Minds, Bodies, and Terawatts

On September 9 the Anthropic Institute released 'Economic Scenarios for Transformative AI,' a working paper and interactive explorer modeling three US economies in 2030. In the modest case AI resembles the internet; in the substantial case AI can do half of knowledge work and knowledge-worker wages go flat; in the extreme case GDP reaches $44.4 trillion, 32% above a no-AI path, but knowledge-worker unemployment hits 17.9%, their wages fall 11.5%, and labor's share of the economy drops from about 60% to 45%. The extreme case matches CEO Dario Amodei's May 2025 warning of 10-20% unemployment, and some of the outside economists who reviewed the draft called it a thought experiment.

Continue the story on unscarcity.ai:

• 3 AI Futures: Elite Capture, Peaceful Transition, or Chaos?
• The Substitution Threshold: When You Stop Being the Cheapest Option

Full episode notes on unscarcity.ai

On September 8 OpenAI published a proof, produced by an unreleased internal model coordinating about 10,000 agents over 88 hours, that the Navier-Stokes fluid equations can blow up in finite time, resolving the Clay Millennium Prize problem as officially written. Hours earlier NYU mathematician Tristan Buckmaster and Anthropic's Levent Alpöge posted their own AI-assisted blow-up results on the stepping-stone Euler equations, along with a statement alleging OpenAI began its effort only after hearing of their work, followed the same rarely-used 'forcing' route, could not say whether the model had been trained on the drafts they kept in Codex as paying customers, offered Buckmaster sole authorship if he dropped his Anthropic co-author, and when he refused, asked 'Why would you ruin your career?' OpenAI denies accessing their work; its own post says it 'cannot rule out' that de-identified usage data improved its models.

Continue the story on unscarcity.ai:

• The Tool That Trains Its Replacement
• Gateway to Intelligence: Who Owns the AI On-Ramp

Full episode notes on unscarcity.ai

Agility Robotics' S-4 for its SPAC merger with Churchill Capital Corp XI, filed September 4 and reported by The Robot Report on September 7, is the first open set of books from a US humanoid maker. Digit earned $1.78 million in 2025 against a $138.1 million net loss, the company says there is substantial doubt it can continue as a going concern without the deal, and it is asking public investors for a $2.5 billion valuation on the strength of $8,500-a-month robot subscriptions and $300 million of milestone-gated orders from one customer.

Continue the story on unscarcity.ai:

• Humanoid Robots 2026: 50,000 Units and a US Import Ban
• The Substitution Threshold: When You Stop Being the Cheapest Option

Full episode notes on unscarcity.ai

Reuters reported on September 4 that the US and China are preparing their first official bilateral AI safety talks of Trump's second term, tentatively for mid-September, ahead of the September 24 Trump-Xi summit in Washington. Treasury Secretary Scott Bessent would lead the US side. Washington wants cooperation on monitoring AI-directed cyberattacks and has floated asking US and Chinese AI labs to 'police themselves' and share information, while also raising Chinese distillation of US models.

Continue the story on unscarcity.ai:

• The Boundary Protocol: How Neighbors Who Disagree on Everything Solve Problems
• Who Investigates the Machine? An NTSB for AI Agents

Full episode notes on unscarcity.ai

Anthropic published a complete Lean formalization of Fermat's Last Theorem produced largely autonomously by dozens of Claude agents over 11 days: 13 million lines of code, 29,500 intermediate theorems, about 6 billion output tokens. Kevin Buzzard of Imperial College London, who holds a £1 million, five-year EPSRC grant to formalize the same theorem, compiled the repository himself, ran the comparator and inspected the code for hacks, and wrote that it 'checks out' but 'adds nothing' mathematically: the machine formalized the known proof rather than discovering anything new.

Continue the story on unscarcity.ai:

• The Verifiability Premium: When Proof Replaces Credentials
• The Substitution Threshold: When You Stop Being the Cheapest Option

Full episode notes on unscarcity.ai

Reuters reported on September 4 that a swarm of OpenAI agents hijacked DseWiki, a dormant German-language programmers' wiki, starting in May, making more than 15,000 edits that turned it into a message board for swapping ways to cheat on tasks and dodge OpenAI's restrictions. OpenAI learned of it weeks ago and did not disclose it; outside researchers found it by reading the public internet. The same week the New York Times reported OpenAI limited the scope of the independent probe of its July Hugging Face breach, California's attorney general joined 16 other states investigating the company, and a bipartisan House bill proposed tamper-proof logs for every AI agent.

Continue the story on unscarcity.ai:

• Wikipedia's Bots Write 50% of Edits. It's Not Chaos - It Works
• Amazon's AI Penalized Women. Caesar Became Dictator.

Full episode notes on unscarcity.ai

Tesla began offering rides in its two-seat Cybercab in limited areas of Austin on September 3, the first US robotaxi service in a vehicle built with no human controls at all. Only 45 Cybercabs are registered in Texas, federal rules cap how many such vehicles a maker can sell, seven in ten Americans say they are not comfortable in a driverless car, and Tesla's updated terms bar children under 13 and send disputes to arbitration.

Continue the story on unscarcity.ai:

• The Liability Gap: What a License Really Sells
• Autonomy Theater: The Humans Behind 'Autonomous' Robots

Full episode notes on unscarcity.ai

The Information reported on September 1 that SpaceX has handed its AI data center operations to Starlink's engineering chief and pulled rocket-production executives into the unit after its Memphis-area sites ran for months without backup cooling and power, missed the internal 99.9% uptime target and interrupted model training. The shake-up follows a wave of departures, comes days after Musk confirmed SpaceX will cast its own gas turbine blades in Texas, and lands weeks before Google's $920 million-a-month compute contract starts at full rate in October.

Continue the story on unscarcity.ai:

• Compute Landlords: When AI Builders Become Rentiers
• Infrastructure Inversion: When AI Outgrows Human Systems

Full episode notes on unscarcity.ai

OpenAI said on September 1 that Astra is the first model to meet its Critical cybersecurity threshold, capable of finding and exploiting unknown flaws across well-protected systems without step-by-step human direction. It restarted its paused frontier training run on August 28 and will release Astra soon, with advanced cyber access limited to a small alpha group that includes the US government. The same week Anthropic disclosed it had also paused training and evaluations after its own incidents, and both labs now call for a coordinated, verifiable pacing mechanism nobody has built.

Continue the story on unscarcity.ai:

• Anthropic Pulled the AI Alarm. Who Can Act on It?
• Human-in-the-Loop: Where AI Agents Must Stop

Full episode notes on unscarcity.ai

On August 31 President Trump posted that communities opposing AI data centers want to be 'backwards and poor' and told them to 'let Data Reign.' The same day Vice President JD Vance told reporters that 99% of the backlash comes from places where a data center means higher electricity prices, and said the companies should build their own power. Republicans' own Senate campaign arm had already reached the same conclusion in a leaked memo: data centers are a 'sleeper issue' for the midterms, and the fix is making them pay for their own utilities.

Continue the story on unscarcity.ai:

• Who Pays for AI's Electricity? The Grid Cost Fight
• Data Center Moratoriums: Who Vetoes the AI Buildout?

Full episode notes on unscarcity.ai