October 21

Your Daily Note,

Anthropic wants a legal threshold for blocking dangerous AI models. OpenAI is moving deeper into drug development. Google is making it easier for developers to leave Nvidia’s chips behind, and Meta’s Muse app has passed 10 million downloads.

In today’s Noted:

Anthropic proposes a law that could block an AI model over catastrophic risks

We noted:

Anthropic is circulating a draft bill among US lawmakers that would allow the government to block a model’s release or use, but only if it poses a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people.

The proposal is not yet law. Anthropic says it reflects the Advanced AI Framework it has promoted with federal and state lawmakers since June.

Why it matters

The draft prevents federal agencies from stopping a release on their own, so the legal process is the central safeguard. Questions remain about independence: companies could pay the evaluators, as long as the payment is not tied to the outcome. The draft would also preempt certain state requirements on risk disclosures, evaluations and incident reporting.

Google makes it easier to move AI applications away from Nvidia

Image source: Google Cloud

We noted:

Google is working on TorchTPU, a project to make its Tensor Processing Units more compatible with PyTorch, the open-source framework most AI developers use. Reuters first reported the internal project in December 2025, saying Google is considering open-sourcing parts of it. Google has since confirmed the project, and an engineering lead says the TPU should become an obvious choice for any PyTorch user.

PyTorch has been closely tied to Nvidia’s CUDA software, and earlier TPU efforts left many teams with significant extra engineering work. The aim is to make TPUs more developer-friendly for organisations already built around PyTorch.

Why it matters

Developers may get more options for running AI applications on different chip architectures. A simpler migration path does not guarantee identical performance, compatibility or cost, so teams still need to test their workloads before switching.

OpenAI acquires Orchestra Bio to bring AI deeper into drug development

We noted:

OpenAI has acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex work of drug development. The company says it is bringing Orchestra’s team and technology together with Rosalind, its AI models and tools for life sciences research. The announcement, dated 20 October, does not include financial terms.

According to OpenAI, developing a new medicine typically takes around 10 to 15 years from target discovery to regulatory approval in the United States. Teams must coordinate programmes across internal departments and external partners, often relying on disconnected systems.

Why it matters

The acquisition does not mean AI-generated treatments will reach patients sooner, or that clinical trials can be skipped. It does show AI companies moving beyond general-purpose chatbots into specialised industries, where software must understand complex workflows. Buying a team with existing domain expertise can be faster than building that knowledge from scratch.

Meta’s Muse passes 10 million downloads

Image source: Meta Muse

We noted:

Meta’s AI agent app Muse has more than 10 million downloads, according to a BNP Paribas analysis of third-party data published on 16 October. Muse launched on 8 September and is available only in the US and Canada.

Sensor Tower had earlier placed the app above 5 million downloads by 30 September, about 22 days after launch. Meta began ramping advertising spend for Muse on 16 September, and Sensor Tower said downloads rose 73% day over day that day.

Why it matters

Muse’s early growth shows how distribution across Meta’s own platforms can bring an AI app to a large audience quickly. These are third-party estimates, and different trackers disagree, so they are not audited company figures. Strong early downloads do not guarantee growth at the same rate or lasting use.

Anthropic’s revenue is growing fast, and so are its infrastructure needs

We noted:

Anthropic told investors on 14 October that its third-quarter 2026 revenue rose to $16.8 billion, according to reports by the Financial Times and Reuters. Reuters reported first-quarter revenue of $4.7 billion and second-quarter revenue of $11.6 billion, which together make $16.3 billion for the first half.

Growth is still strong but slowing. Revenue rose about 147% between the first and second quarters, and about 45% between the second and third.

Why it matters

Revenue is not profit, and the larger loss in the prospectus comes mostly from accounting charges, so read the headline numbers carefully. For businesses building on AI APIs, the practical question is whether providers can keep supplying the capacity, performance and reliability they need.

Products worth knowing

 

Everything else in AI

NVIDIA releases Cosmos3-H-Dreams, a real-time surgical simulator

NVIDIA has released Cosmos3-H-Dreams, a real-time generative simulator for robotic surgery. It generates surgical scenes that respond to instrument movements, and it is trained on video and robot kinematics from clinical cholecystectomy procedures performed on CMR Surgical’s Versius platform.

The model builds on Cosmos3-H-Embodied-Nano, a 16-billion-parameter model, and runs interactively on a single RTX PRO 6000 GPU. It succeeds Cosmos-H-Dreams, which focused on tabletop suturing tasks.

NVIDIA calls it a research simulator, not a validated system for patient care. Intended uses include evaluating surgical robotics policies in simulation, exploring interactive world models and generating synthetic data. A convincing simulation is not, by itself, evidence that a system is safe for surgery.

 

ChatGPT now lets you lock sensitive chats behind a PIN

OpenAI has launched Locked Chats, which lets users protect selected conversations with a four-digit PIN, Face ID or Touch ID. Locked chats move out of regular chat history, pinned chats and search results, and are grouped under a “Locked chats” section.

Locked chats are not end-to-end encrypted. OpenAI says locking a chat does not change how long it is retained or the “Improve the model for everyone” setting. With Memory enabled, locked chats have a separate memory that does not flow into regular chats. Resetting the PIN permanently deletes all locked chats and their protected files.

The feature is available on Free, Go, Plus, Pro, Business, Enterprise and Edu plans, but not on ChatGPT for Healthcare. Temporary, archived, group and shared-link chats cannot be locked. OpenAI says the rollout is gradual.

Why it matters: locked chats protect against people who can see your screen or device, but they do not change what OpenAI stores or how it trains on data.

Worth your weekend

See you tomorrow
same time, different note.

~ Stackari

 

P.S. Hope this made your day a little more interesting. More News Tomorrow:)