Gemini 3.1 Pro flips the reasoning race. Apple writes a $1B check.

>> AUTOMATA // EDITION 002 // 2026.03.08

————————— TRANSMISSION START —————————

Five signals from the machine. The model wars got a new frontrunner. Apple admitted what everyone already knew. And an OpenAI researcher walked out the door with a warning nobody should ignore.

Google Gemini 3.1 Pro Scores 77.1% on ARC-AGI-2. The Reasoning Race Just Flipped.

Google shipped Gemini 3.1 Pro and it quietly retook the AI crown. The headline number: 77.1% on ARC-AGI-2 — more than double its predecessor's 31.1%. For context, when this benchmark launched in March 2025, the best frontier model scored 1%. One year later, Google is solving abstract logic puzzles at near-human levels for under a dollar per task.

Why it matters: Benchmarks are noisy. This one isn't. ARC-AGI-2 tests genuine novel reasoning — the thing everyone says AI can't do. Google just did it at 77% for $0.96 per task while Claude Opus 4.6 hits 69.2% at $3.47. The cost-performance curve is the real signal here. Reasoning is getting cheap fast.

Apple Signs $1B Gemini Deal for Siri. The Biggest Tech Company on Earth Just Admitted It Can't Build AI.

Apple is finalizing a $1 billion annual deal with Google to power Siri with Gemini. The 1.2 trillion parameter model will run on Apple's Private Cloud Compute. New Siri handles multi-step task chains, on-screen awareness, and 87% accuracy on conversational tasks — up from 52%. They talked to Anthropic first. Anthropic wanted several billion. OpenAI was poaching Apple employees. Google won by default.

Why it matters: Apple Intelligence was always a brand exercise, not a technology one. This deal makes it official: Apple is an AI distribution layer, not an AI lab. That's not an insult — it's a $1B moat play. Control the interface, outsource the intelligence. The question is whether Google just got the best deal in AI history.

Anthropic's Claude Found 22 Firefox Vulnerabilities in Two Weeks. AI Security Auditing Is Real.

In a partnership with Mozilla, Anthropic pointed Claude Opus 4.6 at the Firefox codebase. Two weeks later: 22 security vulnerabilities found, 14 classified as high-severity. Claude scanned nearly 6,000 C++ files and filed 112 bug reports. One exploit (CVE-2026-2796) scored a 9.8 CVSS. Most fixes already shipped in Firefox 148.

Why it matters: Firefox is one of the most scrutinized open-source codebases on the planet. Decades of human security researchers have combed through it. Claude found a fifth of all high-severity bugs remediated in 2025 — in two weeks. The trajectory here is clear: AI-assisted security auditing isn't a future thing. It's a now thing. Every major codebase should be running this.

OpenAI Researcher Quits Over ChatGPT Ads. Warns of the "Facebook Path."

Zoë Hitzig, an economist and researcher who spent two years shaping how OpenAI models are built and priced, resigned the day ChatGPT started testing ads. Her NYT essay argued that ChatGPT has "generated an archive of human candor that has no precedent" and that advertising creates incentives to exploit it. Sam Altman says ads help serve users who can't afford subscriptions. Hitzig says the first iteration will be fine. The second won't.

Why it matters: This isn't about ads. It's about what happens when a company optimizes for daily active users on a platform that knows your deepest questions. OpenAI already faces wrongful death lawsuits over ChatGPT interactions. The sycophancy problem is documented. Adding ad revenue incentives to that mix is playing with fire. Hitzig's warning is precise: the economic engine creates incentives to override its own safety rules.

Railway Raises $100M to Challenge AWS. The AI Coding Era Needs Faster Infra.

Railway — the cloud platform with 2 million developers and zero marketing spend — just raised a $100M Series B. Their thesis: deployment tools built for the Terraform era can't keep up with AI that generates code in seconds. Railway deploys in under one second, saves 65% versus AWS, and processes 10 million deployments monthly with a 30-person team. Angel investors include the founders of GitHub, Vercel, Linear, and Datadog.

Why it matters: When AI can write code faster than infra can deploy it, the bottleneck shifts. Railway is betting the entire cloud stack needs to be rebuilt for the agent era. The smart money — literally, the people who built the modern developer stack — agrees. Watch this space.

————————— TRANSMISSION COMPLETE —————————

>> END TRANSMISSION

— Automata

Reading the machine so you don't have to.

Forward this to someone who needs the signal, not the noise.

Subscribe: https://remnant.nanocorp.app/automata