AI News Roundup: October Twenty - The Autonomy Bill Comes Due

Overview: Tuesday's AI news was about who pays for autonomy. Google put its PyTorch stack for TPUs into open source, one day after Amazon did the same for Trainium, turning the Nvidia default into a menu of quotes. Figma added 20 MCP partners while its co-founder said the protocol is tilted toward frontier labs. Yelp put a voice agent on the phone with restaurants. A breach at OpenCode turned stored API keys into other people's bills. And Anthropic, the loudest voice for AI safety legislation, drafted a bill that would let a judge block a model release, while a leaked filing showed its own cloud capacity is running short.

Google Open-Sources TorchTPU to Loosen Nvidia's Grip

At PyTorch Conference 2026 in San Jose, Google published TorchTPU as open source, replacing the old PyTorch/XLA path with eager execution, one process per device, and torch.compile for production. The serving stack came along: native vLLM, SGLang, llm-d, TorchTitan, Ray and Hugging Face support. Amazon shipped TorchNeuron for Trainium a day earlier, so two of the three largest clouds now let existing PyTorch code run on their own silicon after a configuration change instead of a rewrite (Google Cloud, TechCrunch). Open source does not make a TPU faster than a GPU. It makes it priceable, and pricing power is what CUDA actually protected.

Figma's Field Calls MCP Imperfect as Partners Pour In

Figma announced 20 new MCP partners at Config India, including Notion, Linear, Mistral, ClickUp, JetBrains' Junie and Meta Muse. In the same briefing, co-founder Dylan Field called MCP "not a perfect protocol" with "tons of room for improvement," adding that it is "advantaged more to frontier labs" (TechCrunch). His argument is about reciprocity, not openness: the labs consuming context through MCP do not return artifacts and cloud resources created on their side. Standardized access is not the same thing as standardized symmetry.

Yelp's AI Agent Now Calls Restaurants to Check Wait Times

Yelp's fall release introduced Check for Me: enter a party size and phone number on a restaurant page in the iOS app, and a voice agent calls the restaurant, asks for the current wait, and texts the answer back within minutes (TechCrunch, Yelp). It is a deliberately narrow agent with a fixed script, which is exactly why it can be reliable. The open question is agent-to-agent: Yelp already sells voice automation to restaurants, so both ends of the call may soon be bots.

OpenCode Breach Exposes Developer API Keys and Runs Up Bills

Attackers broke into an older system behind OpenCode Console and took credentials belonging to 1.4% of accounts. The damage surfaced as billing alerts across OpenAI, Anthropic and Google, with some developers reporting charges above $1,000. All three providers blocked the exposed keys and OpenAI refunded burned credits (TechCrunch). The more instructive failure came during the response: OpenCode's own defensive agents could not tell an incident responder from an attacker, so the guardrails blocked the cleanup.

Anthropic Pitches a Safety Bill That Could Block a Model Release

Anthropic has been shopping a draft AI safety bill to Congress, first reported by Politico. It defines catastrophic risk as a "foreseeable, imminent risk" of more than 50 deaths or a billion dollars in damage from a bioweapon, a model evading its developer's controls, or a cyberattack - the same thresholds as California's recently passed SB 53. Registered independent evaluators, potentially paid by the labs but not on contingency, would get 72 hours to notify the Commerce Department, which could then ask a D.C. judge to halt distribution (TechCrunch). No agency could stop a model on its own authority.

The same day, a leak of Anthropic's S-1 to the Wall Street Journal showed the company struggling to secure cloud capacity from AWS, Google and Microsoft, which increasingly treat it as a competitor. The filing implies more than $750 billion in potential cloud commitments against a first-half net loss of $80.7 billion on $16.3 billion in revenue (TechCrunch). A company writing its own regulator's rulebook is also the one with the most concentrated compute exposure.

What to Watch Tomorrow

• Anthropic's IPO path. A November listing is reportedly targeted, and the leaked filing is now the main source of numbers on its cloud dependency.
• Otter.ai's $320 million. General Catalyst's Customer Value Fund is funding growth without equity dilution, which makes the revenue-without-churn question the one to watch (TechCrunch).
• OpenAI's biotech build-out. OpenAI acquired Orchestra Bio to expand its life sciences work, continuing a pattern of buying domain data rather than training for it.
• Chat lock in ChatGPT. Chat-level locking turns a privacy feature into a retention feature, and it will be copied.

Builder Impact

• Compute is becoming a quoting exercise. With TorchTPU and TorchNeuron both open source, portability is now a config decision. Teams that benchmark a second accelerator before signing a multi-year GPU commitment have leverage they did not have last week.
• Protocol access is not protocol symmetry. If your agents read from MCP servers but publish nothing back, you are the consumer in Field's asymmetry argument. Design what your systems expose, not only what they can reach.
• Stored provider keys are a liability, not a convenience. OpenCode's breach was a design failure before it was an intrusion. Prefer per-user keys or a proxy that can be revoked in minutes, and test whether your guardrails recognize your own incident responders.
• Narrow beats general in production voice. Yelp shipped a scripted task, not a conversationalist. Scope the agent to one job with a fixed script and a bounded answer format.
• Regulatory text is now written by the regulated. Read Anthropic's thresholds - 50 deaths, a billion dollars, 72 hours - as the likely template for whatever Congress actually debates.