In partnership with

Good morning. It’s Wednesday, January 14th.

On this day in tech history: In 1994, Vladimir Vapnik introduced Structural Risk Minimization, shifting learning theory from empirical performance to capacity control. SRM directly motivated support vector machines, but its deeper legacy lives on in modern generalization theory, regularization, and margin-based thinking. Long before “overfitting” became a meme, SRM quietly explained why some models should never scale unchecked.

In today’s email:

• OpenAI reportedly plans 5 devices, starting with ‘Sweetpea’

• Anthropic’s Claude ‘Cowork’ just dropped desktop agentic power

• Siri’s long-awaited glow-up is here—and Google’s doing the heavy lifting 

• 5 New AI Tools

• Latest AI Research Papers

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

Speak fuller prompts. Get better answers.

Stop losing nuance when you type prompts. Wispr Flow captures your spoken reasoning, removes filler, and formats it into a clear prompt that keeps examples, constraints, and tone intact. Drop that prompt into your AI tool and get fewer follow-up prompts and cleaner results. Works across your apps on Mac, Windows, and iPhone. Try Wispr Flow for AI to upgrade your inputs and save time.

Try Wispr Flow

OpenAI reportedly plans 5 devices, starting with ‘Sweetpea’

According to a report citing supply-chain sources, Foxconn has been instructed to prepare manufacturing capacity for up to five OpenAI-branded hardware devices by the fourth quarter of 2028. The full lineup has not been disclosed, but the report says a home-style device and a pen-like form factor are still under consideration.

The first device expected to launch is Sweetpea, a dedicated audio wearable positioned as an AI-native alternative to AirPods and developed with Jony Ive’s design team. Sweetpea is described as a metal, “eggstone”-shaped core with two removable, behind-the-ear modules. A launch as early as September is reportedly planned, with first-year shipment estimates of 40 to 50 million units.

The report claims Sweetpea will use a 2nm smartphone-class processor, with Samsung Exynos currently favored, alongside a custom chip designed to execute voice-based commands capable of triggering iPhone-level actions. Its bill of materials is said to resemble a smartphone more than traditional earbuds.

Separately, OpenAI has acquired Torch, a four-person startup focused on unifying medical data, in a deal reportedly worth $100 million in equity. Torch’s “medical memory for AI” technology will power OpenAI’s new ChatGPT Health service. Read more.

Anthropic’s Claude ‘Cowork’ just dropped desktop agentic power

Claude Cowork is now available in research preview for Max subscribers on macOS. The desktop AI agent is designed for non-technical users, allowing Claude to read, edit, organize, and create files within a user-selected folder. It can handle tasks like reports, spreadsheets, and media management without any coding.

Built on the Claude Agent SDK, Cowork autonomously plans workflows, executes multi-step tasks, and integrates with browser automation and existing Claude connectors. Anthropic cautions that risks like file deletion and prompt injection remain, though folder-level permissions help limit exposure.

Anthropic is also shaking up its C-suite to expand its internal incubator. Former CPO Mike Krieger will co-lead Anthropic Labs, the company’s experimental AI incubator, while Ami Vora takes over as CPO to scale Claude products. Labs plans to double its team within six months.

On the ecosystem front, Anthropic has pledged $1.5 million over two years to the Python Software Foundation to enhance PyPI security, maintain CPython, build malware datasets, and fund community programs, signaling a dual focus on AI advancement and open-source stewardship. Read more.

Siri’s long-awaited glow-up is here—and Google’s doing the heavy lifting 

Apple’s Siri finally gets the upgrade it’s been promising since 2024. Apple is tapping Google’s Gemini to power the next-gen version, using heavyweight models for complex reasoning and lighter ones for everyday tasks. It runs on-device or through Apple’s Private Cloud Compute for maximum privacy. Expect deeper personal context, on-screen awareness, and real cross-app actions. ChatGPT stays in the mix for tougher queries. The rollout is expected with iOS 26.4 this spring (likely March/April).

