Rosa Del Mar

Daily Brief

Issue 31 2026-01-31

Adjacent Ai Ops Products Claims From Host

  • Finn is described as an AI customer service agent that can automatically resolve up to 93% of customer queries.
  • Wix has not yet cracked TikTok and sees large upside in TikTok and LinkedIn despite imperfect targeting on LinkedIn and past disappointment with sports endorsements.
  • Omer Shai rejects LTV as a primary marketing metric because it is unknowable and too slow for fast decisions, and prefers TROI (time to return on investment).

Cta Dispersion Drivers Speed Universe And Volatility Targeting

  • CTA return dispersion was described as being explainable by implementation choices including trading speed, market set, and volatility-adjusted position sizing, even when return correlations are high.
  • It was asserted that 2025 trend performance was driven by a very narrow band of assets, while early 2026 shows broader trends across metals and other commodities.
  • Central banks have been net buyers of gold every year since 2011, after being net sellers each year from 2000 to 2009.

Version Lineage And Preservation Targets

  • Crimsonland’s release lineage includes a 2002 freeware prototype series, a 2003 shareware v1.8–v1.9 line, and a GOG “classic” build v1.9.93 (Feb 2011) that was later bundled as a bonus alongside the 2014 remaster.
  • A Ghidra-driven workflow maintains a name_map.json to iteratively rename and type functions based on evidence such as strings, call patterns, and struct sizes, allowing improved types to propagate across decompilations.
  • Crimsonland assets are stored in custom PAQ archives with magic 'paq\0' and entries consisting of filename, size, and payload, using Windows-style backslash paths.

Collective Efficacy Framing

  • Matt Webb published a piece arguing that people can "just do things" to improve their communities.
  • A community can bootstrap a shared public-good project by organizing collectively, producing needed infrastructure, securing subsidies, and redistributing costs to include lower-income participants.
  • Small collective projects can escalate into sustained political engagement, including contacting representatives and tracking legislation to embed the change into building requirements.

Training Cost/Time Baselines For Gpt-2-Level Capability

  • In 2019, GPT-2 training reportedly used 32 TPU v3 chips for 168 hours at about $8 per TPUv3-hour, totaling roughly $43,000.
  • A roughly 600× reduction over seven years (about 2.5× per year) in the cost to train a GPT-2-level model is claimed based on comparing the GPT-2 cost baseline to the nanochat cost estimate.
  • With recent improvements merged into nanochat (many originating from modded-nanogpt), a higher CORE score than GPT-2 can reportedly be reached in about 3.04 hours for roughly $73 on a single 8xH100 node.