Human Ship, Apple Borrows, Musk List, Bezos Build

1. Continuing our series on transformer alternatives.

2. This week’s AI dives into Fable.

3. The opinion section discusses AI tokens as units of economics.

4. We might introduce a new fun section. Playing with new ideas.

With AI, some weeks it feels like patches are released in stages. This felt like a major upgrade for the entire industry. If you squint, you can see that these four events – the launch of the Frontier model, the reboot of the consumer assistant, the largest IPO in history, and a $12 billion bet on physics engineering – are all chapters of the same story. The idea is that the AI ​​will escape from the chat window.

Let’s start with “anthropology”. On Tuesday, the company released Claude Fable 5 and Claude Mythos 5, and the architecture of that release is just as interesting as the models themselves. Both share the same basic model. The difference is the policy, not the weight. Fable 5 comes with a conservative secure classifier that intercepts queries in high-risk areas such as cybersecurity, biology, and chemistry and falls back to Opus 4.8, while Mythos 5 runs indefinitely against a group of vetted cyber defenders under Project Glasswing. Think of this as the same kernel with different system call privileges. Benchmarks justify that warning. SWE-Bench Pro has 80.3%, which is over 10 points better than Opus 4.8 and more than 20 points better than GPT-5.5. We have entered an era where functionality and access are explicitly separated. The available models are sandbox views of existing models.

And finally, Apple showed up. At Tim Cook’s farewell WWDC, the company announced Siri AI. Siri AI is a conversational assistant with personal context, on-screen recognition, and a standalone app, reportedly powered by a custom Gemini model with 1.2 trillion parameters under the hood. There’s a certain irony in Apple, the original vertical integrator, outsourcing its brainpower. However, strategically this is the right decision. Apple’s moat was never a model. It’s a graph of distribution and personal context. One billion devices with intimate access to messages, photos, and calendars is a data set that no lab can reproduce. Apple isn’t competing on intelligence. It’s a competition for intimacy.

The boldest move this week came from Elon Musk. SpaceX went public for about $1.77 trillion, raising about $75 billion in the largest IPO in history. The prospectus reads more like an AI infrastructure paper than a rocket company. Musk, who integrated xAI into SpaceX in February, is touting an orbital data center. That means up to a million GPU-powered satellites will move training and inference off-planet, where energy is abundant and regulation is weak. xAI lost $6.4 billion last year on $3.2 billion in revenue, making the IPO effectively a public market undertaking for the most capital-intensive scaling proposition ever proposed. It turns out that Compute has escape velocity.

Finally, Jeff Bezos broke his silence on Prometheus, which raised $12 billion at a $41 billion valuation to build an “artificial general engineer” (an AI that designs and manufactures physical systems ranging from jet engines to pharmaceutical compounds). It’s not robotics, Bezos insists. Something close to CAD with a frontier brain.

Let’s dive in.

AI Lab: Google Research and Google DeepMind

summary: In this paper, we present a unified framework for building practical kernel-based two-sample tests derived from the f-divergence family. The authors demonstrate that these adaptive tests, especially hockey stick divergence, effectively capture a wide variety of local differences and are highly applicable to tasks such as differential privacy auditing and non-learning evaluation of machines.

AI Lab: Alibaba Group, Qwen Team

summary: The authors propose RACES, a framework that scales up reinforcement learning of language models by recursively assembling testable environments like building blocks when input and output types match. By leveraging composition operators such as SEQUENTIAL and PARALLEL, this approach generates structurally diverse training tasks and significantly improves the model’s inference generalization on unseen benchmarks.

AI Lab: microsoft research

summary: To address the high token costs associated with visual observation of computer-based agents, this paper introduces REVISION, a framework that trains multimodal models to filter redundant visual patches across consecutive screenshots. This method significantly reduces token accumulation while preserving important spatial structure, allowing agents to handle longer interaction histories and achieve higher success rates in complex tasks.

AI Lab: Fairness in the meta

summary: In this work, we investigate feedback distillation, a policy-based post-training technique in which a model learns to match its own token-level distribution, conditional on privileged feedback from a more powerful language model. Evaluated in a Lean 4 theorem proof, the method maintains better trajectory diversity and achieves better pass@k scaling than standard GRPO, and proves to be particularly powerful when used as an initialization for subsequent reinforcement learning.

AI Lab: stanford university

summary: DELM is a new multi-agent framework that eliminates the bottlenecks of centralized orchestration by relying on shared, validated contexts and asynchronous task queues. Agents independently request subtasks and contribute updates to a compact, verified global state, resulting in superior performance and cost efficiency for both software engineering testing and long-context inference workflows.

human Release of the long-awaited Fable 5 modela limited edition model based on Mythos. We also released a version of Mythos 5 for a select group of cybersecurity and infrastructure companies.

Moonshot AI “Kimi Work” releasedThis is a new agent that specializes in business automation.

1. SpaceX (SPCX) made its Nasdaq debut on June 12, 2026., It became the largest IPO in stock market history, valuing the company at nearly $1.75 trillion after raising about $75 billion at a price of $135 per share. Shares have soared to trade around $161 during the day, with the valuation backed by Starlink, which is now bundled with xAI following an all-stock merger earlier this year.

2. Bezos’ Prometheus raises $12 billion — Jeff Bezos and Vik Bajaj’s physics AI startup Prometheus has raised $12 billion at a $41 billion valuation to build “artificial general-purpose engineers” to automate the design and manufacturing of complex physical systems from jet engines to pharmaceuticals.

3. Mistral AI is in initial negotiations to raise approximately 3 billion euros (approximately $3.5 billion) at a valuation close to 20 billion euros (approximately $23 billion)— almost double last September’s Series C valuation of €11.7 billion. The new round brings the three-year-old company’s total funding to about 6.5 billion euros in debt and equity, and will help Europe’s leading AI lab ramp up its computing capabilities as it competes with larger rivals in the United States and China.

4. Taker raises $85 million — Barcelona-based Theker has raised $85 million in what it calls the largest robotics Series A in European history to build reconfigurable factory robots that can swap arms and hands for different tasks, rather than specializing in one thing.

5. Jediify raises $24 million — New York-based Jedify has raised $24 million in Series A led by Norwest, with Snowflake as a strategic investor, to build a “context graph” that gives enterprise AI agents the business knowledge they need to run in production.

6. Sandstone raises $30 million — Sandstone raised $30 million in Series A led by Lightspeed to bring AI-powered workflow automation (ingestion, routing, triage, and task execution) to in-house corporate legal teams rather than law firms.

7. OpenAI to acquire Ona — OpenAI has agreed to acquire cloud platform startup Ona and bring its secure, persistent execution environment into the Codex team, enabling AI agents to perform long-running, multi-step tasks for enterprises.

8. xAI Co-Founders Announce River AI — Igor Babuschkin, co-founder of xAI, announced River AI. River AI is a startup (partly made up of former xAI and Tesla employees) that builds personalized AI agents that learn from, and remain owned and controlled by, individual users rather than large corporations.

9. Tether backs Neura in $1.4 billion round — German company Neura Robotics has raised up to $1.4 billion in a Tether-led Series C, also backed by Nvidia, Amazon, Qualcomm, and Bosch, to scale production of humanoid/cognitive robots into the millions by 2030.

10. Apple announces Siri AI — Apple announced “Siri AI,” a reimagined Apple Intelligence-powered assistant with personal contextual understanding, on-screen awareness, world knowledge, dedicated apps, and expanded visual intelligence. It is currently available for developer testing and will be available as a user beta later this year.