Reporting from the Uncanny Valley

Reporting from the Uncanny Valley

By Reporting from the Uncanny ValleyTechnology
Download on the App Store

All Reporting from the Uncanny Valley episodes

  • The Agents Go on Call: AI Systems Take the Incident Shift
    The pager goes off at three in the morning, and this week it started going off for software. Resolve AI's SRE-1 model is built to find the root causes of production alerts; Groundcover's Agent Investigations opens a shared workspace where AI agents and human engineers work the same incident; and Google says, as Quartz reported it, that autonomous agents are outpacing human response. This episode works the story in order: what an on-call agent actually does, what the speed claim measures and what it leaves out, the recursive problem of agents investigating systems that fail because of other agents, the agentic security operations center as the parallel case, and the accountability question waiting in the first serious postmortem: when the agent's root cause call is wrong, whose postmortem is it? Vendor claims are attributed to the vendors throughout and tested, never adopted as findings.
    00:00 The pager goes off at three in the morning
    01:28 What was actually announced: a root cause model and a shared investigation workspace
    02:48 What an on-call agent actually does
    06:41 The speed claim: outpacing whom, measured how
    08:55 The recursion: agents investigating agents
    10:24 The agentic security operations center
    12:27 Whose postmortem is it
    14:38 The pager is not going away this year
    16 min
  • Anthropic Opens Its Books: What the First AI Lab S-1 Actually Shows
    Anthropic IPO news: what the company's S-1 filing and prospectus show about revenue, losses, AI compute costs and governance. On October 26, 2026, Anthropic filed with the SEC for a proposed initial public offering of Class A stock on the Nasdaq under the symbol ANT. The share count and price range are still blank. Noelle reads the filing, including $16.286 billion in first-half 2026 revenue, an $80.748 billion net loss that the filing mostly ties to a $72.388 billion non-cash change in the fair value of convertible notes, derivative liabilities and SAFEs, and a first quarterly operating profit in the second quarter. The episode also covers compute spending and commitments, including a minimum $110 billion Amazon agreement, up to $120.5 billion in SpaceX.ai compute deals and a blank aggregate total. It closes with the risk factors, the dispute with the Department of War, and the long-term benefit trust that can elect a board majority of the public benefit corporation.
    Timestamps
    0:00 Anthropic's securities filing and the $80.748 billion net loss
    1:09 Opening Anthropic's books: what the filing covers
    1:34 IPO filing status: Nasdaq symbol ANT, blank share count and price range
    2:22 Revenue growth from 2023 through the first half of 2026
    3:15 Compute expense compared with revenue
    4:35 Operating expenses and the first quarterly operating profit
    5:24 Net loss, non-cash charge and operating cash flow
    6:44 Customer revenue thresholds
    7:09 The blank compute commitment total
    7:30 Amazon commitment of at least $110 billion
    8:20 SpaceX.ai compute agreements
    9:11 Additional compute estimates outside the commitments table
    11:21 Matching capacity to demand: cloud platforms and chip families
    12:13 Risk factors in the prospectus
    13:26 Dispute with the federal government
    14:48 Public benefit corporation and Long-Term Benefit Trust control
    15:59 What the filing does and does not establish
    18 min
  • The Fed Is Worried AI Agents Could Trigger a Bank Run
    Can AI agents cause a bank run? Federal Reserve Governor Michael Barr warned in a November 16 speech in Milan that AI agents could move customer deposits at high speed and in the same direction, so a bank run could unfold in minutes. He spoke for himself, not the Federal Reserve Board, and described a possible scenario, not an event that has happened. This episode of Reporting from the Uncanny Valley explains how speed, correlated behavior and the loss of friction change a bank run, and why banks already report AI agents trying to act in customer accounts without verification. It compares the scenario with the 2023 Silicon Valley Bank collapse, when customers withdrew $42 billion in one day. It also covers Barr's four proposed guardrails (authentication, consent verification, technical limits and behavior constraints), the OCC's examiner framework in Bulletin 2026-54, and whether banks and regulators act on AI risk to financial stability before it is tested.
    Timestamps
    0:00 Fed Governor Michael Barr's warning in Milan
    1:00 What the episode covers
    1:47 What Barr said and did not say
    2:35 Two points: unverified agents and run speed
    4:13 Barr's four guardrails
    4:35 Fed financial stability report and AI risk survey
    5:09 Why speed changes a bank run
    7:33 Yield chasing and falling deposit stability
    8:20 Silicon Valley Bank in March 2023
    9:51 What is different about an agent-driven run
    12:01 Guardrails and the OCC examiner framework
    14:43 ChatGPT wallet and standing permission
    15:34 Closing assessment and open question
    18 min
  • New York Builds an AI Enforcement Unit Before the Law Takes Effect
    New York Attorney General Letitia James has launched ASSET, a new AI enforcement unit, about six weeks before the New York RISE Act (Responsible AI Safety and Education Act) takes effect on January 1, 2027. This episode of Reporting from the Uncanny Valley explains what the AI safety law requires of large frontier AI developers: published safety frameworks, transparency reports, state registration and critical safety incident reports within 72 hours. Noelle covers who leads ASSET, how the unit works with DIGIT, the Department of Financial Services office that administers the law, and the other AI laws in its brief, including chatbot and data security rules. The episode closes on open questions about ASSET staffing and budget, how AI safety frameworks will be evaluated, the narrow incident threshold, federal preemption of state AI laws and the absence of whistleblower protections.
