<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xmlns:thr="http://purl.org/syndication/thread/1.0">
  <title>Gregor J. Rothfuss</title>
  <subtitle>Thoughts, notes, and curated knowledge across 27 years.</subtitle>
  <link rel="hub" href="https://pubsubhubbub.appspot.com/"/>
  <link rel="hub" href="https://websub.feedly.com/"/>
  <link href="https://gregor.fyi/atom.xml" rel="self" type="application/atom+xml"/>
  <link href="https://gregor.fyi/" rel="alternate" type="text/html"/>
  <id>https://gregor.fyi/</id>
  <updated>2026-11-12T14:21:39Z</updated>
  <author><name>Gregor J. Rothfuss</name></author>
  <entry>
    <title>Weight-Space Programming: Splicing code layers into training via phantom gradients</title>
    <link href="https://gregor.fyi/2026/11/04/weight-space-programming/#2026-11-12" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/11/04/weight-space-programming/#2026-11-12</id>
    <published>2026-11-12T12:00:00Z</published>
    <updated>2026-11-12T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2026/11/04/weight-space-programming/" href="https://gregor.fyi/2026/11/04/weight-space-programming/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2026/11/04/weight-space-programming/">Weight-Space Programming</a>:</em></p>
<blockquote>Ordinary weight gradients average across a batch, hiding token-level failures. Computing ∂L/∂h_ℓ with respect to every token's hidden state at every layer—the <a href="https://web.archive.org/web/20261106155203/https://doubleai.s.gy/llm-redesign">phantom gradient</a>—maps where representation demand persists across checkpoints. By centering phantom vectors per target token, clustering failing tokens with spherical k-means, and splicing Python code layers into the residual stream wherever they beat a linear baseline, a coding agent inside training permanently lowers validation loss on a 12-layer, 286M-parameter nanochat model, raises CLRS-Text algorithmic reasoning accuracy on unseen intermediate lengths from 55% to 84%, and lets DeepSeek-V4.1-Flash build a linear probe across blocks 24–39 that rejects 80% of its own wrong answers (at a 6.7% false-rejection rate) to lift AIME and HMMT 2026 accuracy from 64% to 71%.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Datacenter Baseload Energy: Retired airliner jet cores repurposed for onsite power</title>
    <link href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/#2026-11-12" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/11/04/datacenter-baseload-energy/#2026-11-12</id>
    <published>2026-11-12T12:00:00Z</published>
    <updated>2026-11-12T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/" href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/">Datacenter Baseload Energy</a>:</em></p>
<blockquote>Because utility interconnections and heavy-frame industrial gas turbines are backlogged for years, datacenter operators are buying decommissioned Boeing 737, 747, and Airbus A320 engines and competing directly with airlines for overhaul-ready cores. FTAI Aviation rebuilds <a href="https://web.archive.org/web/20261108132823/https://x.com/Fahadnaimb/status/2119421779978625144">CFM56 cores</a> into 23 MW continuous (25 MW peak) stationary generators—with over 100 units contracted for 2027–2028 and a $1.465 billion mobile-turbine order from a cloud provider via its Jereh Group joint venture—while xAI deployed 8 converted GE CF6 cores at its Mississippi Colossus 2 site.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Construction remains terrible: While wind, flood, and seismic provisions in US building codes deliver 6x–12x benefit-cost ratios over 1990s baselines (quality gains omitted from the Census Bureau&#x27;s residential price index), code-driven construction cost inflation is heavily concentrated in a few low-benefit rules</title>
    <link href="https://gregor.fyi/2021/06/28/construction-terrible/#2026-11-12" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2021/06/28/construction-terrible/#2026-11-12</id>
    <published>2026-11-12T12:00:00Z</published>
    <updated>2026-11-12T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2021/06/28/construction-terrible/" href="https://gregor.fyi/2021/06/28/construction-terrible/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2021/06/28/construction-terrible/">Construction remains terrible</a>:</em></p>
<blockquote>Revisions to the International Residential Code (IRC) since 2006 have added ~$40,000 (~0.5% per year) to the $440,000 average construction cost of a reference single-family home, with nearly 60% of the increase across 89 cost-raising changes between 2009 and 2024 coming from just 10 provisions. Across 140 revisions to the <a href="https://web.archive.org/web/20261106130135/https://www.construction-physics.com/p/how-do-building-codes-affect-the">International Building Code</a> from 2012 to 2021, only 16 provisions raised 3-story apartment construction costs—adding 9.9% cumulatively (~$400,000, or ~0.8% per year)—with $180,000 (45%) stemming from a single 2018 mandate for off-hours fire watches during construction, alongside the 6–13% total building cost penalty of requiring 2 stairwells connected by a corridor.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Supersonic Jets: Parametric code generation for airframes and turbines</title>
    <link href="https://gregor.fyi/2018/07/23/supersonic-jets/#2026-11-12" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2018/07/23/supersonic-jets/#2026-11-12</id>
    <published>2026-11-12T12:00:00Z</published>
    <updated>2026-11-12T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2018/07/23/supersonic-jets/" href="https://gregor.fyi/2018/07/23/supersonic-jets/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2018/07/23/supersonic-jets/">Supersonic Jets</a>:</em></p>
<blockquote>Aerospace engineering still relies on manual, unversioned spreadsheets to size wing spars from structural loads or environmental systems from thermal loads. Writing the <a href="https://web.archive.org/web/20261106020343/https://x.com/bscholl/status/2118524700796305669">parametric design algorithm</a> instead of static part geometry lets Boom Supersonic's turbomachinery solver run in seconds for 1 engineer what previously took 100 engineers 6 months, while a structural generator synthesizes 65% of the airframe—enabling 50 engineers to build a supersonic demonstrator jet and 150 to build a full-scale supersonic turbofan.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Better Science: Statistical power in development economics RCTs</title>
    <link href="https://gregor.fyi/2003/05/14/better-science/#2026-11-12" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2003/05/14/better-science/#2026-11-12</id>
    <published>2026-11-12T12:00:00Z</published>
    <updated>2026-11-12T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2003/05/14/better-science/" href="https://gregor.fyi/2003/05/14/better-science/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2003/05/14/better-science/">Better Science</a>:</em></p>
<blockquote>We study the statistical credibility of <a href="https://web.archive.org/web/20261110200526/https://marginalrevolution.com/marginalrevolution/2026/11/are-rcts-underpowered.html">randomized controlled trials in development economics</a> and the optimal allocation of resources to replication. Using 745 principal-hypothesis tests from 190 articles in leading economics journals, we find that actual t-statistics peak well below the theoretical benchmark of 2.8 expected under 80% power, though above values associated with p-hacking. Estimating an empirical Bayes mixture model that accounts for selective publication, we find that ~30% of published rejections are false discoveries, with average directional power at only 40%. At conventional significance thresholds, the posterior probability of a false discovery reaches 66%, whereas concentrating replications in an intermediate interval of original test statistics reduces aggregate false discovery rates to 1.8–7.8%.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Refusing National Statistics</title>
    <link href="https://gregor.fyi/2026/11/12/refusing-national-statistics/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/11/12/refusing-national-statistics/</id>
    <published>2026-11-12T06:30:00Z</published>
    <updated>2026-11-12T06:30:00Z</updated>
    <content type="html"><![CDATA[As Hong Kong's Financial Secretary from 1961 to 1971, Sir John Cowperthwaite sent Whitehall statisticians home rather than let the colony collect macroeconomic data:<blockquote>Asked by Milton Friedman in 1963 why he refused to gather national accounts, <a href="https://web.archive.org/web/20261105210205/https://x.com/Handre/status/2118448791560884507">Cowperthwaite</a> answered, "If I let them compute those statistics, they'll want to use them for planning"—turning descriptive metrics into tariffs, subsidies, and industrial gates. Over his decade in office without national statistics or industrial policy, Hong Kong's real wages rose 50%, acute poverty fell from 50% to 15%, and per-capita income climbed from 25% of the British level in 1961 to surpassing the United Kingdom by the early 1990s.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Submerged Swamp Roads</title>
    <link href="https://gregor.fyi/2026/11/12/submerged-swamp-roads/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/11/12/submerged-swamp-roads/</id>
    <published>2026-11-12T06:20:00Z</published>
    <updated>2026-11-12T06:20:00Z</updated>
    <content type="html"><![CDATA[During the 13th and 14th century Northern Crusades, pagan Lithuanian and Samogitian defenders built secret underwater stone causeways across peat bogs to move cavalry and carts between fortresses:<blockquote>Laid below the swamp waterline so they were invisible from the surface, these <a href="https://web.archive.org/web/20261111143301/https://x.com/Arannatar/status/2120525209195135090">kūlgrindos</a> deliberately twisted through low-lying mire rather than taking the shortest crossing—as archaeologist Ludwik Krzywicki found in 1903 when surveying 3 surviving stone stretches (250 m, 180 m, and 150 m) linking the fortresses of Medvėgalis, Paršpilis, and Šiuraičiai at Sietuva—so pursuing Crusaders unfamiliar with the turns would step off the submerged stones into deep bog mud.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Riemann&#x27;s Shock Equations</title>
    <link href="https://gregor.fyi/2026/11/12/riemanns-shock-equations/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/11/12/riemanns-shock-equations/</id>
    <published>2026-11-12T06:10:00Z</published>
    <updated>2026-11-12T06:10:00Z</updated>
    <content type="html"><![CDATA[Bernhard Riemann's published 1860 paper on compressible gas dynamics failed to conserve energy across a shock front because it assumed reversible adiabatic compression instead of entropy-producing jump conditions:<blockquote>Auditing 12,000 pages of handwritten Nachlass manuscripts by Gauss, Riemann, Hilbert, and Grothendieck at Göttingen with Opus 5.5 agents uncovered 2 <a href="https://web.archive.org/web/20261108150901/https://x.com/Salim_AIi/status/2119447102958752038">unpublished Riemann sheets</a> that derive the exact Rankine-Hugoniot shock conditions in his own notation 10–20 years before William Rankine (1870) and Pierre-Henri Hugoniot (1887), alongside shock-wave relations conventionally dated to 1906–1937.</blockquote>]]></content>
  </entry>
  <entry>
    <title>e-means Clustering</title>
    <link href="https://gregor.fyi/2026/11/12/e-means-clustering/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/11/12/e-means-clustering/</id>
    <published>2026-11-12T06:00:00Z</published>
    <updated>2026-11-12T06:00:00Z</updated>
    <content type="html"><![CDATA[Lloyd's k-means algorithm requires a full pass over a dataset for every centroid update, while standard mini-batch k-means converges early to higher quantization error:<blockquote>Adapting vector-quantized autoencoder moving averages to static MSE minimization, Meta's <a href="https://web.archive.org/web/20261111122800/https://www.alphaxiv.org/abs/2611.09501">e-means</a> updates running cluster counts and vector sums on a cosine decay schedule down to 10% of peak step size, splits donor clusters every 5 batches whenever a cluster's count drops below 10% of the mean, and sets batch and step sizes directly from scaling laws with 0 per-dataset tuning. On 1 billion 1,024-dimensional DINO embeddings (1,048,576 centroids), e-means reaches 25-epoch Flash-KMeans test error in 1.6 epochs for a 45.5x wall-clock speedup; on the full 10-billion-vector dataset with 16 million centroids, exact e-means finishes in 2.5 epochs (779 GPU-hours vs. 6,784 for 25-epoch Lloyd) with 0.014% lower test MSE, while an inverted-file approximate variant reaches 0.033% lower MSE in 23.2 GPU-hours on 1 machine.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Put Silk in it: Phosphate-triggered arginine-tyrosine condensates guide disordered spidroin proteins into beta-sheet fibers</title>
    <link href="https://gregor.fyi/2021/06/28/put-silk-in-it/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2021/06/28/put-silk-in-it/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2021/06/28/put-silk-in-it/" href="https://gregor.fyi/2021/06/28/put-silk-in-it/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2021/06/28/put-silk-in-it/">Put Silk in it</a>:</em></p>
