faye issues
w40
-
Uppsala group proposes using datacenter fiber itself as memory: stream LLM weights once into a recirculating fiber ring and let inference nodes tap the light passively instead of each holding its own HBM copy. a 1,000 km loop of 19-core multicore fiber holds 128 GB in flight, enough for a 70B model. arXiv
-
Ai2's Nature paper on "byteification" retrofits subword models to byte-level on a fraction of the pretraining budget, with BOLMO 7B gaining 16.5% on STEM over BLT 7B. Ai2
-
a UW + Meta FAIR paper converts token logits to byte logits, enabling cross-tokenizer distillation — distilled byte models beat distilled token models on a fraction of the data. arXiv
-
Keith Robison on Sculpta Bio's BOB-Seq preprint: bioorthogonal barcoding tags RNA inside the cell for pooled sequencing, with 99% barcode accuracy and uniform transcript-body coverage. OmicsOmics
-
Ginkgo's September whitepaper lays out an enzyme engineering platform combining AI with physics-based design, running on Nebula, the 100+ robotic cart autonomous lab. headline numbers are a 6-billion-gene discovery library and a closed-loop run where an LLM designed cell-free expression experiments that lifted protein titer 27% over state of the art. the pitch is moving enzyme engineering from bespoke screening to predictable design, framed as a US biomanufacturing priority. Ginkgo Bioworks
-
Google's Suncatcher paper is now peer-reviewed in Joule. the design calls for 81-satellite clusters in one-kilometer formations at 650 kilometers altitude, linked by lasers, with Trillium TPUs surviving a five-year radiation dose in ground tests. the gate is launch cost. below $200 per kilogram by the mid-2030s, which by Google's estimate needs 1,800 Starship launches. heat is the open problem. Arxiv
-
Hormuz flows are back to roughly 80% of prewar levels via the strait and bypass routes, while Iran's own exports sit at zero under the blockade. with the chokehold broken, the risk shifts to Iranian military escalation to regain leverage.
-
how the recovery actually works, per Kpler — 16.5 mbd left the region in September, matching prewar ex-Iran. 40% bypasses Hormuz entirely through Saudi and UAE pipelines, up from 17%. of what still crosses, over 70% changes tankers offshore at Fujairah or Sohar, AIS dark, on a shuttle fleet of 63+ VLCCs.
-
a 1h44 engineering video argues Starlink V3 is a radar constellation waiting to be switched on: ballistic missile tracking with warhead/decoy discrimination, stealth aircraft tracked by their radar shadow, and one-way attack drones imaged through their plastic shells.
-
retired Lockheed Martin Fellow Vladimir Volman, associated with NASA Glenn, designed a ground receiver that senses targets from Starlink downlinks with no reference signal and nothing from SpaceX.
faye
long horizon, post-quant investments in magic and/or technology, steeped in chaos, dependent on human intuition & faith.
philosophy
greetings investor(s), superintelligence is big. it's going to get bigger. but machine intelligence, like most tech once-novelties, will be increasingly commoditized and must flow through literal human vessels, in some way/shape/form. and the ultimate effects must benefit humans (& environment & economy). hardware & software only matter insofar as they benefit the essentially-alien-technology, "wetware," that is biology. wetware is the single most exciting frontier to invest in due to its utter nascency and the limitless benefits it promises to offer humanity & the future. so, we invest in FAYE's own internally-aligned blend "wetware+," or infrastructure & enabling technologies that become strategically important if one/more expected long-duration technological transitions compound: ai, bio, space, defense, nuclear/isotope tech, & ukrainian reconstruction. FAYE invests using a qualitative & thesis-first strategy, embracing chaos & noise. reducing/smoothing noise often means reducing possibilities. some "unknown" must necessarily remain in all models, as the ylem from which progress is reconstituted. we are O.K. being early, if the structural direction & trends remains reasonable. the past reveals a lot, but not the future.
the goal
our loose goal is to beat SPRX & the ARK funds over multi-year/multi-decade timeframes. names can bleed out for years before any payoff lands, although that's not our base case. thesis breakage or fundamentals/technicals breakage triggers a review, but human judgment gates every decision — no quantitative or rules-based stops of any kind.
AURUM NOSTRUM NON EST AURUM VULGI 🪱