-linked work and Anthropic-linked work were cited as pushing toward faster math and science breakthroughs. The Apple event was framed as the counterweight: a hardware stack around iPhone 18 Pro, Apple Watch, AirPods 5, and a foldable iPhone is being built to funnel on-device AI through cameras, sensors, and app intents, with Siri AI rolling out Monday and a Gemini-backed Apple stack still seen as privacy-first. The last segment tied the two themes together: data concentration, alignment risk, and rapidly improving models make the consumer tradeoff bigger, even as the odds of AI wiping out humanity were argued to be falling below 1%.
Jacob Coxin’s X post, initially from a young account with one tweet, hit 150 million+ views, ~750,000 likes, and about 250,000 followers overnight after being amplified in doomer circles and by legitimate AI researchers.
The post claimed humanity has about a 10% chance of dying over the next decade because Anthropic and OpenAI are being "incredibly reckless" about aligning frontier models.
The resignation post’s core line was: "I resigned from anthropic today. I spent the last 3 years doing pre-training research at both open air and anthropic," followed by the claim that both labs are racing to self-improving superintelligence and gambling with lives.
A pushback claimed alignment has been central for years, citing Ilya leaving OpenAI to start Safe Super Intelligence and arguing the goal is to limit downside risk as much as possible.
Apple’s keynote opened with Tim Cook handing off to John Turnus, who pitched a suite of about five products built around an "agentic" AI future and a "new version of a personal device."
iPhone 18 Pro was presented first: black, on a 2nm architecture described as the first consumer device on two nanometers, with Apple Intelligence pushed before hardware.
Siri AI is slated to roll out Monday, with the phones already in users’ pockets able to get the new assistant after an update; the assistant was described as optimized for those phones and comparable to ChatGPT, Gemini, or Claude.
Apple Watch was framed as the sleeper pick: a new health app taking heart readings every 5 seconds, producing readiness and sleep scores, and using a study of 1,000+ people across chest straps, Whoop, and Apple Watch to claim Apple is the most accurate and versatile wearable sensor.
The watch demo included an opt-in perpetual-listening mode that can recall misheard conversations, ambiently record the day, and summarize key moments; the privacy claim was that it is local, on-device, and not stored anywhere.
Apple said its installed base is 2.5 billion to 3 billion devices, and the AI strategy was to use that data edge plus private cloud processing encrypted "to the extent that it could be."
Apple was also said to have struck a deal with Google to license a proprietary Gemini version for Siri OS, paying about $1 billion a year, while the assistant itself was still described as not feeling intelligent enough and Siri as "the world’s worst personal assistant."
The iPhone Duo was the event’s most impressive new form factor: split screen with Siri AI on one side and an app on the other, letting the phone see what the user does, interact with 300,000+ app intents, and do more than a chatbot.
The foldable iPhone was described as Apple’s second most impressive device in a decade after Vision Pro: thinner than the iPhone Air by about 1/4 mm, with a matte inside screen, a proprietary sprayed finish, a 100-part hinge, two batteries, and a 48MP camera that can show live view while pointed at the user.
Foldable preorders were said to start at the end of October with delivery the week after, around the 23rd, and the starting price was given as $2,000.
AirPods 5 were priced at $129, with two versions around $130 and $150; the upgrade was ANC for an open-ear design, narrowing the gap enough that $250 AirPods Pro looked harder to justify if all you want is wireless earbuds.
Apple Pencil will work on the iPhone’s paper-like screen later this year, and the whole event was framed as a stack of devices feeding into one AI system that sees, hears, and summarizes the user’s day.
The AI/math segment said Navier-Stokes-like work was blown up for a $1 million prize this week after a computer showed stirring milk or coffee in a very particular way can make the rule fail and the swirl go infinite.
A Sept. 1 collaboration by two researchers who had done work at Anthropic and partly at OpenAI produced only a partial solution to a cousin Navier-Stokes problem, before rumors spread that Anthropic had solved Navier-Stokes internally.
OpenAI then used an internal model, not GPT-6 Astra, that had been trained only 12 days, ran nonstop for 88 hours, and produced the perfect solution; Sam was cited as confirming the cost was $10 million+.
The same segment said OpenAI/AI cracked a Millennium problem quickly even though the Millennium set started in 2000 and includes seven of the hardest math problems; solving math was presented as the fastest route into science, DNA, medicine, and AGI-like breakthroughs.
A credit fight followed: two researchers accused OpenAI of stealing user data by using ChatGPT and Claude to solve the problem, with a leaked internal log claiming OpenAI’s head of research offered them a contract and told them to sign up or lose credit.
The closer claim was that models are getting cheaper and smarter so similar results could arrive for a fraction of the cost in months, with math as the most deterministic training target and science next.
The science pivot extended to drug discovery: Anthropic internal models, Google DeepMind, and Isometric Labs were cited alongside a story that Douglas Yao used ChatGPT to design PAC 3310 in a garage lab to treat schizophrenia, avoiding three receptors that caused coben fry’s nausea, sickness, and joint pain.
The episode ended on a data-and-alignment warning: consumers are already giving more data to centralized corporations than ever, robots are already scaling up and China is "already there," and the speaker’s own preference is to trade data for a better quality of life while demanding transparency and imposed privacy.
The risk call was that AI killing humanity is now below 1% and likely falling over the next couple of years or decades, even as more devices, more model use, and more corporate infrastructure make the consumer-AI tradeoff harder to avoid.