In Today’s Issue:

🎧 Spotify engineers stop coding thanks to internal "Honk" AI

🧪 Gemini 3 Deep Think hits 84.6% on reasoning benchmarks

💼 Microsoft AI CEO predicts white-collar automation in 18 months

🚨 Safety researchers exit Anthropic and OpenAI over risk concerns

✨ And more AI goodness…

Dear Readers,

Google's Gemini 3 Deep Think just scored 84.6% on a reasoning benchmark where humans average 60%, and that's not even the wildest headline today. Spotify's best engineers haven't written a single line of code since December, letting an AI system ship 50+ features while they watch from their phones. Microsoft's AI chief is putting a hard deadline on white-collar work as we know it: 12 to 18 months before most office tasks are automated.

Meanwhile, two safety researchers walked out of Anthropic and OpenAI in the same week, one to study poetry, the other over ads in ChatGPT, a signal that the tension between building fast and building responsibly is reaching a breaking point.

Today's issue covers all of it: the benchmark wars, the workforce countdown, the insider exits, and what it all means for anyone paying attention. Let's dive in.

All the best,

An AI safety leader at Anthropic has dramatically resigned, warning that the “world is in peril” from AI, bioweapons, and cascading global crises. Mrinank Sharma, who led research into AI safeguards and bioterrorism risks, is stepping away to study poetry in the UK, while another researcher quit OpenAI over concerns about introducing ads into ChatGPT.

The back-to-back exits spotlight rising tension inside top AI labs: balancing rapid commercialization (including ads) with long-term safety, ethics, and mental health impacts. With Anthropic recently settling a $1.5bn lawsuit and OpenAI defending its ad strategy, this moment feels pivotal for how AI companies prioritize values versus growth.

Anthropic's CEO: ‘We Don’t Know if the Models Are Conscious’

Google just dropped a bombshell upgrade to Gemini 3 Deep Think—and the numbers are hard to ignore. The specialized reasoning mode now scores 84.6% on ARC-AGI-2, a benchmark specifically designed to test abstract reasoning that humans average around 60% on. Let that sink in: an AI system is now outperforming the average human on tasks built to measure genuine intelligence.

But this isn't just about benchmark bragging rights. A mathematician at Rutgers used Deep Think to catch a logical flaw in a research paper that slipped past human peer review. Duke University's Wang Lab used it to design semiconductor crystal growth recipes hitting precision targets previous methods couldn't reach. Google also showed how Deep Think can turn a hand-drawn sketch into a 3D-printable file, bridging the gap from napkin idea to physical object.

The model now holds a 3455 Elo on Codeforces (Legendary Grandmaster territory), earned gold medals at the 2025 Math, Physics, and Chemistry Olympiads, and scored 48.4% on Humanity's Last Exam without tools.

Available now for Google AI Ultra subscribers, with API early access opening for researchers and enterprises.

Why it matters: Deep Think signals a fundamental shift from AI that predicts text to AI that actually reasons through complex problems. For scientists and engineers, this could become the most powerful verification and discovery tool they've ever had access to.

Sources:
🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

Coupon Protection partners with DTC brands like Quince, Blueland, Vessi and more to stop coupon extensions from auto-applying unwanted codes in your checkout.

Overpaid commissions to affiliates and influencers add up fast – Take back your margin.

After months of using KeepCart, Mando says “It has paid for itself multiple times over.”

Now it’s your turn to see how much more profit you can keep.

Try it 2 months free – No risk, Cancel anytime.

Microsoft's AI chief Mustafa Suleyman just put a timer on the office as we know it. In a Financial Times interview this week, the DeepMind co-founder predicted that AI will automate most white-collar tasks—lawyers, accountants, project managers, marketers—within 12 to 18 months. Not five years. Not a decade. By mid-2027 at the latest.

He's far from the only one sounding the alarm. UC Berkeley professor Stuart Russell, who co-authored the most widely used AI textbook in history, warns that political leaders are facing 80% unemployment, and nobody is ready. Russell says no job is safe, not even at the top: boards will pressure CEOs to hand decision-making to AI systems or get replaced, because competitors already have. Even surgeons, he argues, won't be spared once a robot can learn the job in seven seconds.

Anthropic CEO Dario Amodei predicts 50% of entry-level white-collar jobs could disappear in one to five years. A fresh Brookings study found 6.1 million clerical workers lack the savings, skills, or local opportunities to transition, 86% of them women. Entry-level hiring in AI-exposed jobs already dropped 13%. Companies like HP, IBM, Salesforce, and Klarna have cited AI in sweeping layoffs.

Russell raises the question nobody can answer yet: in a world where machines do all the work, how do humans find purpose?

The people building AI are openly warning us to prepare, from Microsoft's C-suite to academia's most respected voices. The gap between AI capability and workforce readiness is widening fast, and the workers least equipped to adapt will be displaced first.

What’s the secret to staying ahead of the curve in the world of AI? Information. Luckily, you can join 2,000,000+ early adopters reading The Rundown AI — the free newsletter that makes you smarter on AI with just a 5-minute read per day.

Sign up to start learning.