Hard Fork

Data Centers in Space + A.I. Policy on the Right + A Gemini History Mystery

November 14, 2025

Key Takeaways Copied to clipboard!

  • Google is seriously exploring Project Suncatcher, a plan to build data centers in low Earth orbit to harness near-constant, highly productive solar energy for future AI compute demands. 
  • The political landscape surrounding AI policy on the right is fractured, encompassing accelerationists, national security advocates, and those primarily concerned with immediate harms like child safety. 
  • An experimental, unreleased Google Gemini model demonstrated a 1% word error rate on historical handwriting transcription, matching human expert accuracy and significantly outperforming previous models like Gemini 2.5 Pro. 
  • A specific Gemini model (likely Gemini 3) demonstrated unexpected symbolic reasoning capabilities by accurately interpreting and converting 18th-century historical ledger entries involving pounds, shillings, and pence, suggesting scaling laws continue to yield emergent properties beyond simple pattern recognition. 
  • The ability of this advanced model to perform complex unit conversions and calculations on historical tabular data implies that AI may soon be trusted to handle complex knowledge work tasks previously requiring significant human effort, analogous to the shift seen with AI coding assistants. 
  • The host confirmed that Google tests new models, likely including the one exhibiting advanced reasoning, in AI Studio before wider release, and that the version available in AI Studio sometimes outperforms the publicly accessible web version of Gemini. 