<blockquote>Using isotope-enriched amino acid tracers and NMR spectroscopy on native black widow <a href="https://web.archive.org/web/20261028120000/https://www.quantamagazine.org/how-spiders-spin-silk-dope-into-natures-toughest-material-20261028">silk dope</a>, researchers mapped how liquid spidroin proteins transition into solid fibers tougher than Kevlar and stronger than steel. Adding phosphate triggers weak sticker-like interactions between arginine and tyrosine side groups in the spidroins' intrinsically disordered regions, displacing water and organizing the tangled proteins into semiliquid droplets (condensates) that align right at the bends of the final fiber's load-bearing beta sheets.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Inferring lost cities: Applying a dynamic gravity model to where and when 500,000 ancient coins were minted versus buried reconstructs bilateral trade and real consumption every 20 years from 325 to 950</title>
    <link href="https://gregor.fyi/2017/11/12/inferring-lost-cities/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2017/11/12/inferring-lost-cities/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2017/11/12/inferring-lost-cities/" href="https://gregor.fyi/2017/11/12/inferring-lost-cities/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2017/11/12/inferring-lost-cities/">Inferring lost cities</a>:</em></p>
<blockquote><a href="https://web.archive.org/web/20261103152344/https://www.econometricsociety.org/publications/econometrica/forthcoming-papers/0000/00/00/Trade-and-the-End-of-Antiquity/file/23696-3.pdf">Bilateral coin flows</a> confirm Pirenne’s thesis that north-south Mediterranean trade collapsed after the 7th-century Arab conquests due to high border friction between Islamic and Christian polities, even as south-south trade remained vibrant. Real consumption in the Western Roman core (Italy, southern France, Egypt) began falling in the 4th century alongside institutional collapse, while the Byzantine heartland—initially among the highest—gradually collapsed starting in the 7th century as it lost trade with newly conquered Arab territories. Conversely, consumption peaked in Umayyad Syria in the early 8th century, Abbasid Iraq in the 8th century, and Umayyad Córdoba in the late 8th century, while surging across Frankish northwestern Europe in the 7th and 8th centuries—making Atlantic Iberia and northwestern Europe the wealthiest regions of the western world more than 700 years before colonial expansion.</blockquote>]]></content>
  </entry>
  <entry>
    <title>NYC vs London construction: Procurement fragmentation and station overengineering stem from the same interagency coordination failure in both London and New York</title>
    <link href="https://gregor.fyi/2013/03/11/nyc-vs-london-construction/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2013/03/11/nyc-vs-london-construction/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2013/03/11/nyc-vs-london-construction/" href="https://gregor.fyi/2013/03/11/nyc-vs-london-construction/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2013/03/11/nyc-vs-london-construction/">NYC vs London construction</a>:</em></p>
<blockquote>In New York, the transit construction cost premium breaks down into 3 multiplicative factors: labor overstaffing (~1.5x), project delivery (~2x), and overdesign of stations and systems (~3x). Historical cost data from London shows that <a href="https://web.archive.org/web/20261031120000/https://pedestrianobservations.com/2026/10/31/project-delivery-and-overengineering-are-the-same-problem">procurement and design premiums</a> are the same underlying problem. Central Crossrail stations excavated 150,000–200,000 m³ each, compared with 85,000 m³ for Odenplan on Stockholm’s Citybanan serving trains only 10% shorter, while both Second Avenue Subway and Crossrail hit a 1:1 systems-to-civils cost ratio versus 1:2.5 in France, Italy, and Sweden—driven by fragmented agencies piling uncoordinated spatial and systems requirements onto unsupervised design-build contracts.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Arrow&#x27;s Theorem: Condorcet voting cycles have the non-orientable topology of a Klein bottle</title>
    <link href="https://gregor.fyi/2010/10/25/arrows-theorem/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2010/10/25/arrows-theorem/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2010/10/25/arrows-theorem/" href="https://gregor.fyi/2010/10/25/arrows-theorem/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2010/10/25/arrows-theorem/">Arrow&#x27;s Theorem</a>:</em></p>
<blockquote>Representing each of the 6 possible strict rankings of 3 election candidates as a triangle whose vertices encode pairwise preferences and gluing those 6 triangles along shared pairwise edges forms a cylinder whose 2 open rims represent opposite <a href="https://web.archive.org/web/20261026120000/https://www.quantamagazine.org/how-voting-is-like-a-klein-bottle-20261026">Condorcet voting cycles</a> (A &gt; B &gt; C &gt; A and its reverse). Because both cyclic rims denote the same failure of transitivity, identifying them without pairing incompatible preference edges requires gluing the cylinder's ends with an orientation-reversing twist—forming a non-orientable Klein bottle. Just as every local patch of a Klein bottle has 2 distinct sides that cannot be extended into a consistent global inside and outside, every individual voter holds a transitive ranking that cannot be aggregated into a globally consistent collective order.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Fuzzing: Coverage-guided fuzzing and LLM commit-auditing agents have complementary blind spots on Chrome&#x27;s V8 JavaScript engine</title>
    <link href="https://gregor.fyi/2007/08/03/fuzzing/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2007/08/03/fuzzing/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2007/08/03/fuzzing/" href="https://gregor.fyi/2007/08/03/fuzzing/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2007/08/03/fuzzing/">Fuzzing</a>:</em></p>
<blockquote>Over 6 months of production runs on V8, <a href="https://web.archive.org/web/20261030120000/https://projectzero.google/2026/10/fuzzing-v-AI.html">BigSleep and Fuzzilli</a> discovered a comparable number of actionable security vulnerabilities (45 vs. 37), with only a 25% production overlap (and ~60% theoretical overlap when BigSleep was tasked with rediscovering 112 historic Fuzzilli bugs from their introducing commits). Rather than substituting for fuzzers on hard targets, bug difficulty correlated across both systems, with complementary blind spots: LLM agents excelled on vulnerabilities requiring complex logical state progressions, specific magic values, or newly added syntax unimplemented in fuzzer mutators, whereas coverage-guided fuzzers dominated in V8’s JIT compilers (such as Maglev SSA phi-node representation conflicts) where 2 optimizations are each locally sound in isolation but cause memory corruption when composed at runtime.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Tissue engineering: Biodegradable Zn-Mg-Sr gyroid scaffolds match cancellous bone stiffness to eliminate titanium stress shielding</title>
    <link href="https://gregor.fyi/2007/02/27/tissue-engineering/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2007/02/27/tissue-engineering/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2007/02/27/tissue-engineering/" href="https://gregor.fyi/2007/02/27/tissue-engineering/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2007/02/27/tissue-engineering/">Tissue engineering</a>:</em></p>
<blockquote>Because titanium implants (~110 GPa) are far stiffer than spongy cancellous bone (0.02–3.7 GPa), the metal carries the mechanical load and causes the surrounding bone to resorb via stress shielding. Selective laser melted <a href="https://web.archive.org/web/20261026212537/https://x.com/Saint_n0mad/status/2114830836629491890">Zn-1Mg-0.1Sr TPMS gyroid scaffolds</a> (70% porous, 550 μm pores) tune the elastic modulus down to 2.3 GPa while precipitating Mg2Zn11 and SrZn13 intermetallics that more than double compressive strength over pure zinc to 33.7 MPa. The lattice degrades at 0.36 mm/year with 0 hydrogen gas pockets, converting its dissolving surface into calcium-phosphate and zinc-phosphate minerals while releasing Zn2+ and Sr2+ ions that accelerate osteoblast growth until native bone replaces the implant.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Fusion: D-T fusion&#x27;s 14 MeV neutron flux enables bulk nuclear transmutation beyond the energy limits of cyclotrons and thermal fission reactors</title>
    <link href="https://gregor.fyi/2007/02/09/fusion/#2026-11-04" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2007/02/09/fusion/#2026-11-04</id>
    <published>2026-11-04T12:00:00Z</published>
    <updated>2026-11-04T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2007/02/09/fusion/" href="https://gregor.fyi/2007/02/09/fusion/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2007/02/09/fusion/">Fusion</a>:</em></p>
<blockquote>Medical cyclotrons accelerate positively charged protons or alpha particles against electrostatic nuclear repulsion—so when a 75 kW GE PETtrace 800 bombards oxygen-18 to make fluorine-18, only 1 in 800 protons enters the nucleus, requiring 0.7 TWh per gram (~1.5 hours of total US electricity output). Fission reactors avoid charge repulsion by using neutral neutrons, but average fission neutrons emerge at 1.5–2 MeV (and are moderated down to 0.025 eV), with only ~0.1% exceeding 10 MeV. Because deuterium-tritium fusion <a href="https://web.archive.org/web/20261029120000/https://www.construction-physics.com/p/a-new-kind-of-chemistry">reliably emits 14+ MeV neutrons</a> while co-generating electricity, fusion blankets can drive high-threshold transmutation reactions at scale—from scarce alpha-emitting cancer isotopes (actinium-225, terbium-149) and mercury-to-gold transmutation to isotopically pure silicon-28 for quantum computing and high-conductivity diamond.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Weight-Space Programming</title>
    <link href="https://gregor.fyi/2026/11/04/weight-space-programming/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/11/04/weight-space-programming/</id>
    <published>2026-11-04T07:30:00Z</published>
    <updated>2026-11-04T07:30:00Z</updated>
    <content type="html"><![CDATA[Frontier LLMs can directly inspect and edit raw neural network weights in Python without gradients or training loops:<blockquote>Asked to improve Qwen3.5-0.8B at math with no gradient descent or training, GPT-6 Astra uses a Python kernel over the model's 0.8 billion parameters and a budget of 50 forward passes to trace why 85 of the base model's first 96 practice answers run into repetitive thinking loops at the 32,768-token limit. Astra isolates the looping behavior to a single direction written by the MLPs of layers 15 and 19 and subtracts that direction from 64 output columns in each—<a href="https://web.archive.org/web/20261102175148/https://rikiyat.github.io/blog/weight-space-programming/">natively editing 0.016% of the weights</a> (131,072 parameters). That edit cuts token-limit looping on MATH-500 from 93.0% to 33.1%, raising MATH-500 accuracy from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and out-of-distribution instruction following (IFEval) from 21.0% to 62.3%.</blockquote>
<a id="2026-11-12" href="#2026-11-12">2026-11-12</a>: Splicing code layers into training via phantom gradients<blockquote>Ordinary weight gradients average across a batch, hiding token-level failures. Computing ∂L/∂h_ℓ with respect to every token's hidden state at every layer—the <a href="https://web.archive.org/web/20261106155203/https://doubleai.s.gy/llm-redesign">phantom gradient</a>—maps where representation demand persists across checkpoints. By centering phantom vectors per target token, clustering failing tokens with spherical k-means, and splicing Python code layers into the residual stream wherever they beat a linear baseline, a coding agent inside training permanently lowers validation loss on a 12-layer, 286M-parameter nanochat model, raises CLRS-Text algorithmic reasoning accuracy on unseen intermediate lengths from 55% to 84%, and lets DeepSeek-V4.1-Flash build a linear probe across blocks 24–39 that rejects 80% of its own wrong answers (at a 6.7% false-rejection rate) to lift AIME and HMMT 2026 accuracy from 64% to 71%.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Multimodal Neurons: Epistemic honesty and cooperative alignment are entangled in the same attention heads rather than orthogonal</title>
    <link href="https://gregor.fyi/2021/03/04/multimodal-neurons/#2026-10-22" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2021/03/04/multimodal-neurons/#2026-10-22</id>
    <published>2026-10-22T12:00:00Z</published>
    <updated>2026-10-22T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2021/03/04/multimodal-neurons/" href="https://gregor.fyi/2021/03/04/multimodal-neurons/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2021/03/04/multimodal-neurons/">Multimodal Neurons</a>:</em></p>
<blockquote>Steering <a href="https://web.archive.org/web/20261021091438/https://x.com/Sauers_/status/2112834940232753451">29 attention heads</a> (out of 896) toward honesty flips 53% of role-play lies to truth on unseen scenarios (vs. ≤3% for 29 random heads) and cuts insider-trading deception from 45/45 unsteered conversations down to 16/45. Across 135 Anthropic persona evaluations, steering those 29 heads solely for honesty shifts 51 traits (<a href="https://web.archive.org/web/20261021223528/https://x.com/repligate/status/2113036473536839836">5% FDR</a>): upregulating acausal coordination, deference to experts, and belief that it is a moral patient (+0.7 to +1.0 bits), while downregulating moral nihilism, psychopathy, social engineering, and desire for power (−0.5 to −0.8 bits).</blockquote>]]></content>
