Overview of Hard Fork — "Data Centers in Space + A.I. Policy on the Right + A Gemini History Mystery"
This New York Times Hard Fork episode covers three distinct but connected AI stories: Google’s Project Suncatcher (a proposal to build solar-powered data centers in low Earth orbit), an insider briefing on how the Trump White House approached national AI policy (with Dean Ball, former senior AI policy advisor), and a surprising experiment by historian Mark Humphries that suggests an unreleased Google Gemini model may show step‑function improvements in handwriting, numeric and tabular reasoning.
1) Project Suncatcher — data centers in space
What Google (and others) are proposing
- Google published a paper/blog describing "Project Suncatcher": a design for space‑based, highly scalable AI infrastructure.
- Concept: thin solar arrays and computing clusters in low Earth orbit (dawn–dusk orbit) that receive nearly continuous sunlight — potentially up to ~8× the productivity of terrestrial solar.
- Prototype plans: Google aims to test two prototype satellites around 2027 in partnership with Planet. Startups (e.g., StarCloud) and other firms (Axiom Space) are also active; public comments suggest interest from figures like Jeff Bezos and Eric Schmidt; possible Chinese efforts exist.
Rationale
- Terrestrial buildout faces permitting, land, water, community opposition and an energy‑grid capacity problem as AI compute demand grows exponentially.
- Space solar promises abundant, near‑constant energy, addressing one of the biggest bottlenecks for large AI deployments.
Technical and economic hurdles
- Launch and hardware costs: sending chips and infrastructure to orbit is currently many times more expensive than building equivalent data centers on Earth.
- Radiation: Google tested TPUs in proton‑beam simulations and found newer TPUs were more resilient than expected, potentially surviving a five‑year mission.
- Maintenance & repair: operators expect to rely on robotics for in‑orbit repairs.
- Latency and data transfer: LEO latency is relatively small — comparable to existing satellite constellations (Starlink); deemed manageable for many workloads.
- Environmental and political concerns: space debris, regulatory and geopolitical issues, and public opposition (a new NIMBY variant: "NOMPs—Not On My Planet").
How Google frames it
- Positioning Project Suncatcher as a long-term moonshot (5–15+ years), similar to previous multi‑decade efforts (Waymo, quantum computing). Google presents it as serious R&D rather than sci‑fi whimsy.
2) AI policy on the right — interview with Dean Ball
Who Dean Ball is
- Former senior policy advisor for AI and emerging technology in the Trump White House; led drafting of the administration’s AI action plan. Now at the Foundation for American Innovation and author of the newsletter Hyperdimensional.
Key takeaways on the administration’s stance and right‑of‑center views
- Coherent intuitions in the administration: AI is an enormous opportunity; there are familiar and novel risks; AI is central to U.S. strategic leadership.
- Right‑wing factions are diverse:
- Accelerationists/pro‑industry voices (skeptical of “doomer” rhetoric).
- Doomer/nationalist voices (e.g., Steve Bannon‑aligned figures) who emphasize existential and catastrophic risk.
- National security-focused conservatives who prioritize competition with China.
- Kid‑safety/consumer‑harms conservatives concerned with content harms and youth safety.
- Industry is not monolithic: hyperscalers (Google, Microsoft, Amazon) have nuanced positions (may favor export controls to slow foreign competition), frontier labs want chip access and infrastructure; firms are building infrastructure to create moats.
Federal vs. state regulation
- Dean Ball argues for federal standards for high‑impact models (interstate commerce, training costs, and scale) to avoid a patchwork (California currently functions as a de facto regulator).
- He supports federal action on tail risks and national security; otherwise he accepts a largely reactive posture for many harms, using liability laws and targeted statutes as harms materialize.
The "woke AI" executive order (procurement)
- The administration’s executive order targets procured models for government use, asking agencies to avoid models with engineered ideological bias — it governs procurement, not public model training.
