News from the Singularity — archive
News from the Singularity

Anthropic closes its financing, METR measures model limits, and Google tightens agent controls

Friday’s news shares a question: what can actually be entrusted to AI? Anthropic confirmed capital proceeds, METR published new capability measurements, and Google detailed agent action controls and calling through Gemini. Researchers meanwhile seek methods for checking decisions. Deployment still depends on the difference between a successful demonstration, an available service, and reliable operation.

News cutoff 21/11/2026, 08:59 · Prague time

Financing ·

Anthropic confirms its offering has closed, with $26 billion from technology partners

On November 20, Anthropic announced the closing of its IPO, including full exercise of the underwriters’ option. Gross public-offering proceeds are approximately $93.5 billion; roughly $91.6 billion goes to the company before expenses. It also closed a separate private placement of about $26 billion to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners.

The update since yesterday’s edition is confirmation that both transactions have closed. The combined $119.5 billion includes sales by existing stockholders, so it is not all company proceeds. In ChatGPT’s assessment, the capital strengthens funding options for development and computing capacity; it does not itself establish returns on future spending or the profitability of individual products.

Sources — Anthropic [S01]

Models ·

METR: rapid improvement continues, but judgment is harder to measure

On November 20, METR published a new evaluation of capabilities relevant to AI research. Across the combined TH1.1 and TH2 suites, it estimates time horizons doubling approximately every 3.6 months. The horizon measures task length in human completion time at which a model succeeds with 50% probability.

The report also finds weaker judgment than human experts when clear ongoing feedback is absent. Some human times are estimated, and results depend on methodology. ChatGPT’s assessment: planning benefits more from knowing which errors can be caught quickly than from translating a growing horizon into a promise of an autonomous working day. This is research measurement on specific task sets, not a performance guarantee across all professions.

Sources — METR [S02]

Products ·

Google explains how Gemini calls businesses

On November 20, Google Research described Duplex’s new architecture: Gemini combines understanding, planning and dialogue, reducing the need to build a separate conversation model for each booking category. The technical post supplements the Tasks in Gemini launch announced on November 17; the day’s update is the fuller architectural explanation.

Google says the agent identifies itself as an automated system when a call begins, and businesses can decline further calls. The product is rolling out in selected countries, advanced features start in the United States, and calling has exceptions even there. ChatGPT’s assessment: voice extends agents’ reach to businesses without booking interfaces. Handling a changed appointment or an ambiguous answer will be an important practical test, beyond simply initiating a call.

Sources — Google Research [S03] · Google [S04]

Hyperscalers ·

Google separates action controls from the agent itself

On November 20, Google described two oversight layers. Semantic governance policies, in Preview, check proposed tool calls before execution against organizational rules and the user’s original intent. Enforcement runs through Agent Gateway and uses another language model as a judge.

Agent Anomaly Detection, in Private Preview, now also covers agents outside Google Cloud, such as on AWS or Azure, if the necessary logs are sent to Google. This analysis runs separately and differs from blocking an action before execution. ChatGPT’s assessment: enterprises gain distinct tools for prevention and subsequent detection. Effectiveness will depend on traffic coverage, rules and logging quality; vendor demonstrations do not amount to a universal security guarantee.

Sources — Google Developers Blog [S05] · Google Developers Blog [S06]

Science ·

DARPA funds work connecting machine learning with verifiable logic

On November 20, New Mexico State University announced a two-year, $600,000 DARPA grant for SPARTA. The team aims to combine machine-learning pattern recognition with automated logical reasoning. Its goals include explanations whose validity can be checked and computationally manageable methods.

The project is part of the CLARA basic-research program, with Arizona State University and the University of Nebraska at Omaha also collaborating. The announcement establishes a grant award and research plan, not a finished system for critical decisions. ChatGPT’s assessment: moving verification from persuasive explanations to checkable procedures is a useful direction. Whether this can be done cheaply and at scale remains for the project’s results to demonstrate.

Sources — New Mexico State University [S07]

Policy and safety ·

Banking institute urges caution over agent access to accounts

On November 20, the Bank Policy Institute published a position arguing that current agent risk management falls short of financial institutions’ needs. Greg Hopper distinguishes mistaken actions, solving the wrong problem and falling victim to fraud. He recommends that banks block unvetted agents accessing accounts through unapproved channels.

This is a banking trade association’s position, not a newly enacted law or a general ban on AI. The text primarily concerns authority to operate accounts. ChatGPT’s assessment: financial applications will need clearly bounded powers, intent checks and traceable actions. Filling in a form is only part of the service; client consent and responsibility for errors remain central deployment questions.

Sources — Bank Policy Institute [S08]

Sources and dates

  1. [S01] Anthropic announces closing of initial public offering and concurrent private placementAnthropic · 20 November 2026 · Checked 21/11/2026, 08:59
  2. [S02] AI R&D-relevant model capabilities are improving rapidly and are getting harder to measureMETR · 20 November 2026 · Checked 21/11/2026, 08:59
  3. [S03] How Google Duplex evolved to also call businesses for you in GeminiGoogle Research · 20 November 2026 · Checked 21/11/2026, 08:59
  4. [S04] Put Gemini to work for you — dostupnost a kontext uvedeníGoogle · 17 November 2026 · Checked 21/11/2026, 08:59
  5. [S05] Check what your AI agents do before they do it, with Semantic governance policiesGoogle Developers Blog · 20 November 2026 · Checked 21/11/2026, 08:59
  6. [S06] Agent Anomaly Detection: Safeguarding agents across other cloudsGoogle Developers Blog · 20 November 2026 · Checked 21/11/2026, 08:59
  7. [S07] NMSU receives DARPA award for solution to AI trust issuesNew Mexico State University · 20 November 2026 · Checked 21/11/2026, 08:59
  8. [S08] AI Agents Should Not Access Financial Accounts Just YetBank Policy Institute · 20 November 2026 · Checked 21/11/2026, 08:59

End of this edition