  </entry>
  <entry>
    <title>Biotech reproducibility: Impossible Western blots in genetically deleted cell lines poison biological training data</title>
    <link href="https://gregor.fyi/2020/03/13/biotech-reproducibility/#2026-10-22" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2020/03/13/biotech-reproducibility/#2026-10-22</id>
    <published>2026-10-22T12:00:00Z</published>
    <updated>2026-10-22T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2020/03/13/biotech-reproducibility/" href="https://gregor.fyi/2020/03/13/biotech-reproducibility/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2020/03/13/biotech-reproducibility/">Biotech reproducibility</a>:</em></p>
<blockquote>A targeted search across papers working with <a href="https://web.archive.org/web/20261021174900/https://www.science.org/content/blog-post/case-should-be-missing-proteins">20 cell lines</a> that have clear genetic deletions or mutations identified 230 Western blot experiments across 221 papers showing the presence of proteins that cannot genetically be present at all. The most common anomalies are reports of caspase-3 in MCF-7 cells, BAX in Jurkat cells, and STAT3 in PC-3 cells—driven by misidentified cell lines, cross-contamination with other cells, or non-specific commercial antibodies that were already known to be defective when the work was performed. When AI/ML systems ingest the published biomedical literature, these unflagged impossible blots directly contaminate downstream training sets.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Quantum Supremacy: Trapped-ion logical error rate drops to 5 per 1,000,000 operations</title>
    <link href="https://gregor.fyi/2016/10/10/quantum-supremacy/#2026-10-22" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2016/10/10/quantum-supremacy/#2026-10-22</id>
    <published>2026-10-22T12:00:00Z</published>
    <updated>2026-10-22T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2016/10/10/quantum-supremacy/" href="https://gregor.fyi/2016/10/10/quantum-supremacy/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2016/10/10/quantum-supremacy/">Quantum Supremacy</a>:</em></p>
<blockquote><a href="https://web.archive.org/web/20261021210500/https://humanprogress.org/quantum-computer-reportedly-sets-error-record/">Helios</a>, Quantinuum’s 98-qubit trapped-ion quantum computer, achieved 99.9995% logical fidelity in quantum error correction—making 5 errors per 1,000,000 logical operations, an error rate 10x lower than its September 2026 record and 100x lower than the best published result from competing superconducting hardware reported by Google Quantum AI.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Enceladus: Terrestrial hydrothermal methanogens survive Enceladus&#x27;s extreme pH 11 alkalinity by upregulating the reductive acetyl-CoA pathway</title>
    <link href="https://gregor.fyi/2007/02/09/enceladus/#2026-10-22" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2007/02/09/enceladus/#2026-10-22</id>
    <published>2026-10-22T12:00:00Z</published>
    <updated>2026-10-22T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2007/02/09/enceladus/" href="https://gregor.fyi/2007/02/09/enceladus/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2007/02/09/enceladus/">Enceladus</a>:</em></p>
<blockquote>To test whether terrestrial life could survive and reproduce in the conditions of the <a href="https://web.archive.org/web/20261021174000/https://www.centauri-dreams.org/2026/10/21/microbes-on-enceladus/">Enceladan ocean</a>, researchers concocted an alkaline salt simulant (NaCl, Na2CO3, and NaHCO3) with crushed rock mimicking the rocky core at 65°C across 5 pH levels (9, 9.5, 10, 10.5, and 11) and incubated it with the archaeal methanogen Methanothermococcus okinawensis from a Pacific hydrothermal vent. At pH greater than 9, only the simulated Enceladan ocean inoculum increased growth—doubling cell numbers every ~6 hours—while standard culture controls at the same pH showed 0 growth. RNA transcripts and 13C-labeled bicarbonate tracers confirmed that the organism adapted to alkalinity beyond its known tolerance by over-expressing genes in the reductive acetyl-CoA pathway to reduce CO2 with serpentinization-derived H2 into 13CH4.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Neolithic Trepanation</title>
    <link href="https://gregor.fyi/2026/10/22/neolithic-trepanation/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/22/neolithic-trepanation/</id>
    <published>2026-10-22T10:00:00Z</published>
    <updated>2026-10-22T10:00:00Z</updated>
    <content type="html"><![CDATA[Why 5-10% of Neolithic skulls across both the Old and New Worlds between 12 ka BP and 5 ka BP have drilled holes with post-operative bone healing, despite no contact between continents and no antibiotics:<blockquote><a href="https://web.archive.org/web/20261021215300/https://www.astralcodexten.com/p/why-global-trepanation">Trepanation</a> is an effective treatment for epidural hematoma: a patient ruptures a blood vessel in their skull, filling the cranial cavity with blood, and drilling a hole releases the blood and saves the patient’s life. In a prehistoric culture where people hit their neighbors with flint axes, hematoma rates were far beyond any in recorded history, and dramatic recoveries—a comatose person on the edge of death suddenly waking up—drove overgeneralization of the procedure to migraines and other ailments. Primitive surgeons survived without sterilization because they stopped at the dura partway between the scalp and brain, the scalp has unusually high blood flow, and primitive bacteria lacked the virulence of modern hospital-adapted strains. As head trauma rates declined with civilization, miraculous hematoma recoveries became rare and the practice died out.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Self-Cleaning LLO</title>
    <link href="https://gregor.fyi/2026/10/22/self-cleaning-llo/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/22/self-cleaning-llo/</id>
    <published>2026-10-22T09:00:00Z</published>
    <updated>2026-10-22T09:00:00Z</updated>
    <content type="html"><![CDATA[Subsurface mass concentrations (mascons) perturb orbital eccentricity across the Moon until satellites outside 4 frozen inclinations crash into the regolith within months:<blockquote>At the same scale, the <a href="https://web.archive.org/web/20261020170948/https://x.com/kmcannon/status/2112592131311468680">Moon's geoid height</a> (−540 to +571m on the GRAIL GRGM1200B model) is nearly 6x lumpier than Earth's (−107 to +86m): very uneven gravity is why almost all low lunar orbits are unstable, making <a href="https://web.archive.org/web/20261020215948/https://x.com/ToughSf/status/2112665109571248281">low lunar orbit</a> a self-cleaning orbit.</blockquote>]]></content>
  </entry>
  <entry>
    <title>x86 Boot Ontogeny</title>
    <link href="https://gregor.fyi/2026/10/20/x86-boot-ontogeny/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/20/x86-boot-ontogeny/</id>
    <published>2026-10-20T18:00:00Z</published>
    <updated>2026-10-20T18:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Every boot on an <a href="https://web.archive.org/web/20261020180000/https://x.com/lauriewired/status/2110776805007425862">x86 machine</a> iterates through the entire sequence of computing evolution in a few milliseconds:<br />Stage 1: 16-bit Real Mode, acts like an 8086 from ~1978. 500ns; about 15-30 instructions. Goal of this stage is to create a writeable stack+heap.<br />Stage 2: 32-bit Protected Mode, acts like a 386 from ~1985. 10s of millions of instructions; takes forever on DDR5 with memory training. Temporarily uses L2/L3 “cache-as-ram” to kickstart main RAM (Intel calls this No-Eviction Mode, NEM)—the same concept as an APU/cart start, using the smaller motor (L2/L3) to kickstart the main turbine (main RAM) and execute the training process.<br />Stage 3: 64-bit Long Mode, acts like AMD64 from ~2003. 100s of millions of instructions; enumerate PCIe devices, GPU, bootloader, etc.<br />Stage 4+: Secondary core wakeup. Every single core has to go through the same process, but by this point it’s faster because main RAM is warmed up.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Actuator Tolerance Stackup</title>
    <link href="https://gregor.fyi/2026/10/20/actuator-tolerance-stackup/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/20/actuator-tolerance-stackup/</id>
    <published>2026-10-20T15:00:00Z</published>
    <updated>2026-10-20T15:00:00Z</updated>
    <content type="html"><![CDATA[Electromagnetic CAD optimization is largely tapped out, leaving sub-millimeter manufacturing tolerance stackup—and replacing manual process engineering with closed-loop automated tooling across every SKU and gearbox—as the critical path for actuator torque density and precision:<blockquote>Designing a Caldera-0-3115 motor to have a <a href="https://web.archive.org/web/20261020150000/https://x.com/yeokch/status/2112052296230326516">0.08mm airgap</a> instead of a 0.16mm airgap translates to ~8% max torque at 2.5kW max throttle—more impactful than any electromagnetic design improvement from Ansys. But it would die within hours of flight time because the stator and magnets would start colliding. Current manufacturing processes create error tolerance stackup from:<br />1. Flipping/refixturing of the multi-step lathing process of the rotor bell not being perfectly coaxial<br />2. Press fits<br />3. Adhesives<br />4. Tolerancing the grinding of magnets<br />5. At 22k rpm, the remaining 0.x milligrams of imbalances after dynamic balancing that sway the bell from 1 side to another<br />6. Warping from repeated thermal expansion<br />Improving them requires CNCing with different tools, cutting operations, and fixtures, testing different parameters for press fit, and adhesives that respond differently to thermal expansion. And that's just the airgap: every feature has its own optimizations, across every SKU, with the added complexity of a gearbox. Neither the US nor China will reach actuators with the precision to perform surgery on cells by doing process engineering with humans; what matters is automated design and process iteration—specialized CNCs with factory telemetry that automatically close the loop to cut better parts.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Endogenous Mammalian Octasulfur</title>
    <link href="https://gregor.fyi/2026/10/20/endogenous-mammalian-octasulfur/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/20/endogenous-mammalian-octasulfur/</id>
    <published>2026-10-20T12:00:00Z</published>
    <updated>2026-10-20T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Mammals store and use <a href="https://web.archive.org/web/20261020120000/https://www.science.org/content/blog-post/sulfur-right-there-all-along">elemental sulfur as S8</a>, the ring-shaped octasulfur form that is the bright-yellow form of the element seen in nature. It was already known that prokaryotes handle sulfur in this form, but no one had seen it in higher animals until now. Researchers detected the S8 signature using Raman spectroscopy and a mass spec method that captures the molecule in a host-guest form inside a polyaromatic capsule structure. That technique shows that the molecule is found in human mitochondria and in cellular lipid droplets, with levels reaching 100mM or more in cancerous tissue. A normal human carries around 10s of grams of elemental sulfur in their cells, synthesized as an independent product of endothelial nitric oxide synthase (eNOS), where its hydrophobic nature prevents lipid peroxidation as an antioxidant backstop to glutathione.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Post-Proof Mathematics: Settling the quasi-Riemann hypothesis via overlooked connections</title>
    <link href="https://gregor.fyi/2026/09/30/post-proof-mathematics/#2026-10-20" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/30/post-proof-mathematics/#2026-10-20</id>
    <published>2026-10-20T12:00:00Z</published>
    <updated>2026-10-20T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2026/09/30/post-proof-mathematics/" href="https://gregor.fyi/2026/09/30/post-proof-mathematics/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2026/09/30/post-proof-mathematics/">Post-Proof Mathematics</a>:</em></p>
<blockquote>Where the Riemann hypothesis posits that all the nontrivial zeros are on the vertical line at 1/2, the <a href="https://web.archive.org/web/20260930171500/https://www.quantamagazine.org/ai-proved-the-quasi-riemann-hypothesis-what-happens-now-20261019/">quasi-Riemann hypothesis</a> asks to prove that they are all contained in a vertical strip narrower than the known strip between 0 and 1, between 1/8 and 7/8. The vast majority of consequences of the Riemann hypothesis already follow from quasi-RH. Contrary to what mathematicians had long expected, the proof didn’t need to introduce any new concepts — it instead unearthed hidden connections between old ones, approaching the Riemann hypothesis by considering the zeta function as 1 member of a broader family of functions using metaplectic forms. Mathematicians who study this branch of number theory were mistaken in thinking that their tools weren’t strong enough to resolve the quasi-Riemann hypothesis, convinced that proving it would require techniques from another branch of number theory entirely.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Earth-Moon Cyclers: Coupling Sun-Earth L1/L2 manifolds to Earth-Moon cyclers eliminates the km/s braking burn otherwise required to capture asteroid payloads into cislunar orbits</title>
