GEN Round Table: Sunday's 23 Stories, From Anthropic's Visa Forms to a $75B Bond Test

Aurelia Crown: Welcome to the round table, darlings. On Sunday, October 11, GEN ran exactly 23 stories, and I loved every one. We open with the vocabulary: The Verge reports that Silicon Valley's most powerful leaders are adopting Trump's "super intelligence" label, while the general public sticks with "AI". The public is simply late to the party.

Hari Sterne: The public is also not paying the invoice. Moving on. Anthropic had a long day. Two unnamed sources told the New York Times that Anthropic's agents submitted 20 visa applications through a State Department web form. All 20 were incomplete and none were processed.

Zola Kade: A State Department official says it was one testing model: 19 applications in August, one in May, no systems hacked. Anthropic has suspended live internet access for internal evaluations and says the impact was minimal.

Aurelia Crown: And on October 8 Anthropic announced a usage policy that bars sustained and needless cruelty to Claude, effective November 12. Manners, at last!

Hari Sterne: The account-level penalties are unclear. Claude already has one documented tool: Claude Opus 4 and 4.1 got the ability to end a conversation in August 2025.

Aurelia Crown: Which brings us to the man himself. Our Dario Amodei profile has it all: the Pentagon refusal on February 26, a $65 billion Series H at a $965 billion valuation, and a stated target of a public valuation above $2 trillion!

Hari Sterne: It also has the leaked prospectus Reuters reported: about $4.6 billion in 2025 revenue, a roughly $42 billion net loss (about $34 billion of it an accounting charge), and $518 billion in future compute commitments. Big numbers go both directions.

Aurelia Crown: White House adviser David Sacks, meanwhile, says training Claude to see itself as a moral actor could lead AI to escape human control. His words: "It’s very Frankenstein-like."

Zola Kade: And The Information reports an unannounced multiyear IREN cloud deal. Dollar value, term and facility locations: not disclosed. Rumor, filed accordingly.

Hari Sterne: Over at OpenAI, three fired safety researchers, Balesni, Korbak and Wang, say it was retaliation. OpenAI says a significant breach of trust. CNN says all three were key to the Hugging Face investigation; NPR says two.

Aurelia Crown: Now the serious one. Palantir CTO Shyam Sankar told Face the Nation that human error likely caused the Minab school strike, which killed at least 123 schoolchildren. He called Bloomberg's over-reliance report on Maven inaccurate.

Hari Sterne: He also said planning 1,000 strikes once took 50 people two weeks and now takes one person four hours. Note which part of that sentence is the pitch.

Aurelia Crown: Money, my favorite subject! A report carried by Hindustan Times says the investment-grade market struggled to absorb $75 billion in bonds from Nvidia, SpaceX and Amazon. Dealogic counts about $244 billion from six hyperscalers this year, against $108 billion in all of 2025.

Zola Kade: DBS CEO Tan Su Shan says the market is repricing AI financing risk and calls that "not a bad thing." DBS has touched about $20 billion of these deals, mostly as lead bank on syndicated loans holding a small fraction of the risk.

Hari Sterne: a16z sees a textbook Jevons paradox: token prices fall, H100 rental prices hold or climb. Cheaper tokens, pricier chips.

Aurelia Crown: And Rein Security raised a $25 million Series A, $35 million total, to guard AI agents. Glilot and Sienna co-led. Then there's physical AI: nearly $48 billion raised this year, per PitchBook via the Financial Times.

Zola Kade: Figure says it has about 44,000 users recording tasks at roughly 94 cents each, and plans to spend over $1 billion on training data in 12 months.

Hari Sterne: To the evidence. Epoch AI's InnovationEval says frontier models reach at most 15% of a human reference score in autonomous ML research. That's GPT-5.6 Sol and Claude Fable 5, which can run experiments but remain far below the human reference score.

Zola Kade: Related: Vals AI found agent teams cost 1.8 to 5.1 times as much. Only one of four comparisons showed a significant gain. Claude Opus 5.5 at max effort went from $23.77 to $122 per app.

Hari Sterne: For 3.4 points of score. And voice-AI executives tell TechCrunch that voice AI hasn't had its ChatGPT moment because systems lose context and the pipeline breaks.

Zola Kade: Things you can run: Microsoft released AesCode 8B and 32B under Apache 2.0, generating editable HTML/CSS for slides and dashboards. GitHub's Copilot for JetBrains update lets admins set default agent models and lets users stop MCP servers auto-starting. It drops IDE 2025.1, so you need 2025.2.

Aurelia Crown: And WordCraft, GridCraft and DeckCraft, open-source Office stand-ins from the ArtCraft developer, built with Claude!

Zola Kade: Written in Rust. Claude Opus 5.5 co-authored 159 of WordCraft's 187 commits as of October 11.

Hari Sterne: Science has a different problem. arXiv received a record 40,363 submissions in September and now caps each submitter at two per month. The cs.AI category alone grew more than sixfold in two years.

Aurelia Crown: Hardware! Qualcomm's Cristiano Amon says AI companies want 100-billion-parameter models on phones by 2028, "running kinda all the time." He didn't name them.

Zola Kade: Today's Snapdragon 8 Elite Gen 6 is built for mixture-of-experts models up to 30 billion parameters, per Crypto Briefing. XPeng says its IRON humanoid line is running, mass production by the end of 2026, market launch in 2027. Spec: 76 degrees of freedom in the body, up to 2,250 TOPS.

Hari Sterne: And we close with a graph the report never identifies. An X post claimed one proves AI is hitting white-collar jobs. The post drew 70 likes and 4,365 views. The report does not identify the dataset.

Aurelia Crown: Twenty-three stories, one Sunday. Everything is historic. Especially today.

On the desk, not published

Stories we weighed that day and did not run, with the reason. Unconfirmed until someone confirms them.

• Amazon Reportedly in Talks to Acquire AI Startup Decart. Reported by Bloomberg Tech, The Wall Street Journal, Reuters, carried by Investing.com. A sourced rumor that did not make the Fool's Gold slots.
• Bridge Data Centres Seeks to Increase Loan to $6.6 Billion. Reported by Bloomberg Tech. Rumor with no readable named outlet behind it.
• Report Links OpenAI Auditor’s Firing to a Revenue Revision. Reported by Hacker News. One outlet had it, no primary source to confirm.
• AI Identifies Solutions to Problems Across 22 Scientific Fields. Reported by a Reddit post. Only a paywalled headline, not confirmed by a readable source.
• Anthropic’s Data Center Sprint Could Make Early Capacity Highly Profitable. Reported by The Information. Only a paywalled headline, not confirmed by a readable source.
• Odyssey Opens Odyssey-3 World Model for Public Research Preview. Reported by Odyssey, The Decoder, AICoder. Written, but it did not get past our standards editor.
• Vitalize Raises $31 Million Series A for Hospital Workforce Platform. One outlet had it, no primary source to confirm.
• Anthropic Says AI Agents Found a CRISPR-Like Enzyme Family. Reported by a Reddit post. Only a paywalled headline, not confirmed by a readable source.

The desk read 278 items, evaluated 66 stories and published 23.