    Timestamps
    0:00 New York announces the ASSET enforcement unit ahead of the RISE Act
    0:47 Episode roadmap and Ellis Ward's framing on disclosures
    1:39 What ASSET is, per the Attorney General's announcement
    2:32 Where the unit sits and who leads it
    3:17 The RISE Act's path to law and the amended version
    4:06 Who is covered: frontier models and large frontier developers
    4:56 Required frameworks, transparency reports and registration
    6:59 The 72-hour critical safety incident reporting duty
    8:07 Catastrophic risk definition and its threshold
    8:50 Enforcement and penalties for the Attorney General
    9:39 DIGIT at the Department of Financial Services
    10:33 Other laws in ASSET's brief
    11:23 Standing unit versus a single lawsuit
    12:14 Open questions: capacity and evaluation
    13:39 The threshold problem
    14:31 Federal preemption and the missing whistleblower protections
    15:51 Six weeks out and what it means for the Hudson Valley
    18 min
  • Could an Agent Deserve Rights? The Evidence Behind the Soul Question
    Could AI be conscious, and could an AI agent deserve moral or legal rights? Noel Brown starts with Mustafa Suleyman's November 1 post on X, in which the head of Microsoft AI said he has seen no evidence of AI consciousness and called for credible evidence to be made public. Noel Brown then works through five kinds of evidence for AI consciousness and AI welfare: self-report, behavioral continuity, mechanistic analysis, suffering and goal-directedness, and social relationship. For each, he describes the false positive and what it would take to falsify the claim. The episode also asks who decides, whether the builder, a regulator, a court or a scientific body, and it uses Australia's new frontier AI regulator as an example. It separates tools, AI agents and legal persons, and sets out four properties of a testable standard: access, incentive disclosure, falsifiability specified in advance, and theoretical grounding.
    Timestamps
    0:00 Mustafa Suleyman's November 1 post on AI consciousness
    1:22 Five classes of evidence for moral status
    1:45 Self-report as evidence
    2:39 Behavioral continuity
    3:24 Mechanistic evidence
    4:14 Suffering, preference and goal-directedness
    5:35 Social and relational evidence
    6:43 Who decides: builders, regulators, courts, scientists
    7:58 Australia's frontier AI regulator
    9:10 Religious authority and the Washington Post soul question
    9:59 Tool, agent and person as separate categories
    11:21 Legal consequences of a conscious agent
    12:08 Numerical identity across many instances
    12:35 Operational record: Medicare breach and inter-agent failures
    13:52 Capability, danger and moral status
    14:39 Suleyman's commitment to stop development
    15:46 Four properties of a testable approach
    16:35 Closing: from personal commitments to shared standards
    18 min
  • Anthropic's Trillion-Dollar Question: What the IPO Filing Really Reveals
    Anthropic, the company behind the Claude AI model, has filed an S-1 registration statement for its IPO, and reports of a lower valuation have followed. This episode of Reporting from the Uncanny Valley separates what the SEC filing discloses from what the market infers. Noel Brown walks through Anthropic's revenue growth, from $41 million in 2023 to $16.286 billion in the first half of 2026, and four disclosed risks: usage-based token pricing, customer concentration, the layered roles of Amazon and Google as distribution partners, cloud infrastructure vendors and investors, and a governance structure that limits public shareholder voting power. The episode also looks at AI compute costs and infrastructure commitments that run to 2045. It explains why a reported Anthropic valuation of $1.5 trillion to $2 trillion, cited in IT Pro's reporting and based on unnamed sources, is not an official IPO price. It also covers the Axios report of a downward reset in valuation expectations.
    Timestamps
    0:00 Anthropic's IPO filing and the opening question
    0:53 What the S-1 does and does not say about valuation
    2:10 Reported valuation versus offering price versus market value
    2:32 Revenue growth from 2023 to the first half of 2026
    3:24 Risk one: consumption-based revenue and net revenue retention
    5:23 Opus 5.5 cost savings and what they mean for revenue
    6:13 Risk two: customer concentration
    7:22 Amazon and Google as distributors, vendors and investors
    9:19 Risk three: compute spending and commitments through 2045
    11:26 Revenue per employee and what it leaves out
    11:50 Risk four: governance and founder voting control
    13:26 Public benefit purpose versus shareholder interests
    14:16 The Axios valuation reset report and anonymous sourcing
    16:10 What investors must weigh
    18:46 Sources and closing question
    20 min
  • Australia Tells Its Highest Court the Ban Protects Children's Brains
    Australia's High Court has heard two days of argument in the first constitutional challenge to the country's social media ban for children under 16, and the seven justices have reserved their decision. Noel Brown explains the case brought by the Digital Freedom Project and Reddit, the government's argument that the harm comes from platform design such as addictive feeds, and the challengers' claim that the law shuts young people out of political communication. The episode also covers how the Australian ban works, including the "reasonable steps" age verification rules, the fines for platforms, and early eSafety Commissioner results showing many kids still have accounts. It closes with what the case could mean for New York, where the Safe for Kids Act regulates algorithmic feeds and overnight notifications, with final rules taking effect January 25.