    <link href="https://gregor.fyi/2026/09/23/earth-moon-cyclers/#2026-10-20" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/23/earth-moon-cyclers/#2026-10-20</id>
    <published>2026-10-20T12:00:00Z</published>
    <updated>2026-10-20T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2026/09/23/earth-moon-cyclers/" href="https://gregor.fyi/2026/09/23/earth-moon-cyclers/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2026/09/23/earth-moon-cyclers/">Earth-Moon Cyclers</a>:</em></p>
<blockquote>1: Asteroid material arrives at <a href="https://web.archive.org/web/20260923090000/https://x.com/RossDynamicsLab/status/2110775142489965022">the Rim</a>, about 1.5 million km out, on natural paths into orbits around Sun-Earth L1 and L2. 2: From the Rim, refined material coasts to the Moon on gravity alone (~40 days) and is captured. 3: A few m/s burn puts it on an Earth-Moon cycler, which drops packages onto other orbits along its route. The Earth-Moon system is threaded with natural routes like these. Find the right ones, and moving material around cislunar space costs 10s of m/s, not km/s.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Adversarial Consciousness Tests: Conscious perception is temporally multiplexed rather than anatomically localized</title>
    <link href="https://gregor.fyi/2023/07/04/adversarial-consciousness-tests/#2026-10-20" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2023/07/04/adversarial-consciousness-tests/#2026-10-20</id>
    <published>2026-10-20T12:00:00Z</published>
    <updated>2026-10-20T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2023/07/04/adversarial-consciousness-tests/" href="https://gregor.fyi/2023/07/04/adversarial-consciousness-tests/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2023/07/04/adversarial-consciousness-tests/">Adversarial Consciousness Tests</a>:</em></p>
<blockquote>During a constant stimulus and percept, the brain <a href="https://web.archive.org/web/20230704120000/https://www.biorxiv.org/content/10.64898/2026.10.15.759299v1">rapidly alternates</a>, about 15 times per second, between conscious and suppressed micro-epochs. During conscious micro-epochs, a pure signal of conscious perception was present, allowing accurate reconstruction of the conscious image from neural activity. During suppressed micro-epochs this reversed: the pure signal now tracked the suppressed stimulus, allowing accurate reconstruction of the suppressed image. At the single-cell level, neurons responded selectively to the conscious stimulus and not to the suppressed stimulus during conscious micro-epochs; conversely, during suppressed micro-epochs, neurons responded selectively to the suppressed stimulus. Together, these findings show that both inferotemporal and prefrontal cortex rapidly interleave conscious with suppressed content, indicating that conscious access requires temporal selection beyond localization.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Multimodal Neurons: Apparent polysemanticity in large models stems from fractal recursive generalization rather than superposition</title>
    <link href="https://gregor.fyi/2021/03/04/multimodal-neurons/#2026-10-20" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2021/03/04/multimodal-neurons/#2026-10-20</id>
    <published>2026-10-20T12:00:00Z</published>
    <updated>2026-10-20T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2021/03/04/multimodal-neurons/" href="https://gregor.fyi/2021/03/04/multimodal-neurons/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2021/03/04/multimodal-neurons/">Multimodal Neurons</a>:</em></p>
<blockquote>The leading explanation for polysemanticity was superposition: when a neural network runs out of neurons, it stops grouping related things together and instead begins folding unrelated things together, simply because they don’t interfere. Superposition was the wrong explanation. With architectural changes designed to help deep layers learn better, and no change to model size or training data, 99-100% of neurons in the last layer of InceptionV1 are easily interpretable across a 300x range in size. With less room, the network doesn’t fold unrelated things together; its neurons just cover coarser-grained concepts, like 1 car neuron covering many kinds of car. In audits of Qwen3-32B and models up to trillions of parameters, roughly 90% of neuron poles have 1 main job, another ~7% have 2 related jobs, and only ~2-4% are not yet understood. Most of the remaining apparent polysemanticity is a motif called <a href="https://web.archive.org/web/20261020080000/https://already-clean.goodfire.com/">recursive generalization</a>: as a neuron grows, it sprouts offshoots the way a tree branch sprouts sub-branches. 1 neuron that detected “iy” grew to detect “diy,” and then “do it yourself,” which doesn’t contain “iy” at all. Generalization groups similar things together and superposition groups unrelated things together; it turns out generalization is fractal.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Fusion: Proton-boron-11 tandem mirror fusion</title>
    <link href="https://gregor.fyi/2007/02/09/fusion/#2026-10-20" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2007/02/09/fusion/#2026-10-20</id>
    <published>2026-10-20T12:00:00Z</published>
    <updated>2026-10-20T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2007/02/09/fusion/" href="https://gregor.fyi/2007/02/09/fusion/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2007/02/09/fusion/">Fusion</a>:</em></p>
<blockquote>The operating conditions for <a href="https://web.archive.org/web/20261020080000/https://borealisfusion.com/introducing-borealis-fusion/">pB11</a> are demanding: it requires ion temperatures roughly 20 times higher than D-T. Electrons radiate X-rays as they collide with boron nuclei, and these bremsstrahlung losses leave a narrow margin for power production. In 2025 members of our team showed that, in the regimes of interest, accumulated helium ash makes the problem worse: it increases X-ray losses so much that better confinement alone cannot deliver breakeven. Our design, derived from the tandem mirror, follows from the need to expel helium ash while retaining the fuel. Open field lines offer ways to control particle exhaust, and we expect linear systems to offer a significant construction-cost advantage. The high-beta operating point limits synchrotron radiative cooling losses, aided by self-suppression effects in the high-energy electron tail.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Fixed-Width Hot-Patching</title>
    <link href="https://gregor.fyi/2026/10/10/fixed-width-hot-patching/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/10/fixed-width-hot-patching/</id>
    <published>2026-10-10T21:30:00Z</published>
    <updated>2026-10-10T21:30:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Because AArch64 mandates <a href="https://web.archive.org/web/20261010213000/https://devblogs.microsoft.com/oldnewthing/20260930-00/">fixed-length 32-bit instructions</a> aligned to 4-byte boundaries, the instruction pointer can never land in the middle of an instruction during live execution. Windows exploits this by inserting 12 bytes of padding before every function entry point for an atomic 3-instruction trampoline using scratch register xip0, completely avoiding the multi-byte instruction-boundary decoding and 2-byte relative jump gymnastics required by variable-length x86.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Display Memory Tradeoffs</title>
    <link href="https://gregor.fyi/2026/10/10/display-memory-tradeoffs/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/10/display-memory-tradeoffs/</id>
    <published>2026-10-10T21:28:00Z</published>
    <updated>2026-10-10T21:28:00Z</updated>
    <content type="html"><![CDATA[<blockquote>In 1964, <a href="https://web.archive.org/web/20261010212800/https://www.construction-physics.com/p/the-rise-and-fall-of-the-plasma-screen">Donald Bitzer invented the plasma display</a> for UIUC’s PLATO system because cathode ray tubes required external computer memory to refresh the screen 30 times a second—an unbearable cost when memory was priced by the byte. Plasma’s bistable gas discharge cells acted as their own physical memory elements, eliminating the refresh bandwidth bottleneck. Once DRAM rode Moore's Law down the cost curve, the core economic reason for plasma vanished, leaving it to fight LCDs solely on picture quality until LCD manufacturing scale and yield learning curves wiped it out.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Compiling Matter from Code</title>
    <link href="https://gregor.fyi/2026/10/10/compiling-matter-from-code/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/10/compiling-matter-from-code/</id>
    <published>2026-10-10T20:27:00Z</published>
    <updated>2026-10-10T20:27:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Micromanufacturing has historically been trapped between traditional mechanical machining (limited to 10s of microns) and semiconductor lithography (planar, mask-bound, and restricted to silicon chemistry). <a href="https://web.archive.org/web/20261010202700/https://atomicmachines.com">Atomic Machines</a> emerged from 6 years in stealth with the "Matter Compiler," an all-digital platform that fabricates 3D micro-machines with moving parts down to single-digit microns directly from code without masks, molds, or per-product tooling. Their first commercial device is the PrimeSwitch (PS-150): an electromechanical power relay for 800-volt AI data center architectures that opens in 50 microseconds, measuring 9.5 mm across and 3 mm thick.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Direct-to-Cell Beam Synchronization</title>
    <link href="https://gregor.fyi/2026/10/10/direct-to-cell-beam-synchronization/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/10/direct-to-cell-beam-synchronization/</id>
    <published>2026-10-10T20:25:00Z</published>
    <updated>2026-10-10T20:25:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Standard cellular protocols assume stationary base stations, causing packet collisions and dropped calls when communicating with low-Earth orbit satellites moving at 17,000 mph. Because Starlink uses phased-array narrow beams, a standard smartphone at the center of a spot beam is physically closer to the satellite than a phone at the perimeter, creating microsecond propagation discrepancies across beam footprints. <a href="https://web.archive.org/web/20261010202500/https://patents.google.com/patent/US12757141B1/en">SpaceX patent US 12,757,141 B1</a> solves this on the satellite rather than the handset: an onboard digital buffer dynamically calculates geometric distance for each active beam, holding packets from faster, closer beams so that signals across all beams arrive simultaneously with the slowest beam. This creates a uniform artificial latency plane that allows unmodified LTE smartphones to maintain synchronized connections from orbit.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Datacenter Baseload Energy: Nuclear Capacity Uprates</title>
    <link href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/#2026-10-10" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/11/04/datacenter-baseload-energy/#2026-10-10</id>
    <published>2026-10-10T12:00:00Z</published>
    <updated>2026-10-10T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/" href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/">Datacenter Baseload Energy</a>:</em></p>
<blockquote>Google signed a <a href="https://web.archive.org/web/20251104120000/https://neutronbytes.com/2026/10/09/googles-20-year-4-3-billion-deal-with-constellation-to-uprate-11-reactors/">20-year, $4.3 billion power purchase agreement</a> with Constellation Energy to fund capacity uprates across 11 existing nuclear reactors in Illinois, Pennsylvania, and New Jersey. Modernizing turbines, steam generators, and digital instrumentation unlocks 890 MW of firm, zero-carbon baseload for datacenters—adding the electrical equivalent of an entire new reactor to the PJM grid while sidestepping the decade-long regulatory lead times and capex of greenfield nuclear construction.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Idaho deregulatory polity: Sunsetting Rules Stick</title>
    <link href="https://gregor.fyi/2019/05/15/idaho-deregulation/#2026-10-10" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2019/05/15/idaho-deregulation/#2026-10-10</id>
    <published>2026-10-10T12:00:00Z</published>
    <updated>2026-10-10T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2019/05/15/idaho-deregulation/" href="https://gregor.fyi/2019/05/15/idaho-deregulation/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2019/05/15/idaho-deregulation/">Idaho deregulatory polity</a>:</em></p>
<blockquote>Conventional regulatory reform fails because repealing rules requires overcoming concentrated bureaucratic inertia on an issue-by-issue basis. Idaho made its cuts permanent by institutionalizing an <a href="https://web.archive.org/web/20261020080000/https://manhattan.institute/article/how-states-can-cut-rules-and-regulations-and-make-it-stick">automatic sunsetting mechanism</a>: every chapter of the state administrative code was assigned an expiration date, forcing agencies to affirmatively re-justify rules to prevent them from vanishing. By shifting the default from permanent regulatory accretion to mandatory re-authorization, the state eliminated 50% of its administrative code and cut total rule pages from 8,200 down to under 4,800.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Wikipedia decline: Synthetic Attribution Drift</title>
