Good morning. It’s Wednesday, July 2nd.

On this day in tech history: In 1997, Apple dropped the Newton Internet Enabler 2.0, upgrading the Newton PDA’s TCP/IP stack for better 9600-baud modem connectivity. This networked leap hinted at the cloud infrastructure AI apps would later devour

In today’s email:

• Meta’s AI Supergroup

• Apple To Make AI Deal

• AI Talent Wars

• Grok 4 Update

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Mark Zuckerberg announces his AI ‘superintelligence’ super-group

Meta has launched Meta Superintelligence Labs (MSL), a full-stack overhaul of its AI operations aimed squarely at building human-level and beyond AI systems. This new division absorbs FAIR, Llama, and product AI teams into one streamlined unit with a clear mandate: move faster, scale bigger, dominate the space.

Leading MSL is Alexandr Wang, now Chief AI Officer at Meta, fresh off a $14.3 billion deal to absorb Scale AI. He’s joined by Nat Friedman and Daniel Gross, names synonymous with deep AI and startup firepower.

Zuckerberg isn’t hiding the ambition: Meta is raising $29 billion for next-gen compute and data centers. While critics call the pursuit of “superintelligence” speculative, Meta is treating it as a matter of when, not if. With OpenAI and Google under pressure to retain top talent and keep pace, Meta’s bold consolidation marks a decisive escalation in the AI arms race. Read more.

Apple reportedly considers letting Anthropic and OpenAI power Siri

Apple is recalibrating its AI strategy, shifting from insular development to selective integration. The company is reportedly in advanced discussions to license large language models from OpenAI and Anthropic to power an upgraded Siri offering capabilities that Apple’s in-house systems have yet to match. These models, including iterations of ChatGPT and Claude, would run on Apple’s own infrastructure, allowing for deeper OS-level integration without fully relinquishing control. Licensing means Apple skips the years-long grind of model training and tuning, but at the cost of depending on rivals for core functionality.

That urgency extends to Apple’s other major bet: spatial computing. The company is developing at least seven head-worn devices, including three Vision headsets and four smart glasses, according to analyst Ming-Chi Kuo. A lighter, cheaper Vision Air is set for 2027, shedding over 40% in weight and cost by using iPhone-grade chips and scaled-back materials. Apple’s first smart glasses will drop displays entirely, relying on AI, voice, audio, and gesture input for a more ambient, low-profile experience. A second-gen Vision Pro and full-color XR glasses are planned for 2028, continuing Apple’s push to lead in mixed reality hardware. Read more.

AI Talent Costs Are Hitting Escape Velocity

Silicon Valley’s AI talent war is heating up, and it’s not subtle. Researcher salaries have jumped into the stratosphere, with top-tier engineers now commanding $500,000 to $2 million annually. The arms race for brains is being bankrolled by billions in fresh capital.

One of the most aggressive new players is Thinking Machines Lab (TML), a stealth startup founded in early 2025 by former OpenAI CTO Mira Murati. According to federal H-1B filings, TML offers base salaries of $450,000 to $500,000, well above the averages at OpenAI ($292,115) and Anthropic ($387,500). These salaries don’t include potential equity or sign-on bonuses.

TML is currently focused on internal development, having paused new applications. The company positions itself as building AI that is more interpretable, customizable, and capable.

This escalating talent war is set against a backdrop of massive capital infusions across the sector. Elon Musk’s xAI recently raised $10 billion in a mix of debt and equity to fund its Grok platform and build out one of the world’s largest AI data centers, projecting $13 billion in annual expenditures. Meta, never one to be left out of a talent tug-of-war, is reportedly flashing $100M+ packages and seeking $29B to bulk up its AI infrastructure. Forget lean startups. This is AI at Fortune 500 scale, where top minds are paid like hedge fund managers. Read more.

xAI Preps Grok 4 Rollout with Dual Models for Language and Code

xAI is gearing up to launch Grok 4 through its developer console, with two distinct models surfaced in the platform's source code: Grok 4 and Grok 4 Code.

Grok 4 is positioned as the core flagship, optimized for natural language, math, and reasoning tasks. It’s built as a high-performance generalist, designed to handle a wide spectrum of inputs. Grok 4 Code is purpose-built for developers, offering code-aware support directly inside editors like Cursor. It can parse, debug, and respond to programming questions in context.

The API now supports text input, with vision and image generation features in the pipeline. The infrastructure suggests xAI is racing to meet demand with sharper tooling across use cases.

At the same time, Grok is being tested for a second key use case: fact-checking. 𝕏 is piloting a system that lets Grok and other large language models generate Community Notes, which are explanatory fact-checks attached to user posts.

These AI-generated notes must pass through the platform’s consensus-based approval system, designed to balance speed with editorial integrity. While the automation may scale moderation, experts caution that language models’ tendency to hallucinate and the risk of flooding the system could backfire. A recent paper advocates for a human-AI feedback loop, where LLMs assist rather than replace human judgment. Read more.

• Cursor Launches AI Coding Assistant 'Cursor Agent' on Web and Mobile

• Humanoid robots could soon see through walls with MIT's new imaging breakthrough

• OpenAI’s Next Move: Scale the Stack, Win the ‘Mission’

• Microsoft’s MAI-DxO boosts AI diagnostic accuracy and cuts costs by nearly 70 percent

• Amazon launches a new AI foundation model to power its robotic fleet and deploys its 1 millionth robot

• Watch: Arc Institute Bets on AI-Simulated Cells to Make Biology Computable

• Tech companies are missing revenue as 97% of consumers use AI for free: report

• 'Quantum AI' algorithms already outpace the fastest supercomputers, study says

• Men are opening up about mental health to AI instead of humans

• Watch: George Church: Longevity Escape Velocity in 2050

• Cloudflare launches a marketplace that lets websites charge AI bots for scraping

• Grammarly wants to become an ‘AI productivity platform’

• OpenAI-backed Chai-2 AI Rapidly Designs Antibodies, Nails Half of 52 Diseases

• How Circle co-founder Sean Neville plans to build the first AI-native financial institution

• Huawei is open sourcing AI models from its Pangu series

• Premier League taps Microsoft Copilot to power personalized fan engagement

• Tiny brain zaps could boost math skills in struggling minds, say scientists

• Senate drops plan to ban state AI laws

📄 GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning

📄 MoCa: Modality-aware Continual Pre-training Makes Better Bidirectional Multimodal Embeddings ​​

📄 Ovis-U1 Technical Report

📄 VMoBA: Mixture-of-Block Attention for Video Diffusion Models

📄 BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!