Apple first explored partnerships with OpenAI and Anthropic but neither panned out, showing its in-house AI still lags frontier-level. Gemini will embed top-tier reasoning into Siri, acting as a temporary bridge to finally deliver the personalized, agentic features stuck in development since 2024.

Apple also launched Creator Studio, a $12.99/month (or $129/year) subscription bundling pro apps: Final Cut Pro, Logic Pro, Pixelmator Pro (now on iPad with full Apple Pencil support, advanced layers, masking, and hardware-accelerated AI like Super Resolution), Motion, Compressor, and MainStage.

The bundle also adds AI tools and premium content in Pages, Numbers, and Keynote (Freeform coming soon). AI features accelerate workflows, delivering professional tools at a fraction of the cost, family sharing included. A clear shot across the bow at Adobe. Read more.

• NVIDIA fixes LLM memory with real-time learning

• Zuckerberg launches ‘Meta Compute’ and stacks the leadership deck

• Google’s Veo 3.1 drops the upgrade creators needed

• Microsoft vows self-funded energy, no tax breaks, and net-positive water use at data centers

• Meta lays off 10 percent of Reality Labs to chase smart glasses

• Watch: Hyundai shows off K-pop dancing robot dogs and humanoid robot Atlas at CES

• McKinsey now has 25,000 AI agents working alongside 40,000 humans

• Signal creator launches Confer to make AI conversations untouchable by hackers, law enforcement, or servers

• Watch: Google Cloud + NVIDIA Blackwell unlock on-prem AI and high-performance robotics for industries with strict data needs

• Google co-founders may be leaving California

• Advanced AI in China hits a speed bump as Nvidia H200 imports are restricted to “special cases.”

• The AI analyst that connects to all your data and learns your business

• Peking University builds a fourfold faster computing system, hinting at the future of AI hardware

• Watch: China’s MATRIX-3 robot moves like a human and learns new tasks without training

• The new biologists treating LLMs like an alien autopsy

• Watch: Xiaomi EV powered by advanced AI traction tech just beat a Ferrari in the desert

• Senate passes landmark bill letting victims sue over AI deepfakes

• Your Slackbot just got a huge AI agent upgrade 

• AI age verification on Roblox backfires as kids, adults, and predators all slip through

• Researcher affirms human creativity's value amid AI

• DeepSeek has hidden an Easter egg in the code

• This AI dashcam sees the road with human-like depth perception

• Pentagon outlines AI-first future

• Doctors think AI has a place in healthcare — but maybe not as a chatbot

• Manus partnered with Similarweb to bring access to global web traffic data directly to Manus users

• Neo humanoid maker 1X releases world model to help bots learn what they see

• Engram by DeepSeek promises faster, smarter memory access for future LLMs

• Nvidia and Eli Lilly are teaming up on a $1 billion AI lab to automate drug discovery from molecule to production

• Pentagon embraces Musk’s Grok AI for unclassified and classified operations

• This new tool could tell us how consciousness works 

• Watch: Meet LimX COSA, the first agentic OS built for real-world humanoid robots

• Watch: NEO humanoid robot can now teach itself new skills using video-based AI models

• Former Google CEO Eric Schmidt is building a space telescope bigger than Hubble—for a fraction of the cost

• By blocking third-party coding agents, Anthropic may have handed OpenAI a billion-dollar opening

• California prisons are turning VR into a “hope machine” for inmates reentering society

• Samsung wants AI to vanish into your devices

• Trump warns Big Tech: Americans won’t foot the bill for AI’s energy appetite

• Hands-on with Bee, Amazon's latest AI wearable

• A New Jersey lawsuit exposes how nearly unstoppable deepfake porn apps evade the law

📄 VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice

📄 RoboVIP: Multi-View Video Generation with Visual Identity Prompting Augments Robot Manipulation ​​

📄 Over-Searching in Search-Augmented Large Language Models 

📄 Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning 

📄 BabyVision: Visual Reasoning Beyond Language

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!