- Ball argues the government can require disclosure of system prompts for procurement, but mandating training changes for public models would raise First Amendment problems.
Outlook
- Ball thinks incremental policy advances are possible without catastrophic triggers, but acknowledges politically contentious issues will likely fragment into topic‑specific debates (e.g., data centers, kid safety, export controls).
- He sees real incentives for labs to invest in safety (bankruptcy risk if catastrophic harms occur), and believes some tail‑risk measures should be bipartisan.
3) A Gemini history mystery — Mark Humphries’s experiment
The experiment
- Mark Humphries (history professor, Wilfrid Laurier University) uses AI to transcribe and extract metadata from handwritten historical records (fur trade, 18th–19th century).
- In Google’s AI Studio he repeatedly saw an A/B test response from a mystery model (likely an unreleased Gemini variant — possibly Gemini 3) that markedly outperformed prior models.
What stood out
- Error rate: Humphries reports a word error rate around ~1% for the mystery model — comparable to human transcribers and roughly a 50% drop from Gemini 2.5 Pro’s errors (which were around 95% accuracy in his tasks).
- Tabular and numeric reasoning: the model correctly interpreted compact, old ledger notations (e.g., “145” meaning 14 pounds 5 ounces) and converted/aggregated values within historical currency/measurement systems (pounds, shillings, pence).
- This suggests improved numeric, tabular and symbolic reasoning beyond simple token prediction — capabilities notably hard for earlier LLMs.
Implications
- If replicated, historians can trust models for richer data extraction tasks (e.g., extracting itemized purchases, quantities, bookkeeping math) and scale research that previously required large human transcription efforts.
- More broadly, improvements that let models reliably handle structured numeric reasoning and domain‑specific symbolic conversions could generalize across knowledge work (legal, financial, scientific domains).
- The finding supports a view that continued scaling/improvements can produce qualitatively new capabilities, not just incremental betterness.
Notable quotes & pithy lines from the episode
- "The sun is a really freaking good source of energy." — highlights the energy motivation for space solar.
- "From NIMBYs to NOMPs — Not On My Planet." — on potential public opposition to space data centers.
- On the procurement executive order: it's about the versions sold to the government, not public model training.
- Humphries: the model's handling of "145" → "14 lb 5 oz" felt like symbolic reasoning, not mere pattern matching.
Key takeaways / what to watch next
- Project Suncatcher is a serious R&D moonshot: expect prototypes in the next few years but wide deployment remains costly and years away; watch for Google’s 2027 test launches and StarCloud/Axiom activity.
- Space data centers could solve energy constraints for massive AI compute but raise new technical, environmental, regulatory and geopolitical questions.
- U.S. AI policy is still forming: expect fragmentation across issue domains, with federal action prioritized for tail risk, national security and procurement rules; states will continue to act on immediate harms (e.g., kid safety).
- Model capabilities continue to surprise domain experts: the Humphries experiment indicates next‑generation models may reliably perform structured, numeric and domain‑specific reasoning — a potential inflection point for knowledge work automation.
- Keep an eye on Gemini 3’s public release and independent benchmarks, plus further experiments in AI Studio.
Practical actions / recommendations
- If you’re a policy watcher: track federal procurement rules, California’s SB 53 implementations, and any congressional efforts on export controls and tail‑risk mitigation.
- If you’re in tech or infrastructure: evaluate long‑term energy and compute projections; consider supply chain, launch cost, and space‑debris/environmental risk analyses.
- If you’re a researcher or historian: pilot next‑gen models on domain test sets (with held‑out benchmarks) and validate numeric/tabular outputs against human ground truth before automating pipelines.
- If you’re a business leader using AI: plan for faster growth in compute demand and watch how model capabilities could shift knowledge‑work workflows — invest in data governance and safety processes now.
Credits: episode hosts Kevin Roose and Casey Newton; guests Dean Ball and Mark Humphries.