Thursday, June 11, 2026

Apple Finally Reboots Siri

Apple Finally Reboots Siri

Today’s Overview

Good morning, Apple’s long-promised Siri overhaul is finally here, OpenAI’s IPO clock now appears tied to self-improving AI, and SpaceX is pitching orbital compute as an answer to the data center crunch. There’s also a sharp warning from Anthropic on how fast policy needs to move. Let’s dive in.

Top Stories

Apple Introduces Siri AI

Apple announced its long-delayed Apple Intelligence update for Siri, now branded Siri AI. The more conversational assistant is planned for OS releases this fall, with deeper AI integration across Apple platforms and Google-powered changes to Apple’s on-device Foundation Models.

  • Apple is positioning the revamp as Siri’s biggest AI shift since the assistant launched, moving it closer to chatbot-style interaction.
  • The assistant is expected to appear across iOS 27 and macOS 27 as part of Apple’s fall software cycle.
  • Coverage of the rollout points to a dedicated app, richer voice controls, and system-wide integration as key parts of the new experience.

OpenAI’s IPO Timing May Hinge on Self-Improving AI

Altman told employees OpenAI expects to go public within the next year, though progress toward RSI takeoff and compute needs could affect timing. OpenAI has filed a draft S-1 with the SEC, while the company is also weighing infrastructure needs that may require hundreds of billions in financing. Chief scientist Jakub Pachocki reportedly told staff that a model codenamed 5.6, described as a meaningful improvement on flagship GPT-5.5, is coming this month.

  • Altman’s reported logic suggests IPO timing is being treated as a strategic AI milestone rather than only a market-window decision.
  • The compute discussion points to infrastructure as a potential financing bottleneck even for one of the best-funded AI labs.
  • The Information page identifies the report as an exclusive briefing focused on OpenAI’s new model plans and expected public-market path.

SpaceX Previews Orbital AI Data Centers

SpaceX shared the first preview of AI1, a solar-powered satellite designed to run AI chips in orbit. Musk said each satellite could carry roughly the computing power of a top Nvidia server rack, with replaceable chip payloads, solar panels, radiators, and laser data links. Production is targeted before 2028 if timelines hold, and Google and Anthropic have signed on as orbital compute customers.

  • AI1 is described as a roughly 70-meter-wide satellite built around solar generation, cooling, communications, and compute hardware.
  • Reports say the design targets up to 150 kW peak compute per satellite, roughly framing each craft as a rack-scale orbital node.
  • SpaceX’s proposed Bastrop buildout is framed around an 11-million-square-foot factory intended to vertically integrate much of the AI1 supply chain.

Research & Analysis

Anthropic Calls for Faster AI Policy

Anthropic argues that AI is advancing faster than policymaking, creating rising risks in cybersecurity and job displacement. The piece recommends FAA-style regulation for frontier models, mandatory testing, stronger security standards, and broader reforms in macroeconomics, tax policy, biomedicine, and democratic governance.

  • The proposal calls for third-party testing in four risk areas: cybersecurity, biological weapons, loss of control, and automated R&D that could accelerate those risks.
  • Anthropic says government should be able to block or deter deployment when a qualified assessment finds unacceptable risks.
  • The essay also calls for faster labor-market measurement, including expanded statistics to track AI job displacement as it happens.

LLMs Make Patch Gaps More Dangerous

Anthropic’s red-team analysis says AI can accelerate the reverse-engineering of publicly disclosed vulnerabilities that are only partially patched across real-world systems. That makes the patch gap more dangerous because models can automate exploit development from public patches.

  • Across 18 recent Firefox security patches, Claude Mythos Preview built 8 working code-execution exploits autonomously.
  • On 21 Windows kernel patches, the same model produced 8 full exploit chains that escalated a low-privilege user to SYSTEM control.
  • In the Firefox tests, Mythos Preview’s first proof of concept arrived in about 12 minutes and 13 arrived within 40 minutes.

Probing LLMs Instead of Letting Them Speak

The article argues that answers can already exist in an LLM’s hidden state before token generation begins. It describes skipping generation by taking the final prompt-token hidden state, feeding it to a tiny MLP, and calibrating the result so a frozen model can act as a classifier over English-described labels.

  • The suggested extraction point is around 70% up the layers rather than necessarily at the model’s final layer.
  • The training recipe uses a few thousand criterion-content-label triples so the head learns a general satisfaction judgment instead of one fixed task.
  • The author says the resulting system can return calibrated probabilities in tens of milliseconds at roughly embedding-classifier cost.

Trending AI Tools

  • DiffusionGemma Google’s 26B Mixture of Experts text-diffusion model targets low-latency generation, with up to 4x faster GPU output and efficient local inference when quantized.

  • Gemini 3.5 Live Translate A real-time speech-to-speech translation model for 70+ languages, rolling out through Google products including Meet private preview and Google Translate on Android and iOS.

  • Cloudskill A tool for governing the AI skills a team depends on.

Quick Hits

  • EU orders Meta to open WhatsApp to rival AI chatbots for free after regulators said Meta abused its dominant position by blocking competitors from the WhatsApp Business API.

  • Codex helps automate a broccoli farm where a self-taught Hokkaido farmer uses ChatGPT and Codex for greenhouse controls, satellite crop tracking, disease diagnosis, and custom farm software.

  • OpenAI weighs a 10 GW Ohio campus through a reportedly NVIDIA-backed 20-year lease that would put OpenAI in control of the site’s computing equipment, with operations expected in 2028.

  • Ramp Applied AI Solutions embeds engineers in finance teams to build custom AI solutions and expand Ramp’s AI offering into hands-on implementation support.

  • OpenAI adds Lockdown Mode as a ChatGPT security setting that disables live browsing, agent mode, and deep research to reduce exposure to prompt-injection attacks.

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.