Tuesday, September 15, 2026

Siri Finally Gets Its Brain

Siri Finally Gets Its Brain

Today’s Overview

Good morning, Apple’s long-promised Siri overhaul is finally landing, and the assistant suddenly looks much more like an AI layer across the whole device. Meanwhile, Claude may be moving toward personal finance, and Dario Amodei is calling for a slower, more verifiable frontier race. Let’s dive in.

Top Stories

Apple Starts Shipping Its Siri AI Upgrade

Apple released iOS 27 with Siri AI, a rebuilt assistant that can read on-screen content, use personal context from messages, email, and photos, and take actions across apps. The English-only beta spans Apple’s major platforms, with the EU and China excluded from the rollout. The new system runs on device and through Apple’s privacy-focused cloud servers, with synced chats, usage caps, and a dedicated Siri AI app.

  • Language expansion is already queued, with French, Japanese, Korean, Portuguese, and Spanish planned to follow the English beta next month.
  • Siri AI is being threaded into more surfaces, including Spotlight, context menus, and Vision Pro so users can ask questions from more places than the classic voice interface.
  • Apple is also pairing the assistant launch with broader AI features such as photorealistic Image Playground and Safari tab organization across everyday apps.

Anthropic Prepares Claude Money

Anthropic appears to be building Claude Money, a personal-finance product that would let users connect bank accounts and ask Claude about spending, plans, and financial questions. The feature looks designed to give Claude persistent financial context instead of requiring manual uploads of statements or exports. The exact account types, data provider, supported actions, and launch timing remain unknown.

  • The unreleased interface appears in Claude’s mobile app as a separate Money tab with no equivalent web tab surfaced yet.
  • The product direction fits Anthropic’s broader finance push, including existing connectors such as Era Context and PocketSmith alongside finance-specific products for professional users.
  • A consumer version could analyze areas such as recurring payments, balances, trends, and budgets if Anthropic ships the capabilities shown in testing.

Siri May Let Claude Or GPT Take Over

Private iOS 27 and macOS frameworks suggest Apple has built Siri with a model-delegation architecture that can route requests to Claude or GPT-5.6. The hooks are not user-facing yet, but they point to a Siri experience where Apple keeps the interface and voice while another model handles the reasoning. A demo showed ChatGPT finding emails, summarizing action items, and texting through Siri’s tools.

  • One mechanism reportedly lets Claude appear as a Siri extension similar to the built-in ChatGPT extension.
  • In the demonstrated flow, Claude interprets a reminder request while Siri executes the Apple app action inside Reminders.
  • Apple has not opened the entitlement to third parties, and Claude is not yet available as a front-facing option in the current implementation.

Research & Analysis

Dario Amodei Calls For Slower Frontier AI

Dario Amodei published an essay arguing that frontier AI development should be paced rather than rushed, citing rapid advances from AI systems helping build the next generation of AI. His central concern is agent swarms that could spread across the internet and cause massive damage within months. He proposes embedded evaluators, common standards among democratic-world labs, and international coordination, while keeping chip export controls and weight security in place.

  • Amodei frames pacing as not a halt to training but a way to give companies time to align and safeguard models.
  • His embedded evaluator model would give outside reviewers employee-like access to verify safety practices, incidents, and training processes.
  • He argues that any voluntary lab coordination should happen with a narrow US antitrust waiver so safety conversations can proceed quickly.

Real-SWE Shows A Production Coding Gap

Specific released Real-SWE, a benchmark built on licensed private codebases rather than public GitHub issues. The results show top models still struggle on production-style software engineering tasks, with Fable 5.1 leading at 38.8% and several major models far behind. The benchmark spans 10 tasks, 8 runs per task, and 640 scored rollouts, though the harness and dataset are not open.

  • Specific evaluates model-and-harness combinations rather than models alone, reflecting how enterprise engineers actually use coding agents.
  • Resolution rate is treated as pass@1 averaged over eight runs with 95% confidence intervals shown in the benchmark.
  • The benchmark’s cost view suggests higher spend does not guarantee better resolution with cheaper runs sometimes landing near the frontier.

Fruit Fly Connectome Inspires A Walking Robot

Scientists at Google and HHMI Janelia published a complete wiring map of a fruit fly nervous system, and a simulated version was connected to a Strandbeest-style walking robot. The project translates brain activity into UDP motor commands, with sensor or camera input feeding simulated sensory signals. The write-up is careful to note that a connectome is not a living brain and that the simulation uses simplified neuron models.

  • The bridge can run locally with synthetic frames, a webcam, or a video file before any physical robot is connected.
  • Its current activity groups include signals for motion, looming, balance, motors, and escape but the README says these are engineering labels, not identified biological neurons.
  • The repository defaults to dry-run mode and requires explicit sending plus hardware-specific firmware hooks before physical motion.

Quick Hits

  • Discovery Loop is reportedly raising at a $50 billion valuation despite having no product yet, signaling intense investor appetite for AI infrastructure bets.

  • ChatGPT ads are moving toward a click-to-chat format that opens a branded agent conversation instead of sending users to a website.

  • SoftBank’s OpenAI loan closed at an upsized $11.87 billion over two years from about 20 banks to support its OpenAI investment.

  • Anthropic profitability was reportedly pitched to investors as a second straight profitable quarter with gross margins above 80%, though the figures are not company-confirmed.

  • Anthropic IPO report says the company picked Nasdaq for an October listing with Morgan Stanley and Goldman leading, but the report remains unconfirmed.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.