PLUS: OpenAI's new Mac AI agent and a surprising warning from a transformer creator
Good morning
The playing field for building powerful AI agents just got a lot more level. Meta has open-sourced its comprehensive toolkit, providing developers with the foundational infrastructure to build, train, and deploy complex agentic systems.
The move lowers the barrier to entry for creating advanced agents, a capability once reserved for tech giants. Will this flood of new tools accelerate the development of genuinely useful agents, or will it create a new set of complexities for the broader community?
In today’s Next in AI:
Meta open sources its AI agent toolkit
OpenAI's new Mac AI agent
A transformer creator's warning for AI
New financial rails for the AI agent economy
Meta's Agentic AI Toolkit
Next in AI: Meta just released a comprehensive, open-source toolkit designed to help developers build, train, and deploy the next generation of AI agents using PyTorch. The new stack covers the entire development lifecycle, from low-level hardware interaction to large-scale reinforcement learning.
Decoded:
The toolkit supports end-to-end development with libraries like Helion for writing custom GPU kernels with 4x less code, TorchComms for scaling communication across over 100,000 GPUs, and ExecuTorch 1.0 for deploying models on edge devices.
In a major collaboration, Meta and Hugging Face launched OpenEnv, an open hub for sharing and testing reinforcement learning environments, aiming to standardize how agents are trained and evaluated.
A core design principle is simplifying cluster-scale AI, with libraries like Monarch abstracting away multi-node complexity to make large-scale training feel more like working on a single local machine.
Whit It Matters: Meta is effectively open-sourcing the foundational infrastructure needed to build complex AI agents at scale, a capability previously held by only a few large tech companies. This move provides developers everywhere with the building blocks to create and experiment with powerful new agentic systems.
OpenAI Comes for Your Desktop
Next in AI: OpenAI has acquired Sky, a startup building an AI assistant to automate tasks directly on macOS. This signals a major strategic push to move beyond chatbots and create AI agents that can actively use your computer.
Decoded:
Sky acts as a native macOS assistant that can understand what's on your screen and take actions in your apps, like adding calendar events from a text message.
This move places OpenAI in direct competition with rivals like Anthropic and Google, who are also developing similar "computer use" capabilities to automate desktop workflows.
The acquisition deepens the ties between the two companies, building on the existing partnership with Apple to integrate ChatGPT into Apple Intelligence.
Whit It Matters: OpenAI is aggressively moving from AI that answers questions to AI that takes action directly within your operating system. This points toward a future where you instruct your computer on tasks rather than manually executing them across different applications.
Transformer Co-Creator Moves On
Next in AI: In a striking critique, Llion Jones, one of the eight authors of the seminal "Attention Is All You Need" paper, announced he is moving on from transformers, warning that the AI field's singular focus on the architecture is stifling the next major breakthrough.
Decoded:
Jones argues that today’s high-pressure environment, driven by unprecedented investment and intense competition, forces researchers to pursue safe, incremental gains rather than take creative risks.
He contrasts this with the creation of the transformer itself, which he says was born from an organic, bottom-up process with the freedom to explore without direct pressure from management to publish or hit specific metrics.
Putting his philosophy into action, Jones is now dedicating his time at his new company, Sakana AI, to explicitly searching for the next dominant architecture beyond transformers.
Whit It Matters: This is a powerful warning from a key architect of the current AI boom that the industry’s greatest strength could also be its biggest blind spot. His call to “turn up the explore dial” challenges the prevailing scale-is-all-you-need mindset, suggesting the next great leap in AI may come from a lab that dares to look elsewhere.
The Economy for AI Agents
Next in AI: Visa and Mastercard are partnering with Cloudflare to create the financial rails for AI. They are building secure protocols that allow AI agents to safely browse, compare, and purchase goods on your behalf.
Decoded:
The core challenge is trust, as merchants currently struggle to distinguish helpful AI agents from malicious bots, but this new system gives agents a verifiable identity.
This is built on a proposal called Web Bot Auth, which uses cryptographic signatures in HTTP requests to prove an agent is legitimate and authorized.
To simplify development, Cloudflare will integrate Visa's Trusted Agent Protocol and Mastercard's Agent Pay into its Agent SDK, making it easier to build compliant agents.
Whit It Matters: This collaboration creates the foundational trust layer needed for AI agents to handle real money and real transactions. The entry of major payment networks signals that agent-based commerce is moving from a concept to an impending reality.
AI Pulse
Apple began shipping its US-built AI servers from a new Houston factory ahead of schedule, which are designed to power Apple Intelligence and its privacy-focused Private Cloud Compute features.
Google expanded its partnership with Anthropic, agreeing to supply the AI lab with up to one million of its custom Ironwood TPUs in a major deal for AI model training infrastructure.
Microsoft unveiled Mico, a new animated, emoji-like face and personality for its Copilot assistant, in a "human-centered" rebranding effort that draws direct comparisons to the iconic Clippy.
OpenAI faces an updated lawsuit from the parents of a teen who died by suicide, alleging the company twice weakened ChatGPT's safeguards around self-harm discussions prior to his death.