Mission Control
OP009 The Signal Stack 2026-04-28
← EP008 | Mission | EP010 →

OPERATOR LOG · OP009

The Signal Stack

Systems that compound.

2026-04-28

  • SPACEX + CURSOR
  • UTAH SMRs
  • NICK SHIRLEY
  • SPLC
  • WHCD SHOOTING
  • SAMSUNG TEXAS FAB
  • AI HARNESSES
  • BERT
  • QWEN LOCAL
  • STRATUM ENGINE
  • PERPLEXITY

MARKET INTELLIGENCE

Price Action

DATA GRID

Key Numbers

🏢 Samsung Taylor fab size 4.85M m² Taylor, TX — operational late 2026
👨‍💼 Employees relocating ~1,000 Samsung Austin Semi, first wave
⚡ Utah SMR target 4 GW Holtec — Mountain West plan
🚀 Ward250 output 5 MW Airlifted · Jul 4 2026 first power
💬 Nick Shirley views 135M X platform · MN daycare claims
🤖 Qwen 14B HumanEval ~50% Local · Ollama / LM Studio
☁️ Sonnet 4.6 HumanEval 97.6% Cloud · Cursor / Composer
🏛️ WHCD incident APR 26 2026 Trump evacuated · shots outside Hilton

FIELD REPORT

Current Events

PHYSICAL LAYER

Hardware

🏢 Samsung Taylor

Samsung moves 1,000 staff to Taylor, Texas fab. Campus spans 4.85M m², operational late 2026. Part of US semiconductor reshoring push accelerated by CHIPS Act.

⚡ Utah SMR Pivot

After NuScale/UAMPS collapse, Utah pivots to Holtec SMR-300. Target: up to 10 reactors, 4 GW across the Mountain West. Positions Utah as nuclear energy hub.

🚀 Ward250 Microreactor

5 MW reactor airlifted to San Rafael Energy Lab, Utah. Small enough for a military cargo plane. First power target: July 4, 2026.

MODEL LAYER

AI Field

An AI harness is a wrapper around a model. Input goes in. Rules enforced. Output comes out. The simplest harness is a system prompt + an API call — about 20 lines of Python.

BERT (Google, 2018) reads text in both directions simultaneously. It masks words and predicts them from surrounding context. Example: "My [MASK] is cute, he's playing." → BERT predicts "dog." This is what powers Google Search's semantic understanding layer.

Simple build: pip install transformers → load bert-base-uncased → .encode() → classify. Complex build: fine-tune on domain data, multi-head attention layers, hardware-optimized inference. The difference is 20 lines of code vs. 200,000.

EVALUATION

Local Models

ModelHumanEvalParamsRuns where
Claude Sonnet 4.697.6%APICursor / Perplexity
DeepSeek R197.4%APICloud
Qwen 3 480B92.7%APICloud
Qwen 3.5 9B91.0%9BOllama / Local
Qwen Coder 14B~50%14BOllama / LM Studio

Running Qwen 14B locally through Cline against a repo with our full test stack — ruff, pytest, Playwright, real Postgres — exposes context limits fast. Sonnet in Cursor has full repo awareness. Local models don't. SpaceX chose Cursor. That's the answer.

SHIP LOG

Stratum Engine

PRODUCT LOG

Perplexity

SIGNAL STRENGTH

Live Signal

📡 SpaceX + Cursor

Highest-profile enterprise validation of AI-assisted coding. Sonnet-powered Cursor adopted across SpaceX engineering. The local vs. cloud debate is over for elite teams.

⚡ Energy Reshoring

Samsung, Holtec, and Valar Atomics represent a $100B+ hardware migration to the Mountain West and Texas. CHIPS Act + tariff policy accelerating timeline.

🤖 AI Product Velocity

Perplexity shipped 5 major features in 60 days: Personal Computer, CFO with Plaid, Deep Research deliverables, Comet iOS, Model Council. Speed compounds.

📈 Stratum Engine

9 shipped features, 3 in progress. CodeRabbit AI reviews, nightly auto-deploy, Grafana + Loki logging, Sentry tracking, Playwright E2E. Infrastructure as product.

INTELLIGENCE SOURCES

Sources

    OP009 · The Signal Stack · 2026-04-28
    Arrow keys navigate Descript