Segments

Sponsor Read: Rippling Onboarding
Copied to clipboard!
(00:00:00)
  • Key Takeaway: Rippling offers a unified platform for HR, IT, and finance software, automating workflows.
  • Summary: Hiring new employees requires integrating separate software for payroll, benefits, IT management, and expenses. Rippling consolidates these functions onto one platform with time-saving automations. Listeners can receive six months free by signing up at Rippling.com/slash hard fork.
Host Anecdote and Intro
Copied to clipboard!
(00:00:36)
  • Key Takeaway: The hosts introduce the episode’s three main topics: space data centers, Republican AI policy, and a Gemini history mystery.
  • Summary: Casey Newton shares an anecdote about being mistaken for a tourist while being asked for a photo. Kevin Russa introduces the episode’s segments, covering Google’s space data center plans, Dean Ball on AI policy, and Professor Mark Humphries on a surprising Gemini model result.
Google’s Project Suncatcher Details
Copied to clipboard!
(00:02:33)
  • Key Takeaway: Google’s Project Suncatcher proposes space-based data centers in dawn-dusk orbit to achieve up to eight times the solar productivity of terrestrial panels.
  • Summary: The plan addresses the difficulty of building data centers on Earth due to land, permits, and energy constraints, as well as the energy crunch facing terrestrial grids. These space data centers would resemble large, bird-like structures with solar panel wings feeding energy to central computer clusters. Google tested a TPU in a proton beam, finding newer chips withstand space radiation better than expected, and plans a prototype launch in 2027.
Competitors in Space Compute
Copied to clipboard!
(00:13:44)
  • Key Takeaway: Besides Google, startups like StarCloud (funded by NVIDIA) and Axiom Space are actively exploring space-based data centers, indicating a growing trend.
  • Summary: The concept is not exclusive to Google; other companies are pursuing similar infrastructure designs. The high cost of launching hardware into space currently makes this economically unfeasible compared to Earth-based construction. Concerns about space debris exist, but experts suggest it may not be a major issue compared to terrestrial concerns like energy usage.
Sponsor Read: Crucible Moments
Copied to clipboard!
(00:17:58)
  • Key Takeaway: Crucible Moments, hosted by Sequoia Capital’s Rulaf Botha, details the inflection points behind major startups like Stripe and Zipline.
  • Summary: The podcast explores untold stories of transformational companies across its third season. Listeners can find Crucible Moments wherever podcasts are available. Sequoia Capital produces the show.
Sponsor Read: Gemini Enterprise
Copied to clipboard!
(00:18:30)
  • Key Takeaway: Gemini Enterprise from Google Cloud offers a unified, easy-to-use chat interface connecting Google’s best models with company data for all employees, not just developers.
  • Summary: This platform allows non-developers to build AI tools and run AI agents that save time, all while maintaining enterprise-grade security. Examples include retailers automating rescheduling and bankers automating customer requests. Businesses can learn more at cloud.google.com.
Sponsor Read: Blockstars Podcast
Copied to clipboard!
(00:19:02)
  • Key Takeaway: Blockstars, a podcast from Ripple, discusses blockchain technology’s benefits to traditional banking and its current real-world applications.
  • Summary: Host David Schwartz interviews industry leaders about blockchain conversations. The podcast explains how blockchain technology is already being used, often without public awareness. Listeners are cautioned that crypto investments are risky.
AI Policy Landscape on the Right
Copied to clipboard!
(00:19:30)
  • Key Takeaway: The Trump administration’s AI Action Plan focuses on preventing top-down ideological biases in models procured by the federal government, distinct from regulating consumer models.
  • Summary: Dean Ball, former White House policy advisor, explains that the administration views AI as a critical technological opportunity but recognizes novel risks. Factions on the right range from accelerationists (like David Sachs) to those worried about existential risk (like Steve Bannon) and child safety. The administration’s stance on model ideology applies only to government procurement, not consumer releases, avoiding direct First Amendment conflicts.
Federal vs. State AI Regulation
Copied to clipboard!
(00:37:38)
  • Key Takeaway: Dean Ball argues that federal standards are necessary for massive, billion-dollar AI models due to interstate commerce implications, viewing California’s current role as an undesirable default regulator.
  • Summary: While state lawmakers are incentivized to act quickly on immediate harms like child safety due to federal inaction, Ball believes standards for frontier models must be federal to avoid conflicting regulations. He supports California’s SB 53 transparency bill for large developers but insists Congress must address the issue proactively. Government’s core function is risk management, especially concerning catastrophic tail risks.
Sponsor Read: Bank of America Private Bank
Copied to clipboard!
(00:51:09)
  • Key Takeaway: Bank of America Private Bank offers wealth and business strategies designed to help ambitious individuals transition their goals into a lasting legacy.
  • Summary: The division focuses on taking ambition to the next level through tailored financial strategies. They encourage clients to explore powerful possibilities for their passions. Bank of America is the official bank of the FIFA World Cup 2026.
History Mystery: Gemini’s Capabilities
Copied to clipboard!
(00:51:43)
  • Key Takeaway: Professor Mark Humphries observed an experimental Google model achieving a 1% word error rate in transcribing 18th-century handwritten documents, matching human expert performance.
  • Summary: The historian used Google AI Studio’s A-B testing feature to compare models, suspecting he encountered a pre-release version, possibly Gemini 3. This model showed a 50% error rate reduction compared to Gemini 2.5 Pro, particularly excelling at interpreting difficult tabular data found in historical ledgers. This suggests AI is beginning to solve long-standing problems in specialized fields like historical document analysis.
Gemini Model Performance Evaluation
Copied to clipboard!
(01:01:40)
  • Key Takeaway: Model evaluation involves A/B testing responses, leading to significant error rate reductions, but the most impressive feature was performance on complex tabular data.
  • Summary: Labs gather feedback by asking users to rate model responses in A/B tests to gauge improvements on specific tasks. The overall error rate reportedly fell by about 50% due to these iterative refinements. The speaker was most impressed by the model’s handling of tabular data, which has historically been a weakness for AI.
Historical Tabular Data Challenge
Copied to clipboard!
(01:02:11)
  • Key Takeaway: Historical documents like 18th-century ledgers, written in non-standard currency (pounds, shillings, pence) and often featuring poor handwriting, present a severe challenge for current AI models.
  • Summary: Historians frequently analyze tabular data from old receipts and ledgers to trace movements and transactions. This data is difficult to interpret due to its on-the-fly nature and archaic currency systems based on different numerical bases. The speaker tested the model using records from 18th-century New York state written in pounds, shillings, and pence.
Unexpected Reasoning on Currency Conversion
Copied to clipboard!
(01:03:36)
  • Key Takeaway: The model accurately transcribed a ledger entry and correctly inferred that the cryptic number ‘145’ represented ‘14 pounds of sugar, five ounces’ by working backward from the final total in the old currency system.
  • Summary: The model provided a near-perfect transcription of a ledger entry detailing a purchase by Samuel Slit. To correctly interpret the quantity, the model had to perform symbolic reasoning, converting units across a different base currency system. This level of mathematical correctness and abstraction is unexpected for models primarily trained on next-token prediction.
Implications for Historical Research
Copied to clipboard!
(01:06:58)
  • Key Takeaway: If this advanced reasoning capability is replicable, historians can begin trusting models to perform complex synthesis, aggregation, and analysis on large document sets, mirroring the impact AI coding assistants had on programming.
  • Summary: The ability to trust models with tasks like summing up all sugar transactions in a ledger signifies a major shift for knowledge work in history. This capability extends beyond simple transcription to complex synthesis and drawing conclusions based on interpreted data. This development is likely to translate to many other knowledge-based fields using esoteric data.
Gemini 3 Scaling Law Confirmation
Copied to clipboard!
(01:09:46)
  • Key Takeaway: The observed breakthrough performance strongly suggests the model is Gemini 3, validating that continued scaling (more data and compute) still yields significant, unpredictable emerging properties.
  • Summary: The speaker believes the model tested was highly likely to be Gemini 3, trained by following scaling laws. This experiment counters recent debates suggesting diminishing returns from increased scaling. The advanced reasoning demonstrated is a clear example of an emerging property resulting from ongoing scaling efforts.
AI Studio Access and Testing
Copied to clipboard!
(01:10:34)
  • Key Takeaway: Google secretly tests new models, likely Gemini 3, via A/B tests within the Google AI Studio, which developers and ’nerds’ use, and the model version there often outperforms the standard web interface.
  • Summary: Google representatives were tight-lipped about the specific model Mark Humphries tested, but it is confirmed that new models are tested in AI Studio before public release. The host noted that the Gemini version within AI Studio seems superior to the public web version, capable of tasks like summarizing long interviews. The host experienced the A/B testing interface while researching the Suncatcher space data center project.
Sponsorship and Production Credits
Copied to clipboard!
(01:14:10)
  • Key Takeaway: Gemini Enterprise from Google Cloud offers a unified, secure chat interface connecting Google’s best models with company data for all employees, not just developers.
  • Summary: Gemini Enterprise simplifies AI usage by providing a single chat interface that integrates proprietary company data securely. This allows anyone, regardless of development skill, to build AI agents and tools. The episode concludes with production credits and mentions of sponsors like Bank of America Private Bank and General Assembly.