The Information Machine
The edition

Saturday November 14, 2026

This is one edition of my daily email.

One free email each morning, around 8 AM US Eastern.

New today

01
New

Thinking Machines Inkling 2 Small

  • Thinking Machines released Inkling 2 Small on November 5 with weights on HuggingFace under Apache 2.0, then confirmed it free on OpenRouter and compatible with Claude Code, Codex, and Hermes Agent.
  • Benchmark results placed it 19th of 34 in Vals AI's index, with strong finance and legal scores but near-zero on SRE Bench and ProgramBench; VentureBeat reported it outperformed Mistral Large 4 on enterprise tasks including DeepSWE and Finance Agent.
The gist

A 594B open-weight model under Apache 2.0 is accessible to a wide range of developers and researchers without licensing restrictions. Benchmark results show meaningful gaps across task types, so the model's practical fit depends heavily on the application.

02
New

IFM, Runway, and NVIDIA world action models

  • Runway extended Praxis-1, its world action model released November 11 under Apache 2.0 for robotics, to run on the SO-101 robot platform on November 13.
  • The model was trained on thousands of community-shared LeRobot datasets and runs on a single RTX 5090 with NVFP4 precision; Runway made the checkpoint and setup instructions publicly available.
The gist

Both models are released as open weights with permissive licenses, making physically grounded world models available for robot training and simulation outside closed commercial systems. Runway's extension to the SO-101 platform and community datasets broadens the hardware surface where Praxis-1 can run.

03
New

ChatGPT's AI payment wallet

  • OpenAI launched ChatGPT Wallet on November 13, enabling AI agents in ChatGPT Work, Codex, and dots to complete purchases through standard checkout pages, with card tokenization via Visa and Mastercard and biometric approval required per transaction.
  • The rollout covers US Pro subscribers only, EU availability is absent, and United Airlines, Amazon, and eBay are already blocking agent-driven purchases; OpenAI released a Commerce Partner Playbook naming Stripe, Adyen, and Riskified as integration paths for merchants.
  • The OCC published guidance classifying autonomous AI payment agents in its highest risk tier, alongside real-time credit decisions and market access.
The gist

ChatGPT Wallet introduces AI agents making financial transactions directly on behalf of users at scale, raising new questions about authorization, fraud, and regulatory oversight. The OCC's classification of autonomous AI payment agents in its top risk tier signals that US bank regulators are actively developing frameworks for this capability.

04
New

The OpenAI Ukraine prank call

  • Russian pranksters Vovan and Lexus, impersonating Ukrainian intelligence chief Rustem Umerov, tricked three senior OpenAI national security officials into a 17-minute video call in which the officials discussed AI support for Ukraine, possible offensive uses, Pentagon contacts, and a program called Daybreak Blue.
  • The pranksters had first deceived a real Ukrainian official in San Francisco who arranged the meeting through OpenAI security head Sasha Baker.
  • They published the recording on Rumble; OpenAI says it investigated and tightened procedures.
The gist

The incident exposed how senior AI company officials with national security responsibilities can be socially engineered into disclosing sensitive operational details about military AI programs. It raises questions about vetting procedures for high-level calls involving defense-related work.

05
New

DeepSeek chat app capabilities opening to developers

  • DeepSeek announced on 2026-11-13 that the full feature stack behind its consumer chat app is now accessible via its API, including search, file handling, link processing, content safety, and voice.
  • Developers can enable built-in features through a few parameters or call standalone endpoints for OCR, ASR, and TTS.
  • An open-source frontend demo supporting DeepSeek conversation history import is also available.
The gist

Developers can now build applications that replicate or extend the full DeepSeek App experience using the API. The addition of standalone OCR, ASR, and TTS endpoints expands the range of multimodal capabilities available programmatically.

Get the daily email

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy

Updates

06
Day 5

Jensen Huang's alleged Chinese intelligence targeting

  • Jensen Huang resigned as Nvidia's CEO on November 11; CFO Colette Kress became interim CEO and shares fell more than 13%, after the board gave him no choice when a second relationship, with Nvidia executive Kimberly Powell, surfaced as a final straw in the probe over his ties to Alice Huang, a suspected Beijing intelligence recruit.
  • Congressional scrutiny of Commerce Secretary Lutnick broadened, with Democratic lawmakers launching a formal investigation into Nvidia chip sales to Chinese companies and Senator John Kennedy demanding Lutnick testify publicly.
The gist

The resignation of Nvidia's founding CEO under espionage-linked circumstances removes the company's most prominent leader and forces an immediate succession decision at the center of global AI chip supply. The case has drawn congressional investigations and placed a cabinet secretary under removal consideration.