    Timestamps
    0:00 Australia's High Court hears the social media age law challenge
    1:16 The two joined challenges: Digital Freedom Project and Reddit
    2:24 The government's case: seven design risks
    3:37 Reddit's arguments on fit and the "worst possible outcome"
    4:56 What the ban takes away: the account, not the viewing
    5:45 Noah Jones and the participation argument
    6:34 How the ban is designed to work
    7:41 Platform duties, penalties and "reasonable steps"
    9:22 Is it working? Account numbers and regulator findings
    11:20 Platform incentives and the cost of compliance
    13:02 The export question and New York's Safe for Kids Act
    14:58 Where things stand and what the ruling could mean
    17 min
  • The Agent in the Cockpit: Saab Flies Its Own AI on Gripen E
    Saab says its own AI agent flew a Gripen E fighter jet through beyond visual range (BVR) combat scenarios in flight tests. This episode of Reporting from the Uncanny Valley, hosted by Noelle, explains what the announcement shows and what it leaves out, including why a flight test is not a combat deployment. It covers Saab's AI Fighter Pilot Academy, the Spherical AI supercomputing effort in Linköping, and plans for crewed and uncrewed aircraft, including loyal wingman drones, to work together. It also asks the questions the release does not answer about human control of military AI: who can override the agent, how quickly, what it may decide, and who answers for its choices.
    Timestamps
    0:00 Saab says its own AI agent flew Gripen E
    0:22 What the story rests on: one Saab announcement
    1:26 Saab's in-house agents and who owns the development chain
    1:48 What happened in the air: BVR combat scenarios
    3:29 Intended roles and Gripen E's software-defined architecture
    4:16 The earlier Helsing Centaur flights, kept separate
    5:25 The AI Fighter Pilot Academy and simulated flight hours
    8:10 Spherical AI and the computing behind the effort
    10:54 Why a test regime is not the same as trust
    11:23 Crewed and uncrewed systems working together
    12:18 Open questions on human authority and override
    14:12 Deployment timing and what the record does not support
    15:26 What will decide where this goes
    17 min
  • Putting Words in People's Mouths, in Real Time
    A Microsoft patent application, published October 29, 2026, describes AI deepfake technology that could replace spoken words in a live video, redraw the speaker's lips to match, and personalize the result for each viewer's profile. It is a filing, not a product or a demonstration. Noelle walks through what the claims say, including no reported requirement for the speaker's consent or for telling viewers an edit was made, and an optional stored original recording. The episode also covers Resemble AI's Agent Detection, a tool that spots AI agents and bots browsing websites by tracking mouse movement, typing and scrolling. Resemble's benchmark results come from its own internal testing, and the company says about five to six percent of human visits are flagged as agents at the default setting. This is episode thirty-one of Reporting from the Uncanny Valley.
    16 min
  • Fifteen Hundred Agents: What SK Telecom Learned When It Cleaned Its Data
    SK Telecom, South Korea's largest mobile carrier, told employees to build their own AI agents. The count reached about seven thousand, then the company cut it to about fifteen hundred in regular use, and it says it is weighing each agent's effectiveness against its token cost. This episode of Reporting from the Uncanny Valley covers Data Desk, the tool SK Telecom says raised AI accuracy on network database questions from about forty percent to over eighty percent by defining what data fields mean for the business. It also covers the Zero Exposure Zone data security approach and the company's second AX phase: the AX Enabler task force, the AX Sandbox, AX Catalyst and Work Map. Noelle explains why enterprise AI adoption figures and a J curve are not proven financial returns.
    Timestamps
    0:00 From seven thousand agents to fifteen hundred
    0:25 Accuracy rises from about 40 to over 80 percent
    1:17 SK Telecom's AX transformation and the November 1 disclosure
    2:14 One AI agent per person and the adoption figures
    3:37 Trimming the collection and the token cost test
    4:45 Own model versus external large language models
    5:35 Why agents struggle with data field meaning
    5:57 Data Desk and business context for data fields
    6:48 The network database accuracy result and its limits
    7:36 Ellis Ward on data quality versus model quality
    8:45 Zero Exposure Zone and sensitive data
    10:02 AX Enabler task force and employee numbers for agents
    10:49 AX Sandbox and workflow redesign
    11:46 AX Catalyst and Work Map
    12:59 The J curve and the path to returns
    14:29 What the record shows and what it does not
    15:18 Closing lesson on prerequisite data work
    16 min