MTS // Drops

A goal in a knockout match creates a near-instant, planet-wide, synchronized spike — everyone refreshing, commenting, and betting in the same two seconds — then it craters. That vertical-spike-then-crater shape is exactly what serverless Postgres is built to absorb. Five live sections: broadcast wall (840k concurrent connections, one Postgres row), EKG autoscale with saturation gap and scale-to-zero, Crypto.com prediction markets on the same timeline, rotating globe following the sun, and 130 years of database history.

It's listening on every channel you own — that's the feature and the problem. A fake-macOS desktop whose one open app is a Safari-style browser: encyclopedia article, agent directory, forum thread, security advisory, RFC, and a working demo inbox, with cat, a resident terminal agent, navigating the tabs and answering the mail. OpenClaw, Hermes, the closed wing, the fork belt, and eighty years of lineage.

Boot into Sound System 11 — a black-and-white Windows-XP desktop — and try three ElevenLabs audio models live: Music V2 composes from a prompt, Voice Changer reskins your voice, and Scribe transcribes whatever you say. MTS × ElevenLabs.

In a single week, AI systems formally proved theorems and disproved an eighty-year-old Erdős conjecture about unit distances in the plane — twice, by two different labs. How AlphaProof, AlphaGeometry 2, and Claude Mythos work, what they proved against Lean's machine checker, and why a compiled proof collapses the gap between convincing and correct.

Eli Lilly's VERVE-102 is a one-shot, base-editing infusion that permanently disables the PCSK9 gene in liver cells. In the Phase 1b Heart-2 trial, a single dose lowered PCSK9 protein by up to 88 percent and cut LDL cholesterol by 62 percent on average at the top dose, with durable effects out to 18 months. A walk through the mechanism, the dose-response, the safety read, and what's still uncertain.

What world models are, how they work, what's next. A learned internal representation of an environment — compress observations into latent state, predict futures in imagination, plan without acting. A four-step playbook for generating a world frame-by-frame with a real world model (Decart's Oasis), the math behind DreamerV3/Genie/Cosmos/DreamZero, and recent signals from NVIDIA's DreamZero and Odyssey's Agora-1.

YC vocabulary across the Winter 2025 → Winter 2026 batches. Every one-liner and long description for ~680 companies, indexed for what's surging (agentic, claude code), collapsing (voice, compliance), and conspicuously absent (world model, MCP server) in early-stage founder pitches.

1,891 open roles across 122 recently-funded accelerator startups (YC W25+, a16z Speedrun, AI Grant). Median comp, equity rates, remote policy, and the must-have skill list — extracted from every job description by parallel Haiku agents.

A guided gallery through the foundational papers that built modern AI — from the 1943 McCulloch–Pitts neuron to DeepSeek-R1. Eleven eras, each with its own canvas, its own visual treatment, its own formula, and the three or four papers that actually mattered.

US-Japan venture capital intelligence — every Tier-1 arrival, deal, and macro signal flowing between Sand Hill Road and Tokyo. 74 records tracking the foreign capital build-out around Sakana, Rapidus, and SoftBank's reverse flow into OpenAI, with ¥/$ dual tracking and per-entity capital rosters.

Every 13F-HR filing from Situational Awareness LP — Leopold Aschenbrenner's hedge fund, built around an explicit AGI thesis. Quarter-over-quarter holdings, entries, exits, and per-position trade histories with live marks and option-leg detection.

Disclosed trades by members of Congress and their families, with the day-of, week-ahead, and 30-day price moves alongside each filing. Placeholder copy until the data freeze lands.

Every 13F-HR filing since mid-2023 from 45 curated funds, joined to Form 4 insider flow and adjusted-close prices, with rotation and consensus calls per quarter.

A twelve-step walkthrough of why AI memory is the next bottleneck — and what the production stack actually does about it today.

Eleven beats — a black-box app, a dependency tree, a maintainer at 11:47pm, a stealer harvest, an attacker session, a 760-org blast radius. Every 2025–2026 marquee breach fits the same chain; a receipts page ties each beat to 19 real named incidents.

Anthropic (May 11) and OpenAI (May 12) just declared every unauthorized secondary transfer void. Nine self-paced steps trace 1,000 ex-employee shares through three layers of SPVs — to 47 retail-accredited LPs at a +39% cumulative premium — before the announcement that hollowed the whole structure out.

A single gigawatt-scale AI data center costs ~$50B all-in — $25–50B of Nvidia silicon plus ~$20B of building, power, and cooling. AI bulls call it a 2-year payback; commercial-real-estate pros call it 10. Both numbers are arithmetically correct. The public filings show the upside staying with the AI companies while the downside has been pre-allocated to people who never saw the cap table.

An editorial frontend over war.gov's first UFO/UAP release — 161 declassified records from the FBI, NASA, and the Departments of War and State, geocoded onto a globe and indexed for full-text search.

Semantic search across all 199 trial-admitted exhibits in OpenAI v. Musk — emails, journal pages, contracts, screenshots — rendered in form-factor viewers that match the original document.

The viral “500 mL per query” figure conflated per-conversation with per-prompt on stale 2022 hardware. A bottom-up audit lands at ~3 mL — about a sip — and a single hamburger is two million prompts.

An interactive walkthrough of the trial — the first MTS subdomained visualization.