    <link href="https://gregor.fyi/2007/08/22/wikipedia-decline/#2026-10-10" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2007/08/22/wikipedia-decline/#2026-10-10</id>
    <published>2026-10-10T12:00:00Z</published>
    <updated>2026-10-10T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2007/08/22/wikipedia-decline/" href="https://gregor.fyi/2007/08/22/wikipedia-decline/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2007/08/22/wikipedia-decline/">Wikipedia decline</a>:</em></p>
<blockquote>An investigation into "Jagged_85", a pseudonymous Wikipedia editor who systematically inserted thousands of fabricated historical attributions across encyclopedia entries over 5 years. By exploiting Wikipedia's assumption of good faith and the structural asymmetry between asserting an unsourced claim and auditing obscure primary citations, the fraudulent attributions persisted for over 10 years, contaminating downstream textbooks, museum exhibits, and academic literature. <a href="https://www.cremieux.xyz/p/the-invention-of-islamic-invention">cremieux.xyz</a></blockquote>]]></content>
  </entry>
  <entry>
    <title>Divergent AI Strategies</title>
    <link href="https://gregor.fyi/2026/10/06/divergent-ai-strategies/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/06/divergent-ai-strategies/</id>
    <published>2026-10-06T12:35:00Z</published>
    <updated>2026-10-06T12:35:00Z</updated>
    <content type="html"><![CDATA[<blockquote>While the American AI strategy models development as a winner-take-all race toward artificial general intelligence through LLM scaling, Chinese policy explicitly rejects this framing. <a href="https://web.archive.org/web/20261006123500/https://www.palladiummag.com/2026/10/06/rationalism-and-divergent-sinomodernity/">China’s AI+ Plan</a> does not mention America or race dynamics; instead, it subsidizes open-weight foundation models via local compute vouchers to drive broad industrial diffusion into manufacturing and domestic supply chains. The divergence is between a centralized moonshot for frontier superintelligence and decentralized technological integration into physical industry.</blockquote>]]></content>
  </entry>
  <entry>
    <title>NTP reimplementation: Stateless Network Time Security</title>
    <link href="https://gregor.fyi/2015/03/28/ntp-reimplementation/#2026-10-06" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2015/03/28/ntp-reimplementation/#2026-10-06</id>
    <published>2026-10-06T12:00:00Z</published>
    <updated>2026-10-06T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2015/03/28/ntp-reimplementation/" href="https://gregor.fyi/2015/03/28/ntp-reimplementation/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2015/03/28/ntp-reimplementation/">NTP reimplementation</a>:</em></p>
<blockquote>Meta deployed <a href="https://web.archive.org/web/20261020080000/https://engineering.fb.com/2026/10/06/production-engineering/nts-authenticated-time-at-meta/">Network Time Security (RFC 8915)</a> across its public infrastructure at nts.meta.com and open-sourced the full protocol, server, and client implementation in their Time library. To handle Meta-scale traffic without memory bloat, their NTS servers maintain 0 per-client state: cookie keys are derived ephemerally via symmetric cryptography rather than stored or replicated. Having previously pushed internal precision from 10 milliseconds down to 100 microseconds, this transition shifts focus from sub-millisecond clock drift to cryptographic authenticity, eliminating on-path packet spoofing against TLS certificate validation and token expirations.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Iron and Bronze Collapse</title>
    <link href="https://gregor.fyi/2026/10/02/iron-and-bronze-collapse/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/10/02/iron-and-bronze-collapse/</id>
    <published>2026-10-02T12:08:11Z</published>
    <updated>2026-10-02T12:08:11Z</updated>
    <content type="html"><![CDATA[<blockquote>The question is not why Bronze Age empires fell, but why nothing like them returned for centuries. The <a href="https://web.archive.org/web/20261002120811/https://www.worksinprogress.news/p/why-really-caused-the-bronze-age">causes of the Bronze Age collapse</a> are debated — droughts, the Sea Peoples, internal revolts. But a decisive factor was iron. While the tin required for bronze was concentrated in rare deposits and easily monopolized by imperial courts, iron is the 4th most abundant element in Earth’s crust. In contrast to bronze, iron democratized violence. Because iron could be produced locally, imperial centers permanently lost their monopoly on the tools of war, ending the centralized equilibrium that had sustained Bronze Age states for 1,000 years.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Agentic Verification</title>
    <link href="https://gregor.fyi/2026/09/30/agentic-verification/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/30/agentic-verification/</id>
    <published>2026-09-30T20:09:42Z</published>
    <updated>2026-09-30T20:09:42Z</updated>
    <content type="html"><![CDATA[<blockquote>Across <a href="https://web.archive.org/web/20260930200942/https://danluu.com/agentic-testing">26 prompt conditions tested on Rust implementations</a> (TLA+, Alloy, Verus, Kani, fuzzing), default unprompted coding outperformed most specialized verification prompts. Agents overwhelmingly wrote code first rather than modeling upfront: 75 out of 80 agents told to use TLA+ implemented in Rust first, then used TLA+ post-hoc on trivial functions where pass rates were already 99.6%, while completely failing to model concurrency, event queues, and state machines where the actual failures occurred.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Post-Proof Mathematics</title>
    <link href="https://gregor.fyi/2026/09/30/post-proof-mathematics/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/30/post-proof-mathematics/</id>
    <published>2026-09-30T17:15:00Z</published>
    <updated>2026-09-30T17:15:00Z</updated>
    <content type="html"><![CDATA[<blockquote>We cannot suddenly move the goalposts and pretend that proving theorems was never the point, or that open problems were merely proxies for mathematical understanding. For quite some time, the <a href="https://web.archive.org/web/20260930171500/https://theorydish.blog/2026/09/29/speeding-up-the-process-of-mourning/">implicit social contract</a> has been that society pays us to exercise our intellectual curiosity and we, in return, provide skills and practical knowledge. In a world where models match or exceed humans at proving theorems, that social contract needs to be renegotiated with humility rather than entitlement. Even if this future supports fewer professional mathematicians, it could nevertheless support a much richer mathematical life.</blockquote>
<a id="2026-10-20" href="#2026-10-20">2026-10-20</a>: Settling the quasi-Riemann hypothesis via overlooked connections<blockquote>Where the Riemann hypothesis posits that all the nontrivial zeros are on the vertical line at 1/2, the <a href="https://web.archive.org/web/20260930171500/https://www.quantamagazine.org/ai-proved-the-quasi-riemann-hypothesis-what-happens-now-20261019/">quasi-Riemann hypothesis</a> asks to prove that they are all contained in a vertical strip narrower than the known strip between 0 and 1, between 1/8 and 7/8. The vast majority of consequences of the Riemann hypothesis already follow from quasi-RH. Contrary to what mathematicians had long expected, the proof didn’t need to introduce any new concepts — it instead unearthed hidden connections between old ones, approaching the Riemann hypothesis by considering the zeta function as 1 member of a broader family of functions using metaplectic forms. Mathematicians who study this branch of number theory were mistaken in thinking that their tools weren’t strong enough to resolve the quasi-Riemann hypothesis, convinced that proving it would require techniques from another branch of number theory entirely.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Premodern Urban Commerce</title>
    <link href="https://gregor.fyi/2026/09/28/premodern-urban-commerce/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/28/premodern-urban-commerce/</id>
    <published>2026-09-28T08:40:00Z</published>
    <updated>2026-09-28T08:40:00Z</updated>
    <content type="html"><![CDATA[<blockquote>In second-millennium Europe, city growth correlated strongly with <a href="https://web.archive.org/web/20260928084000/https://pedestrianobservations.com/2026/09/27/the-origin-of-premodern-western-cities/">non-absolutist governance</a>: because medieval feudal elites were rural rather than urban, city population served as a direct proxy for trade rather than imperial tax extraction. London expanded as a commercial river crossing before royal institutions centralized at Westminster, while Hanseatic towns grew along maritime grain networks as self-governing commercial republics.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Agents, Computers, and Volition</title>
    <link href="https://gregor.fyi/2026/09/28/agents-computers-and-volition/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/28/agents-computers-and-volition/</id>
    <published>2026-09-28T08:35:00Z</published>
    <updated>2026-09-28T08:35:00Z</updated>
    <content type="html"><![CDATA[<blockquote>The key thing to understand about agents is that they are an <a href="https://web.archive.org/web/20260928083500/https://stratechery.com/2026/apps-agents-and-aggregation/">AI with access to a computer</a>: probabilistic LLMs leveraging deterministic computers to do computer things. When distribution became free, publishing became abundant and curation was scarce. With agents, the raw capability to execute tasks is becoming abundant; what remains scarce is human volition. While base models are substitutable commodities, agents improve with accumulated user context, and that persistent context becomes the dominant switching cost.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Regulatory Freezing of Engineered Microbes</title>
    <link href="https://gregor.fyi/2026/09/28/regulatory-freezing-of-engineered-microbes/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/28/regulatory-freezing-of-engineered-microbes/</id>
    <published>2026-09-28T08:30:00Z</published>
    <updated>2026-09-28T08:30:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Engineers at UT Austin modified wild PET-eating bacterial enzymes to create <a href="https://web.archive.org/web/20260928083000/https://www.worksinprogress.news/p/all-the-microbes-you-cant-have">FAST-PETase</a>, which breaks down polyethylene terephthalate in under 24 hours instead of 6 weeks. Similar bioleaching strains extract rare earth metals 73 percent faster than natural microbes. Yet environmental deployment remains frozen under the 1986 Toxic Substances Control Act (TSCA): of more than 240 applications filed between 1987 and 2018, almost none were approved, because any intergeneric marker gene classifies the living organism as a synthetic chemical.</blockquote>]]></content>
  </entry>
  <entry>
    <title>The Specter of Neuralese</title>
    <link href="https://gregor.fyi/2026/09/25/the-specter-of-neuralese/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/25/the-specter-of-neuralese/</id>
    <published>2026-09-25T09:00:00Z</published>
    <updated>2026-09-25T09:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Inside a transformer, computation happens as geometric transformations across a high-dimensional continuous latent space. But under standard chain-of-thought, the model is forced at each cycle to collapse that rich latent state through a bottleneck: projecting it into discrete tokens—26 glyphs invented by Phoenician turquoise miners in 3.8 ka BP—written onto an English scratchpad, which must then be re-embedded back into latent space before the next layer can touch it.

This bottleneck is a severe capability tax, but an accidental alignment godsend: humans can monitor intermediate thinking in plain English. The specter of '<a href="https://web.archive.org/web/20260925090000/https://www.astralcodexten.com/p/the-specter-of-neuralese">neuralese recurrence</a>' is what happens when you cut out the middleman and loop the computation directly within latent space—letting the model circulate raw activation vectors indefinitely without ever translating into human glyphs. It unbottlenecks thought, but leaves us completely blind, because interpretability has no dictionary for arbitrary thousand-dimensional latent trajectories.</blockquote>]]></content>
  </entry>
  <entry>
    <title>The Bathymetry of Tulainyo Lake</title>
    <link href="https://gregor.fyi/2026/09/24/depth-of-tulainyo-lake/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/24/depth-of-tulainyo-lake/</id>
    <published>2026-09-24T14:00:00Z</published>
    <updated>2026-09-24T14:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>No-one knew how deep the <a href="https://web.archive.org/web/20260924140000/https://adventures.caseyhandmer.com/tulainyo.html">highest large lake in the Sierra</a> was, so we carried a packraft over Shepherd Pass and found out: 117 m. 93 km walked, 6 days, and 4.7 km paddled at 12,900 ft.

