[ The Morning Build ]
4 min read5 storiesAIChipsDev ToolsPolicy

The Morning Build for October 22, 2026: Anthropic Duet, Flint on Laptops, TorchTPU vs Nvidia, EPA Ruling, Mercury 3

Today’s stories center on practical shifts engineers face: always-on agents running in the cloud (Anthropic Duet), a laptop-capable coding model released under Apache 2.0 (Recursive’s Flint), vendor pushes to decouple PyTorch from Nvidia with TorchTPU and TorchNeuron, a federal judge blocking an EPA power grab that would have curtailed citizen Clean Air Act suits tied to data centers, and a diffusion language model claiming large speed and token limits (Inception Mercury 3). Each story affects deployment choices, hardware targeting, or legal exposure for teams building AI services.

Anthropic previews always-on agent Claude Duet that runs on cloud VMs and waits for user approval before risky actions

  • What happened: Anthropic released a research preview of Claude Duet, an always-on agent app built on Claude Opus 5.5 that runs in its own cloud virtual machine, connects to more than 6,000 connectors (Gmail, Slack, Stripe, DocuSign, Shopify), and prompts users for approval before sending messages, making payments, or signing documents. Pro and Max subscribers in the US can access Duet through a waitlist; Team and Enterprise plans are not included yet.
  • Why it matters: Engineers integrating persistent agents must account for background checks that count against shared Claude subscription limits, connector-based access patterns, and Anthropic’s credential handling that keeps passwords and tokens out of agent-visible text. The VM-based browser approach shifts compute and security boundaries off users’ devices and onto Anthropic’s cloud VMs.
  • Outlook: The Pro and Max US waitlist rollout and early tester feedback will reveal how frequently Duet polls connected accounts, how subscription limits affect background checks, and which Duet features change before broader Team and Enterprise availability.

Sources: the-decoder.com · arstechnica.com · the-decoder.com

Recursive releases Flint, a 40B-parameter mixture-of-experts coding model built largely by its ‘Eureka Machine’, available on Hugging Face under Apache 2.0

  • What happened: Recursive launched Flint, a free coding model designed to run on laptops that uses a mixture-of-experts architecture (activating about 2.5 billion parameters at a time from a 40 billion parameter base) and is published on Hugging Face under the Apache 2.0 license. Recursive says its automated research system, the Eureka Machine, contributed across model development and used techniques including low-bit training and custom kernels.
  • Why it matters: Teams wanting on-device coding assistants can run Flint offline to avoid cloud usage fees and keep code local; the mixture-of-experts shape and low-bit training aim to lower memory and latency requirements compared with very large dense models. License terms (Apache 2.0) permit modification and redistribution.
  • Outlook: Hugging Face download and community test results, plus independent SWE-Bench Verified scores from external testers, will be the first public checks on Flint’s real-world performance and stability outside Recursive’s benches.

Sources: techcrunch.com

Google publishes TorchTPU to run PyTorch apps on TPUs as it and AWS push to lessen Nvidia’s PyTorch default

  • What happened: Google announced TorchTPU, an open source PyTorch variant that lets developers run existing PyTorch apps on Google TPUs with configuration changes rather than code rewrites, and released a TPU agent suite and LiteRT updates to ease porting. Amazon released TorchNeuron and is contributing code to the main PyTorch project to decouple features from Nvidia hardware.
  • Why it matters: AI app developers gain practical migration paths to TPU and Trainium hardware without rewriting PyTorch code, changing deployment targets and cloud vendor economics; contributions to upstream PyTorch could alter default device assumptions embedded in many models and libraries.
  • Outlook: Upstream PyTorch commits that incorporate Google and Amazon changes and early adoption metrics for TorchTPU and TorchNeuron will indicate whether these projects reduce the default dependency on Nvidia GPUs in deployed AI apps.

Sources: techcrunch.com · the-decoder.com

Federal judge rejects claim that the EPA can unilaterally dismiss citizen Clean Air Act suits challenging data center pollution

  • What happened: US District Judge Debra Brown ruled that nothing in the Clean Air Act authorizes the EPA Administrator to unilaterally dismiss citizen suits, rejecting an effort tied to former President Trump to block an NAACP lawsuit over unpermitted gas turbines powering a SpaceXAI Colossus data center in Memphis. The NAACP sought an injunction requiring best available pollution controls and penalties up to $124,426 per day per violation.
  • Why it matters: The ruling preserves citizens’ ability to sue under the Clean Air Act for alleged data center pollution and keeps civil penalties and injunctive relief on the table for operators; engineering and site teams maintaining or activating on-prem or cloud-adjacent turbines should factor continued citizen suits into compliance risk and remediation planning.
  • Outlook: Further filings and court orders in the NAACP v. SpaceXAI litigation will determine whether injunctive relief or per-day penalties are imposed at the Memphis Colossus site.

Sources: arstechnica.com

Inception launches Mercury 3, a diffusion-based language model claiming up to 1,068 tokens/sec, 1M-token context, and leads on 8 of 12 internal benchmarks

  • What happened: Inception released Mercury 3, a diffusion language model that refines a draft iteratively rather than token-by-token, claims 1,068 tokens per second on Nvidia GPUs, supports up to 1,000,000 token contexts, and reports wins on eight of 12 internal benchmarks versus GPT-6 Luna and Claude Haiku 5.5. The model is available via Inception’s platform and third-party vendors and uses an OpenAI-compatible API.
  • Why it matters: Mercury 3’s diffusion approach and very long context support target real-time agent and retrieval-heavy apps where latency and context window matter; the claimed throughput and token limits may change system architecture for multi-step agent workflows and long-document tasks if independent benchmarks confirm those figures.
  • Outlook: Independent benchmark results and the end date for Inception’s launch price cut will be the next public checks on Mercury 3’s speed, quality trade-offs, and commercial cost profile.

Sources: the-decoder.com