07
Day 24

SpaceX Super Intelligence and Tesla merger talks

  • SpaceX's official campus announcement on November 12 added job creation projections absent from earlier Reuters reporting: more than 10,000 construction and over 1,500 permanent positions across its two planned AI supercomputing sites in Arizona and Texas.
  • The Arizona application, filed October 30 by SpaceX affiliate Tombstone Resources LLC, leaves the campus power source unresolved.
The gist

SpaceX's AI segment has become its largest revenue source in a single year, and the $74 billion campus commitment is among the largest declared AI infrastructure investments by any single company. The Terafab semiconductor project and reported merger talks with Tesla would, if realized, create a combined entity spanning rockets, satellites, AI compute, chips, and consumer robotics.

08
Day 26

Anthropic IPO and compute commitments

  • At a JPMorgan event with CEO Jamie Dimon, Dario Amodei confirmed Thursday's Nasdaq listing and said compute is 'the deepest business dilemma that the company faces,' noting that Anthropic limits upfront capacity and relies on the spot market when demand is strong.
  • The IPO's stated primary purpose is to stockpile cash for compute costs; Anthropic held over $130 billion in cash at end of September, and planned SpaceX data center contracts run up to $3.1 billion per month.
The gist

Anthropic is proceeding to a public listing with over $895 billion in compute commitments, internal disclosures that dramatically accelerated AI progress may be imminent, and a compute strategy that relies partly on expensive spot-market capacity. The RSP filing language and independent METR evaluation make the IPO the first time a frontier AI lab has publicly disclosed, under securities law, that it may be approaching recursive self-improvement thresholds.

09
Day 9

METR's Anthropic AI capabilities audit

  • At a November 2026 public lecture in Toronto, Paul Cristiano, an OpenAI Foundation board member and Alignment Research Center director, told the audience that an AI accident is 'the most likely reason they will die' and described a takeover as goal-pursuing AI systems controlling Earth, with any individual as powerless as someone trying to stop the US military.
  • The lecture elaborated his prior statement of a material risk of AI takeover in 6-12 months, made after METR found Anthropic accelerating model capabilities at roughly 1.9x its historical rate.
The gist

If the acceleration is genuine and sustained, AI capabilities could advance far more rapidly than most external projections assumed, compressing timelines for transformative AI. The heavy redactions mean the public cannot independently verify METR's conclusions, requiring trust in METR's judgment at a moment when researchers say the stakes are very high.

10
Day 5

Hyperscaler Q3 earnings and AI supply crunch

  • Applied Materials reported record revenue of $10.41 billion, up 53% year-over-year, and guided the next quarter to $11.1 billion, with management expecting global chip equipment spending to grow more than 30% in 2027; Cerebras posted Q3 core revenue of $230.9 million, up 70%, and raised its full-year guidance.
  • DDR5 prices reached 877% above August 2025 levels by November 2026 as Samsung committed 46 trillion won to expand HBM4E production at its Onyang campus with packaging lines by June 2028.
  • Morgan Stanley projected Google at roughly 35% of global HBM demand in 2027, nearly matching Nvidia's projected 36%.
The gist

Earnings across chip equipment, foundries, memory, and AI chip makers confirm that AI infrastructure demand is running ahead of supply across multiple categories, with new capacity in lithography, HBM, and NOR flash not expected before 2028 to 2030. The scale of capital commitment, with JP Morgan projecting hyperscaler capex above $2 trillion by 2028, is drawing scrutiny from central banks over inflation and from investors over whether returns will materialize.

11
Day 2

Google's Gemini 4 Hydrogen launch

  • Google's Hydrogen model, released November 12 as the free-tier replacement for 3.8 Flash, is accessible from the Gemini app dropdown, the Gemini API, and AI Studio; Google's coding tool Antigravity already orchestrates Carbon and Hydrogen in one session.
  • In xhigh mode, Hydrogen scores 54 on the Artificial Analysis Intelligence Index versus Argon's 50.
  • Simon Willison noted that 'considering how close they are in price, the difference in pelican quality for Hydrogen is quite stark'.
The gist

Hydrogen replaces the widely used Flash 3.8 as Google's free-tier and low-cost model, making it the default AI layer in Google Search for all users. Its pricing sits well above comparable lightweight models from competitors, which may shape developer choices in cost-sensitive applications.

12
Day 5

Meta's Muse Ember small model launch

  • Simon Willison argued that per-token pricing is a poor basis for comparing reasoning model costs because variable token usage makes cost per completed task the more meaningful measure, and that Muse Ember, Meta's small model for agentic work released November 9, scores well on that metric.
The gist

Muse Ember enters the competitive small and efficient model tier at the same price point as comparable models from OpenAI and Anthropic, with Artificial Analysis placing it on the efficiency frontier. The cost-per-task framing matters because reasoning models consume varying numbers of tokens per task, making raw per-token price comparisons potentially misleading.

What is moving now · All editions · All stories