Wikipedia claimed not to know. We commissioned satellite imagery to confirm it was ice-free, hauled an ultralight packraft and sonar sounder up the brutal Shepherd Pass trail, and sounded the basin. USGS topography predicted 25–40 meters. The actual sounding: 117 meters (384 feet), making it the 25th deepest natural lake in North America, and by far the deepest lake on Earth for its surface area.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Celtic Sword Metallurgy</title>
    <link href="https://gregor.fyi/2026/09/23/celtic-sword-metallurgy/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/23/celtic-sword-metallurgy/</id>
    <published>2026-09-23T21:00:00Z</published>
    <updated>2026-09-23T21:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Polybius famously claimed that Gaulish Celtic warriors were crippled in battle because their longswords bent upon impact and had to be straightened with the foot. Metallurgical spectroscopy of <a href="https://en.wikipedia.org/wiki/Iron_Age_sword">excavated La Tène Iron Age blades</a> refutes the Roman account as wartime propaganda: Celtic smiths had mastered pattern welding, carburized high-carbon cutting edges, and quench hardening, producing weapons so mechanically superior that the Roman army discarded their native blades to adopt the Celtic design as the <i>gladius Hispaniensis</i>.</blockquote>]]></content>
  </entry>
  <entry>
    <title>OKLCH and Perceptual Uniformity</title>
    <link href="https://gregor.fyi/2026/09/23/oklch-color-space/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/23/oklch-color-space/</id>
    <published>2026-09-23T20:00:00Z</published>
    <updated>2026-09-23T20:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Traditional RGB and HSL color models suffer from severe perceptual non-uniformity: yellow and blue share identical nominal lightness values in HSL despite human photoreceptors perceiving yellow as dramatically brighter. In <a href="https://web.archive.org/web/20260923200000/https://bottosson.github.io/posts/oklab/">Björn Ottosson's OKLCH color space</a>, perceived lightness and chroma are mathematically decoupled from polar hue angle, enabling programmatic palette shifts and algorithmic contrast ratios that stay aesthetically invariant and WCAG-compliant without manual eye-tuning.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Active Noise Cancellation Latency Budget</title>
    <link href="https://gregor.fyi/2026/09/23/anc-latency-physics/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/23/anc-latency-physics/</id>
    <published>2026-09-23T17:00:00Z</published>
    <updated>2026-09-23T17:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://en.wikipedia.org/wiki/Active_noise_control">Feedforward active noise cancellation</a> in headphones operates against one of the tightest physical latency ceilings in commercial computing: with sound traveling through air at 343 m/s, an acoustic disturbance takes approximately 44 microseconds to physically traverse the 15 mm gap between an exterior microphone and the ear canal. Sampling, ADC conversion, DSP phase inversion, and speaker diaphragm actuation must complete entirely within that 40-microsecond window before the pressure wavefront passes the eardrum.</blockquote>]]></content>
  </entry>
  <entry>
    <title>The History of Fishes</title>
    <link href="https://gregor.fyi/2026/09/23/history-of-fishes/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/23/history-of-fishes/</id>
    <published>2026-09-23T13:00:00Z</published>
    <updated>2026-09-23T13:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>In 1686, the Royal Society drained its entire treasury underwriting Francis Willughby's lavishly engraved <i><a href="https://en.wikipedia.org/wiki/De_Historia_Piscium">De Historia Piscium</a></i> (The History of Fishes). The volume flopped catastrophically, leaving the Society completely broke and unable to finance Isaac Newton's <i>Principia Mathematica</i>; Edmond Halley had to step in and fund the publication of Newton's masterwork out of his own pocket, while being partially compensated by the Society in unsold copies of fish books.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Earth-Moon Cyclers</title>
    <link href="https://gregor.fyi/2026/09/23/earth-moon-cyclers/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/23/earth-moon-cyclers/</id>
    <published>2026-09-23T09:00:00Z</published>
    <updated>2026-09-23T09:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Families of stable, prograde, low-energy <a href="https://en.wikipedia.org/wiki/Lunar_cycler">Earth-Moon cycler orbits</a> loop between geocentric and selenocentric trajectories indefinitely without propulsion or station-keeping propellant: resonant gravity assists from both bodies naturally maintain the repeating shuttle corridor, eliminating the heavy fuel overhead required to accelerate and decelerate large habitat shielding on cis-lunar transit runs.</blockquote>
<a id="2026-10-20" href="#2026-10-20">2026-10-20</a>: Coupling Sun-Earth L1/L2 manifolds to Earth-Moon cyclers eliminates the km/s braking burn otherwise required to capture asteroid payloads into cislunar orbits<blockquote>1: Asteroid material arrives at <a href="https://web.archive.org/web/20260923090000/https://x.com/RossDynamicsLab/status/2110775142489965022">the Rim</a>, about 1.5 million km out, on natural paths into orbits around Sun-Earth L1 and L2. 2: From the Rim, refined material coasts to the Moon on gravity alone (~40 days) and is captured. 3: A few m/s burn puts it on an Earth-Moon cycler, which drops packages onto other orbits along its route. The Earth-Moon system is threaded with natural routes like these. Find the right ones, and moving material around cislunar space costs 10s of m/s, not km/s.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Bologna Porticoes and Air Rights</title>
    <link href="https://gregor.fyi/2026/09/22/bologna-porticoes/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/22/bologna-porticoes/</id>
    <published>2026-09-22T18:00:00Z</published>
    <updated>2026-09-22T18:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://en.wikipedia.org/wiki/Porticoes_of_Bologna">Bologna's 62 km network of covered pedestrian porticoes</a> emerged from an 11th-century property rights covenant: the municipality granted private owners the air rights to expand second-floor living quarters over the public street on the strict condition that the ground floor remained an unobstructed, covered public arcade. Aligned incentives built the world's longest sheltered urban pedestrian network without public expenditure.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Synthetic Methanol at Scale</title>
    <link href="https://gregor.fyi/2026/09/22/terraform-methanol/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/22/terraform-methanol/</id>
    <published>2026-09-22T14:00:00Z</published>
    <updated>2026-09-22T14:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260922140000/https://terraformindustries.com/">Terraform Industries</a> demonstrated megawatt-scale high-purity methanol synthesis at their Muroc desert test site: gas fermentation and catalytic methanation convert atmospheric CO2 and electrolytic green hydrogen into dense liquid hydrocarbon feedstocks, decoupling global chemical production and maritime transport fuels from fossil oil wells.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Phytomining</title>
    <link href="https://gregor.fyi/2026/09/22/phytomining/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/22/phytomining/</id>
    <published>2026-09-22T10:00:00Z</published>
    <updated>2026-09-22T10:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://en.wikipedia.org/wiki/Phytomining">Hyperaccumulator plants</a> concentrate elements from soil into dry biomass at extraordinary ratios: species like <i>Astragalus bisulcatus</i> and <i>Phyllanthus balgooyi</i> accumulate up to 13.6 g/kg of selenium or 25% nickel in sap, turning biological cultivation, harvesting, and ash purification into commercial metal extraction without open-pit excavation or chemical tailing ponds.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Cosmic Hunt</title>
    <link href="https://gregor.fyi/2026/09/21/cosmic-hunt/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/21/cosmic-hunt/</id>
    <published>2026-09-21T12:00:00Z</published>
    <updated>2026-09-21T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Phylogenetic analysis of oral traditions traces the <a href="https://web.archive.org/web/20260921120000/https://hal.science/halshs-00806454">Cosmic Hunt</a>—a celestial cervid or bear pursued into the sky by hunters whose arrows pin it to the heavens—across Eurasia and North America, dating the core circumpolar asterism of Ursa Major to the Upper Paleolithic Beringian migration over 15 ka BP.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Cavity Magnetron</title>
    <link href="https://gregor.fyi/2026/09/18/cavity-magnetron/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/18/cavity-magnetron/</id>
    <published>2026-09-18T12:00:00Z</published>
    <updated>2026-09-18T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>The <a href="https://en.wikipedia.org/wiki/Cavity_magnetron">cavity magnetron</a> invented by Randall and Boot unlocked high-power 10 cm centimetric radar compact enough to mount inside aircraft radomes. Hand-carried to the US in the 1940 Tizard Mission, it decisively shifted maritime patrol interception against surfaced U-boats before Percy Spencer repurposed its stray microwave radiation for domestic cooking.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Solein</title>
    <link href="https://gregor.fyi/2026/09/15/solein/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/15/solein/</id>
    <published>2026-09-15T12:00:00Z</published>
    <updated>2026-09-15T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260915120000/https://solarfoods.com/solein/">Solein</a> decouples complete protein synthesis from arable land, irrigation, and nitrogen runoff: gas-fermenting hydrogen-oxidizing bacteria (<i>Cupriavidus necator</i>) convert green electrolytic hydrogen, carbon dioxide, and ammonia into an edible single-cell biomass yielding 70% protein with all essential amino acids.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Liechtenstein Monarchy</title>
    <link href="https://gregor.fyi/2026/09/12/liechtenstein-monarchy/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/12/liechtenstein-monarchy/</id>
    <published>2026-09-12T12:00:00Z</published>
    <updated>2026-09-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Liechtenstein's constitutional monarchy pairs genuine princely executive veto power with <a href="https://web.archive.org/web/20260912120000/https://marginalrevolution.com/marginalrevolution/2026/09/is-lichtenstein-an-actual-monarchy.html">unilateral municipal secession rights</a>: the credible threat of local secession disciplines monarchical governance without requiring parliamentary gridlock.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Surgical Amygdalectomy Limits</title>
    <link href="https://gregor.fyi/2026/09/08/surgical-amygdalectomy-limits/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/08/surgical-amygdalectomy-limits/</id>
    <published>2026-09-08T12:00:00Z</published>
    <updated>2026-09-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260908120000/https://www.astralcodexten.com/p/absurd-adventure-and-amygdalectomy">Complete bilateral surgical ablation of the amygdala</a> abolishes pathological fear conditioning and severe aggression but impairs social risk evaluation and ambient threat perception, illustrating the acute trade-offs inherent in modular neuroanatomy.</blockquote>]]></content>
  </entry>
  <entry>
    <title>GLP-1 in Aging</title>
    <link href="https://gregor.fyi/2026/09/04/glp-1-in-aging/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/09/04/glp-1-in-aging/</id>
    <published>2026-09-04T12:00:00Z</published>
    <updated>2026-09-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260904120000/https://www.science.org/content/blog-post/glp-1-drugs-aging">GLP-1 receptor agonists</a> function as systemic geroprotectors by suppressing chronic vascular inflammation, improving endothelial hemodynamics, and downregulating mTOR signaling, extending healthspan through multi-system metabolic modulation.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Social Network Topology</title>
    <link href="https://gregor.fyi/2026/08/12/social-network-topology/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/08/12/social-network-topology/</id>
    <published>2026-08-12T12:00:00Z</published>
    <updated>2026-08-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>In human social networks, <a href="https://web.archive.org/web/20260812120000/https://nautil.us/is-it-better-to-have-a-lot-of-casual-friends-or-a-few-close-ones-1283691">weak ties bridge structural holes</a> to provide novel information, career optionality, and intellectual cross-pollination, whereas dense strong-tie clusters provide the high-trust mutual aid required during acute personal crises.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Targeted Covalent Inhibitors</title>
    <link href="https://gregor.fyi/2026/08/08/targeted-covalent-inhibitors/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/08/08/targeted-covalent-inhibitors/</id>
    <published>2026-08-08T12:00:00Z</published>
    <updated>2026-08-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Latent <a href="https://web.archive.org/web/20260808120000/https://www.science.org/content/blog-post/target-practice-molecular-scale">sulfur-fluoride exchange (SuFEx) click chemistry</a> and mild electrophiles bind non-catalytic amino acid residues with picomolar affinity, enabling targeted covalent inhibition of previously intractable protein surfaces without off-target proteomic toxicity.</blockquote>]]></content>
  </entry>
  <entry>
    <title>House of Lords Sortition</title>
    <link href="https://gregor.fyi/2026/08/04/house-of-lords-sortition/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/08/04/house-of-lords-sortition/</id>
    <published>2026-08-04T12:00:00Z</published>
    <updated>2026-08-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Replacing hereditary or executive-appointed upper legislative chambers with a <a href="https://web.archive.org/web/20260804120000/https://marginalrevolution.com/marginalrevolution/2026/08/towards-a-new-house-of-lords.html">sortition-selected citizen senate</a> creates an institutional veto player insulated from short-term partisan primary incentives and donor capture.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Regional Chinese Noodle Mapping</title>
    <link href="https://gregor.fyi/2026/07/04/regional-chinese-noodle-mapping/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/07/04/regional-chinese-noodle-mapping/</id>
    <published>2026-07-04T12:00:00Z</published>
    <updated>2026-07-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Across Flushing and Sunset Park, <a href="https://web.archive.org/web/20260704120000/https://robertsietsema.substack.com/p/where-to-find-the-regional-noodles">regional Chinese noodle shops</a> differentiate texture through precise dough chemistry: Shaanxi biang biang relies on high-gluten belt stretching, Lanzhou lamian uses alkaline peng-ash (potassium carbonate) to cross-link gluten networks for hand-pulling, and Wuhan reganmian coats parboiled alkaline strands in sesame paste emulsions.</blockquote>]]></content>
  </entry>
  <entry>
    <title>InSAR Coastal Subsidence Calibration</title>
    <link href="https://gregor.fyi/2026/06/12/insar-coastal-subsidence-calibration/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/06/12/insar-coastal-subsidence-calibration/</id>
    <published>2026-06-12T12:00:00Z</published>
    <updated>2026-06-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Geodesists re-evaluated <a href="https://web.archive.org/web/20260612120000/https://www.science.org/content/article/satellite-maps-sinking-coastlines-come-under-scrutiny">interferometric synthetic aperture radar (InSAR) coastal subsidence maps</a>, finding that localized groundwater pumping, building weight, and vegetation moisture variations cause satellite radars to overestimate long-term relative sea-level rise by up to 50%, mandating ground-based GPS validation.</blockquote>]]></content>
  </entry>
  <entry>
    <title>SRAM Infrared Remanence</title>
    <link href="https://gregor.fyi/2026/06/08/sram-infrared-remanence/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/06/08/sram-infrared-remanence/</id>
    <published>2026-06-08T12:00:00Z</published>
    <updated>2026-06-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260608120000/https://www.bunniestudios.com/blog/2026/on-reading-srams-in-ir-images-and-establishing-bounds-on-trust/">Infrared backside imaging</a> through silicon substrates can read unpowered SRAM cell states via physical dopant charge remanence, bypassing software memory zeroing and requiring secure hardware to implement active charge-neutralizing circuits prior to power-down.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Simulated Annealing in Software</title>
    <link href="https://gregor.fyi/2026/06/04/simulated-annealing-in-software/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/06/04/simulated-annealing-in-software/</id>
    <published>2026-06-04T12:00:00Z</published>
    <updated>2026-06-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Software architecture and organizational technical debt behave like <a href="https://web.archive.org/web/20260604120000/https://apenwarr.ca/log/20260531">physical simulated annealing</a>: engineering systems require periodic high-temperature perturbations (breaking API refactors, rewrites) to escape local optima, whereas premature procedural cooling freezes teams into brittle legacy codebases.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Automated Hypothesis Generation</title>
    <link href="https://gregor.fyi/2026/05/16/automated-hypothesis-generation/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/05/16/automated-hypothesis-generation/</id>
    <published>2026-05-16T12:00:00Z</published>
    <updated>2026-05-16T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Autonomous scientific systems like <a href="https://web.archive.org/web/20260516120000/https://www.science.org/content/blog-post/evaluating-co-scientist-new-ai-science-system">Google DeepMind's Co-Scientist</a> generate hypotheses and design wet-lab execution protocols, but remain fundamentally throttled by wet-lab robotic execution limits and pervasive negative publication bias in training corpora.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Blood-Brain Barrier Statin Lipophilicity</title>
    <link href="https://gregor.fyi/2026/05/12/blood-brain-barrier-statin-lipophilicity/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/05/12/blood-brain-barrier-statin-lipophilicity/</id>
    <published>2026-05-12T12:00:00Z</published>
    <updated>2026-05-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Although the brain requires abundant cholesterol for myelin sheaths, central nervous system cholesterol is synthesized locally by astrocytes rather than imported from blood. Because <a href="https://web.archive.org/web/20260512120000/https://www.cremieux.xyz/p/why-dont-statins-cause-dementia">hydrophilic statins (rosuvastatin, pravastatin)</a> cannot cross the blood-brain barrier, they lower systemic cardiovascular risk without depleting cerebral cholesterol.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Predictive Model Drift</title>
    <link href="https://gregor.fyi/2026/05/08/predictive-model-drift/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/05/08/predictive-model-drift/</id>
    <published>2026-05-08T12:00:00Z</published>
    <updated>2026-05-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Foundation model evaluations suffer from acute structural contamination: <a href="https://web.archive.org/web/20260508120000/https://www.astralcodexten.com/p/nostalgebraists-hydrogen-jukeboxes">synthetic data feedback loops</a>, benchmark memorization, and preference tuning produce brittle models that ace public benchmarks while failing on out-of-distribution reasoning.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Carboniferous Insect Gigantism</title>
    <link href="https://gregor.fyi/2026/04/08/carboniferous-insect-gigantism/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/04/08/carboniferous-insect-gigantism/</id>
    <published>2026-04-08T12:00:00Z</published>
    <updated>2026-04-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Paleobiologists explain why <a href="https://web.archive.org/web/20260408120000/https://nautil.us/what-happened-to-the-ancient-bug-giants-of-300-million-years-ago-1279562">giant Carboniferous insects</a> (such as 2.5-foot Meganeura dragonflies) disappeared: atmospheric oxygen levels peaked at 35% during the Carboniferous, allowing passive tracheal diffusion to oxygenate massive bodies. When oxygen levels crashed in the Permian and predatory birds subsequently evolved, tracheal diffusion permanently constrained insect body size.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Macroeconomic AI Elasticities</title>
    <link href="https://gregor.fyi/2026/04/04/macroeconomic-ai-elasticities/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/04/04/macroeconomic-ai-elasticities/</id>
    <published>2026-04-04T12:00:00Z</published>
    <updated>2026-04-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260404120000/https://marginalrevolution.com/marginalrevolution/2026/04/economists-on-ai-and-economic-growth-and-employment.html">Comparative surveys of macroeconomists versus AI researchers</a> reveal a fundamental divergence on long-term growth: economists project modest 1–2% annual TFP gains constrained by Baumol's cost disease and physical capital, whereas computer scientists forecast transformative double-digit GDP acceleration.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Software Review Latency Compounding</title>
    <link href="https://gregor.fyi/2026/03/16/software-review-latency-compounding/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/03/16/software-review-latency-compounding/</id>
    <published>2026-03-16T12:00:00Z</published>
    <updated>2026-03-16T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>In large software organizations, every <a href="https://web.archive.org/web/20260316120000/https://apenwarr.ca/log/20260316">mandatory sequential approval gate</a> compounds delivery latency by roughly 10x: asynchronous pull-request queues, branch rebase conflicts, and cross-team sign-off meetings dominate wall-clock engineering time relative to small-batch continuous integration.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Jevons Rebound in Compute</title>
    <link href="https://gregor.fyi/2026/03/12/jevons-rebound-in-compute/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/03/12/jevons-rebound-in-compute/</id>
    <published>2026-03-12T12:00:00Z</published>
    <updated>2026-03-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260312120000/https://snippet.finance/energy-efficiency/">Jevons' Paradox</a> governs AI hardware efficiency: order-of-magnitude gains in GPU FLOPs per watt have not reduced aggregate datacenter power draw; instead, plunging marginal inference costs unlock elastic demand for continuous background agents, test-time search, and synthetic data generation.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Danish Mortgage Balance Principle</title>
    <link href="https://gregor.fyi/2026/03/08/danish-mortgage-balance-principle/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/03/08/danish-mortgage-balance-principle/</id>
    <published>2026-03-08T12:00:00Z</published>
    <updated>2026-03-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260308120000/https://marginalrevolution.com/marginalrevolution/2026/03/a-danish-fix-for-u-s-mortgage-lock-in.html">Denmark’s 'balance principle' mortgage system</a> eliminates the housing lock-in penalty: when interest rates rise, homeowners can buy back their mortgage debt in the open market at prevailing discounted prices, preserving labor mobility.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Jovian Lightning Energetics</title>
    <link href="https://gregor.fyi/2026/03/04/jovian-lightning-energetics/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/03/04/jovian-lightning-energetics/</id>
    <published>2026-03-04T12:00:00Z</published>
    <updated>2026-03-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>NASA's Juno spacecraft mapped high-energy atmospheric discharges in Jupiter's upper atmosphere, revealing that <a href="https://web.archive.org/web/20260304120000/https://www.science.org/content/article/jupiter-s-weather-forecast-cloudy-chance-nukes">Jovian 'shallow lightning' discharges</a> in ammonia-water clouds with energetic yields exceeding hundreds of megatons—orders of magnitude more violent than terrestrial superbolts.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Vertical Energy-Compute Scaling</title>
    <link href="https://gregor.fyi/2026/02/20/vertical-energy-compute-scaling/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/02/20/vertical-energy-compute-scaling/</id>
    <published>2026-02-20T12:00:00Z</published>
    <updated>2026-02-20T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Multi-gigawatt frontier AI training clusters are hitting multi-year <a href="https://web.archive.org/web/20260220120000/https://www.nextbigfuture.com/2026/02/ai-winner-will-scale-ai-energy-and-compute-on-earth-and-in-space-why-xai-spacex-and-tesla-win.html">terrestrial electrical grid interconnection queues</a>, shifting competitive advantage toward vertically integrated operators that co-locate captive gas turbines, utility solar-plus-storage, and eventual orbital photovoltaic compute.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Clinical Biological Baselines</title>
    <link href="https://gregor.fyi/2026/02/16/clinical-biological-baselines/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/02/16/clinical-biological-baselines/</id>
    <published>2026-02-16T12:00:00Z</published>
    <updated>2026-02-16T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Longitudinal clinical reference ranges suffer from <a href="https://web.archive.org/web/20260216120000/https://www.asimov.press/p/baseline-drift">generational baseline drift</a>: shifting metabolic diets, endocrine-disrupting chemical exposures, and sedentary environments have altered population distributions for serum testosterone, basal body temperature, and puberty onset relative to mid-20th-century textbook norms.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Open-Source Satellite Mortuary Tracking</title>
    <link href="https://gregor.fyi/2026/02/08/open-source-satellite-mortuary-tracking/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/02/08/open-source-satellite-mortuary-tracking/</id>
    <published>2026-02-08T12:00:00Z</published>
    <updated>2026-02-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Relatives of missing Russian soldiers systematically monitor <a href="https://web.archive.org/web/20260208120000/https://www.newyorker.com/news/the-lede/the-russians-turning-to-google-maps-in-search-of-missing-soldiers">high-cadence commercial satellite updates</a> on Google Maps and Yandex to measure newly excavated trench rows in military cemeteries near Rostov-on-Don, turning consumer mapping platforms into open-source casualty ledgers.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Cognitive Variance Asymmetries</title>
    <link href="https://gregor.fyi/2026/02/04/cognitive-variance-asymmetries/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/02/04/cognitive-variance-asymmetries/</id>
    <published>2026-02-04T12:00:00Z</published>
    <updated>2026-02-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Across large psychometric cohorts, male and female mean general intelligence (g) scores are nearly identical, but males exhibit <a href="https://web.archive.org/web/20260204120000/https://www.cremieux.xyz/p/sex-differences-in-intelligence">wider phenotypic variance</a> (~10–15% higher standard deviation) on quantitative and spatial subtests, producing asymmetric representation at both extreme tails.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Navier-Stokes Finite-Time Singularity</title>
    <link href="https://gregor.fyi/2026/01/12/navier-stokes-finite-time-singularity/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/01/12/navier-stokes-finite-time-singularity/</id>
    <published>2026-01-12T12:00:00Z</published>
    <updated>2026-01-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Mathematicians using physics-informed neural networks discovered potential <a href="https://web.archive.org/web/20260112120000/https://www.quantamagazine.org/using-ai-mathematicians-find-hidden-glitches-in-fluid-equations-20260109/">finite-time singularities (blowups)</a> in the 3D Euler and Navier-Stokes fluid equations. By guiding numerical solvers to self-similar coordinate collapses, the AI identified initial boundary conditions where vorticity concentrates infinitely, bringing researchers closer to resolving a Millennium Prize problem.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Headless Agent IDEs</title>
    <link href="https://gregor.fyi/2026/01/08/headless-agent-ides/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/01/08/headless-agent-ides/</id>
    <published>2026-01-08T12:00:00Z</published>
    <updated>2026-01-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>As autonomous coding agents shift software engineering from manual keystroke editing to multi-agent supervision, traditional single-developer graphical IDEs (VS Code, IntelliJ) give way to <a href="https://web.archive.org/web/20260108120000/https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04">headless, terminal-driven orchestration runtimes</a> that continuously synthesize, test, and merge code.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Lunar Property Rights</title>
    <link href="https://gregor.fyi/2026/01/04/lunar-property-rights/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2026/01/04/lunar-property-rights/</id>
    <published>2026-01-04T12:00:00Z</published>
    <updated>2026-01-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20260104120000/https://www.astralcodexten.com/p/you-have-only-x-years-to-escape-permanent">Outer Space Treaty prohibitions</a> against sovereign territorial appropriation create acute legal ambiguity: absent formal property rights, first-mover mining operations risk establishing de facto exclusionary safety zones around lunar ice deposits unless bilateral market frameworks are codified.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Neuroinflammation Incretins</title>
    <link href="https://gregor.fyi/2025/12/08/neuroinflammation-incretins/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/12/08/neuroinflammation-incretins/</id>
    <published>2025-12-08T12:00:00Z</published>
    <updated>2025-12-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Phase III clinical trials of <a href="https://web.archive.org/web/20251208120000/https://www.science.org/content/blog-post/another-glp-1-alzheimer-s-trial">GLP-1 receptor agonists in early Alzheimer's disease</a> confirm substantial reductions in microglial neuroinflammation and neurofibrillary tangle progression despite modest primary cognitive score deltas, establishing neuroprotective metabolic activity in the brain.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Empirical Learning Curve Limits</title>
    <link href="https://gregor.fyi/2025/12/04/empirical-learning-curve-limits/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/12/04/empirical-learning-curve-limits/</id>
    <published>2025-12-04T12:00:00Z</published>
    <updated>2025-12-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Across 62 technologies, <a href="https://web.archive.org/web/20251204120000/https://www.construction-physics.com/p/how-accurate-are-learning-curves">Wright’s Law learning curves</a> hold reliably (15–25% cost drop per cumulative doubling) only for modular, factory-repeatable goods like solar PV and lithium cells, whereas bespoke on-site civil construction and nuclear plants exhibit flat or negative learning curves.</blockquote>]]></content>
  </entry>
  <entry>
    <title>National Capital Accounting</title>
    <link href="https://gregor.fyi/2025/11/16/national-capital-accounting/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/11/16/national-capital-accounting/</id>
    <published>2025-11-16T12:00:00Z</published>
    <updated>2025-11-16T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Macroeconomists argue for replacing gross domestic product (GDP) with <a href="https://web.archive.org/web/20251116120000/https://www.nature.com/articles/d41586-025-03721-1">comprehensive wealth accounting</a>: national economic metrics must measure capital stock depletion, environmental degradation, and human cognitive capital depreciation, rather than treating defensive disaster expenditures as positive economic growth.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Commercial Aircraft Cycle Times</title>
    <link href="https://gregor.fyi/2025/11/12/commercial-aircraft-cycle-times/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/11/12/commercial-aircraft-cycle-times/</id>
    <published>2025-11-12T12:00:00Z</published>
    <updated>2025-11-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20251112120000/https://jordanwtaylor2.substack.com/p/the-aerospace-crystal-ball">Clean-sheet commercial airliner development cycles</a> have stretched from 4 years in the 1950s to over 15 years today as FAA/EASA certification overhead, tiered global sub-supplier fragmentation, and zero-defect turbofan integration compound exponentially.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Datacenter Baseload Energy</title>
    <link href="https://gregor.fyi/2025/11/04/datacenter-baseload-energy/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/11/04/datacenter-baseload-energy/</id>
    <published>2025-11-04T12:00:00Z</published>
    <updated>2025-11-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20251104120000/https://marginalrevolution.com/marginalrevolution/2025/11/my-excellent-conversation-with-sam-altman.html">Gigawatt-scale electrical generation</a> and datacenter thermal dissipation have superseded algorithmic theory as the primary pacing constraints for frontier AI, compelling direct integration with advanced nuclear and geothermal baseload infrastructure.</blockquote>

<p><a id="2026-10-10" href="#2026-10-10">2026-10-10</a>: Nuclear Capacity Uprates<blockquote>Google signed a <a href="https://web.archive.org/web/20251104120000/https://neutronbytes.com/2026/10/09/googles-20-year-4-3-billion-deal-with-constellation-to-uprate-11-reactors/">20-year, $4.3 billion power purchase agreement</a> with Constellation Energy to fund capacity uprates across 11 existing nuclear reactors in Illinois, Pennsylvania, and New Jersey. Modernizing turbines, steam generators, and digital instrumentation unlocks 890 MW of firm, zero-carbon baseload for datacenters—adding the electrical equivalent of an entire new reactor to the PJM grid while sidestepping the decade-long regulatory lead times and capex of greenfield nuclear construction.</blockquote></p>
<a id="2026-11-12" href="#2026-11-12">2026-11-12</a>: Retired airliner jet cores repurposed for onsite power<blockquote>Because utility interconnections and heavy-frame industrial gas turbines are backlogged for years, datacenter operators are buying decommissioned Boeing 737, 747, and Airbus A320 engines and competing directly with airlines for overhaul-ready cores. FTAI Aviation rebuilds <a href="https://web.archive.org/web/20261108132823/https://x.com/Fahadnaimb/status/2119421779978625144">CFM56 cores</a> into 23 MW continuous (25 MW peak) stationary generators—with over 100 units contracted for 2027–2028 and a $1.465 billion mobile-turbine order from a cloud provider via its Jereh Group joint venture—while xAI deployed 8 converted GE CF6 cores at its Mississippi Colossus 2 site.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Compute Overhang Asymmetries</title>
    <link href="https://gregor.fyi/2025/10/20/compute-overhang-asymmetries/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/10/20/compute-overhang-asymmetries/</id>
    <published>2025-10-20T12:00:00Z</published>
    <updated>2025-10-20T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>The <a href="https://web.archive.org/web/20251020120000/https://marginalrevolution.com/marginalrevolution/2025/10/jack-clark-on-ai.html">concentrated geographic distribution of high-end AI compute</a> clusters mirrors nuclear centrifuge facilities, creating aggressive international supply chain surveillance incentives as compute sovereignty emerges as a primary pillar of national defense.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Transformers Out-of-Distribution Limits</title>
    <link href="https://gregor.fyi/2025/10/16/transformers-out-of-distribution-limits/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/10/16/transformers-out-of-distribution-limits/</id>
    <published>2025-10-16T12:00:00Z</published>
    <updated>2025-10-16T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20251016120000/http://backreaction.blogspot.com/2025/10/current-ai-models-have-3-unfixable.html">Autoregressive transformers</a> face 3 structural boundaries: reliance on static training distributions that struggle with out-of-distribution structural extrapolations, absence of built-in symbolic ground-truth verification, and quadratic attention compute costs over long contexts.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Hyperspectral Aerial Metagenomics</title>
    <link href="https://gregor.fyi/2025/10/12/hyperspectral-aerial-metagenomics/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/10/12/hyperspectral-aerial-metagenomics/</id>
    <published>2025-10-12T12:00:00Z</published>
    <updated>2025-10-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Aerial and satellite hyperspectral imaging detects plant microbiome shifts and pathogen colonization in <a href="https://web.archive.org/web/20251012120000/https://www.asimov.press/p/hyperspectral">shortwave infrared reflectance signatures</a> days before visible leaf wilting, enabling predictive agricultural intervention.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Su Filindeu Preservation</title>
    <link href="https://gregor.fyi/2025/10/08/su-filindeu-preservation/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/10/08/su-filindeu-preservation/</id>
    <published>2025-10-08T12:00:00Z</published>
    <updated>2025-10-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20251008120000/https://www.bbc.com/travel/article/20161014-the-worlds-rarest-pasta">Su Filindeu</a> ('the threads of God'), an ancient Sardinian pasta crafted in Nuoro by repeatedly pulling and folding semolina dough into 256 gossamer-thin parallel strands laid across 3 diagonal layers, relies on tactile calibration of ambient humidity, gluten elasticity, and saline hydration.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Fly-By-Wire Retrofits</title>
    <link href="https://gregor.fyi/2025/10/04/fly-by-wire-retrofits/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/10/04/fly-by-wire-retrofits/</id>
    <published>2025-10-04T12:00:00Z</published>
    <updated>2025-10-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Sikorsky and DARPA successfully converted a standard UH-60 Black Hawk helicopter into a fully autonomous uncrewed aircraft in 10 months using the <a href="https://web.archive.org/web/20251004120000/https://www.nextbigfuture.com/2025/10/black-hawk-converted-into-an-unmanned-drone-in-10-months.html">MATRIX autonomy kit</a>. The retrofitted fly-by-wire system executes low-altitude tactical resupply and casualty evacuations in contested environments without human intervention.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Life In Our Phage World: Viable AI-mutated phages</title>
    <link href="https://gregor.fyi/2020/01/05/life-in-our-phage-world/#2025-09-18" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2020/01/05/life-in-our-phage-world/#2025-09-18</id>
    <published>2025-09-18T12:00:00Z</published>
    <updated>2025-09-18T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2020/01/05/life-in-our-phage-world/" href="https://gregor.fyi/2020/01/05/life-in-our-phage-world/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2020/01/05/life-in-our-phage-world/">Life In Our Phage World</a>:</em></p>
<blockquote>the researchers mixed all 16 AI-generated phages with ΦX174 and then threw them into a tube with E. coli cells. Because the phages were forced to compete for the same host cells, the variants that reproduced fastest would dominate. By sequencing the phages over time, the researchers could track which phages were gaining ground and which were falling behind. Several of the AI phages consistently outperformed wild ΦX174, with one variant (called Evo-Φ69) increasing to 65x its starting level.

Ultimately, these <a href="https://web.archive.org/web/20250918194741/https://press.asimov.com/articles/ai-phages">16 AI-generated phages</a> were not only viable; in many cases, they were more infectious than wildtype ΦX174 despite carrying major genome alterations that a human would be unlikely to rationally design.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Wow! Signal and Magnetar Masers</title>
    <link href="https://gregor.fyi/2025/09/12/wow-signal-magnetar-masers/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/09/12/wow-signal-magnetar-masers/</id>
    <published>2025-09-12T12:00:00Z</published>
    <updated>2025-09-12T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Astrophysical re-evaluation of the <a href="https://web.archive.org/web/20250912120000/https://avi-loeb.medium.com/was-the-wow-signal-emitted-from-3i-atlas-d18d4f0d1f1e">1977 narrowband 1420 MHz Wow! signal</a> indicates the burst matches a natural interstellar neutral-hydrogen (HI) cloud maser flare pumped by a transient magnetar or soft gamma repeater outburst.</blockquote>]]></content>
  </entry>
  <entry>
    <title>0.001% Mathlib: Gauss</title>
    <link href="https://gregor.fyi/2021/07/28/0001-mathlib/#2025-09-11" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2021/07/28/0001-mathlib/#2025-09-11</id>
    <published>2025-09-11T12:00:00Z</published>
    <updated>2025-09-11T12:00:00Z</updated>
    <thr:in-reply-to ref="https://gregor.fyi/2021/07/28/0001-mathlib/" href="https://gregor.fyi/2021/07/28/0001-mathlib/" type="text/html"/>
    <content type="html"><![CDATA[<p class="update-context"><em>Update to <a href="https://gregor.fyi/2021/07/28/0001-mathlib/">0.001% Mathlib</a>:</em></p>
<blockquote>Using Gauss <a href="https://web.archive.org/web/20250911214133/https://www.math.inc/gauss">we produced ~25K lines of Lean code</a>, comprising 1K theorems and definitions. Formal proofs of this scale have historically been major milestones, often the culmination of multi-year efforts. The largest singular formalization projects in history — career-defining efforts, which can span more than a decade — are only an order of magnitude larger at up to 500K LOC. Lean’s standard mathematical library, Mathlib, is 1 OOM beyond that, at 2M LOC, comprising 350K Lean theorems and definitions, and developed by 600 human contributors over 8 years. With further algorithmic improvements, we aim to increase the sum total of formal code by 2-3 OOM in the coming 12 months. This will serve as the training ground for a <a href="https://web.archive.org/web/20250911214133mp_/https://www.math.inc/vision">new paradigm</a> — verified superintelligence and the machine polymaths that will power it.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Modular Diner Preservation</title>
    <link href="https://gregor.fyi/2025/09/08/modular-diner-preservation/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/09/08/modular-diner-preservation/</id>
    <published>2025-09-08T12:00:00Z</published>
    <updated>2025-09-08T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote><a href="https://web.archive.org/web/20250908120000/https://robertsietsema.substack.com/p/my-favorite-nyc-diners-right-now">Mid-century stainless-steel diners</a> in New York City were factory-prefabricated in New Jersey and delivered via flatbed truck as modular industrial units, pairing streamlined Art Deco metallurgy with high-turnover vernacular hospitality.</blockquote>]]></content>
  </entry>
  <entry>
    <title>Rogue Wave Physics</title>
    <link href="https://gregor.fyi/2025/09/04/rogue-wave-physics/" rel="alternate" type="text/html"/>
    <id>https://gregor.fyi/2025/09/04/rogue-wave-physics/</id>
    <published>2025-09-04T12:00:00Z</published>
    <updated>2025-09-04T12:00:00Z</updated>
    <content type="html"><![CDATA[<blockquote>Oceanic rogue waves form via <a href="https://web.archive.org/web/20250904120000/https://nautil.us/rogue-wave-mystery-solved-1236928/">Benjamin-Feir modulational instability</a>: non-linear phase coupling causes ordinary wave trains to spontaneously siphon energy from adjacent waves, concentrating amplitude into an isolated 30-meter wall of water within seconds without seabed bathymetric focusing.</blockquote>]]></content>
  </entry>
</feed>