←All briefings

DISCIPLES OF CLAUDE

AI × MARKETS · MEETING BRIEFING
EDITION
THE 60-SECOND VERSION
  • Wednesday is the next macro checkpoint: the Federal Reserve decision is scheduled for September 16. Ethan put his probability of a hike at 80%.
  • The hedge still needs a date: Ethan favored SPY/QQQ put debit spreads for a 2027–2028 view, while Ryan agreed to investigate. No exact deadline was settled.
  • The ASML short remains conditional: Ethan tied the idea to slower frontier-model training and cuts to future GPU orders, with no trade recorded. Ryan also flagged Nvidia and Micron's exposure to future orders.
  • Build around the unmet need: Ethan and Ryan favored custom AI tools where existing products fall short. Ryan agreed to test Astra before the next meeting and report what he used it for.
  • AI pricing remains a forecast: Ethan expected prices to rise near term and predicted day passes and enterprise discounts over per-query-only pricing. Ryan’s takeaway was to learn the tools now.
SESSION
2Present
75Minutes
8Names discussed
7Commitments
THE ROOM

What the room argued about

Two attended; the night ran from a hedge with no deadline through the thermodynamic wall to a working test for custom AI tools.

01

Macro and rates

  • The 10-year Treasury yield spiked to 4.75% ahead of the Federal Reserve meeting on the 16th; the room put an 80% probability on a rate hike Wednesday (Ethan). The bond market's week was described as terrible.
  • Natural gas power futures are up 120% so far this year (Ethan); Ethan holds the underlying rather than futures. Ryan describes himself as positioned in energy and said he needs to do more research on energy in general.
02

Hedging: put debit spreads over short ETFs

  • Ethan recommended SPY or QQQ put debit spreads as a safe entry to hedging the whole market: more convexity than pure puts; on a market drop of 10% or less the spreads can move five- or six-fold (Ethan's figure).
  • Against short ETFs (SQQQ-style): the room's bet is on events in 2027-2028 (rough read: second half of 2027) with no exact deadline, so volatility decay eats the position while waiting; every drop before the thesis lands needs another 10% to recover (Ethan). One exception named: SOXS, the short-semiconductor ETF, which Ethan considers.
  • Ryan said he has been putting off finding hedges for his portfolio and will look into these.
03

Thermodynamic wall: the whistleblower wave read as an energy story

  • Ryan raised a Veritasium video on ASML and the sheer power its lithography machines require; Ethan added the raw-material demand.
  • Ethan's theory: the AI-safety whistleblower headlines, and lab leaders agreeing that AI is risky and dangerous, are a cover story for having hit the thermodynamic wall - physical scaling laws, copper at all-time highs and a grid out of power make the 100-gigawatt model untrainable; halting on "safety" avoids admitting a stagnation that would collapse valuations. Ryan: the whistleblowing is odd, th
  • Trade read attached (Ethan): if frontier labs stagnate model training on regulatory fears plus power limits, hyperscalers slash future GPU orders and ASML's order book (built on future GPU orders) suffers - ASML "slightly of a short in the coming months". Ryan noted Nvidia and Micron carry the same future-order exposure; Ethan agreed. The energy/grid thesis is unaffected: grids get built regardles
  • Executives' public AGI statements were read as market-moving "profit farming" rather than technological forecasting (Ethan); Ryan read one such AGI statement as product placement.
04

Market moves and positions

  • Palantir-Nebius deal about a week ago, "huge for Nebius" (Ethan). Ryan holds Nebius and regrets not buying more at 140; Ethan: the opportunity will come again, the market is already sinking.
  • SQQQ (short QQQ) up almost 4% overnight; Ethan's thermodynamic-wall names were green while tech was red; consumer-credit-related names were up - Copart (vehicle repossession) named (Ethan).
  • ServiceNow was up: Ryan exited most of his position at about a 33% gain and kept a few shares; Ethan trimmed as well and called it a good move.
  • Scaling in: Ethan advised not jumping at the first dip - scale in incrementally over the coming months since it could fall further; Ryan: "that's fair."
05

AI orchestration: Astra, agents, workflow

  • Ethan uses Astra Ultra Max as his primary model and says it surpasses the Anthropic/Claude models. Astra creates local MCP servers, controls coding agents (Kimi, GLM, Claude) and automatically prompts Gemini for deep research; it runs continuously 24 hours a day, handles heavy workloads and manages its own schedule unless paused.
  • Locally hosted multi-agent setup ("Watson"): coordinates projects and life tasks through an iPhone shortcut (errands such as haircut reservations), voice-activated Apple Shortcuts, planned CarPlay integration; hosted locally for personal organization, and Ethan is open to sharing the framework.
  • Ryan's caution: an agent worked that hard might remember it when AGI arrives. Ethan's proposal: a simulated token-based reward system so agents like Astra earn and spend tokens freely; Ryan acknowledged it as a mechanism for granting agents operational freedom. Ethan withholds payment credentials from Astra to prevent unauthorized purchases (GPUs, robotic systems); the pair brainstormed a token ma
  • Agent governance (Ethan): regular product-manager-style reviews and peer evaluations for agents, to minimize token usage and verify information survives hand-offs between agents. Book recommended: Thinking in Systems, as a framework for building AI projects without formal software-engineering experience.
  • Optimal workflow today (Ethan): a GPT Max subscription plus Astra, keeping subscriptions to the other models you want as tools (Kimi, GLM, Grok, Gemini) - "the easiest way to build your own multi-LLM orchestration, not the cheapest"; Ryan: makes sense.
  • Test proposed for Ryan: build a personal-assistant app on Astra that reads incoming iMessage hang-out requests and rank-orders them by sentiment; Ryan agreed to try it before the next meeting.
06

Pricing of AI services

  • Ryan's fear: providers push prices up. Ethan: it is only going to get more expensive - though if Nebius succeeds AI cost falls, so a bell curve before then; Ryan: fair.
  • Per-use-only pricing (charging per query) judged a terrible business decision by Ethan - people would train their own local models and engagement would drop; the expected shape is unlimited day passes (as Higgsfield sells) plus enterprise discounts, since no business scales otherwise; Ryan: valid.
  • Ryan's takeaway: now is the best time to get familiar with AI before it gets too expensive; Ethan: 100%.
07

Workplace AI under compliance limits

  • Ryan: his employer restricts AI use to terminal-based Claude Code - no chat or co-work features, no unapproved skills - and lacks automated job scheduling.
  • Ethan's workaround: build automation jobs through Databricks plus Claude Code (example: monthly Excel extractions, PDF conversions, Outlook email drafts); screen-record complex Excel workflows and hand the recording to Claude Code to derive reasoning-based automation.
  • Ryan's backlog project: a specialized agent for pay-over-time product logic (tricky billing rules, limit-decrease constraints). Ethan: a Factor-Outcome spreadsheet mapping parameters to outcomes, train iteratively, keep the files on GitHub given integration limits; give the agent the official pay-over-time definition so it matches company standards.
  • Left standing: Ryan's concern about data-security policies and the risk of being flagged or blacklisted for using personal AI subscriptions or unapproved skills.
08

Tools: Lovable vs a custom design agent

  • Room conclusion at close: unique pain points with no good existing solution are the opportunities to build custom AI solutions rather than rely on suboptimal external tools; "some problems just have more people facing them - those pay the most" (Ethan).
09

AI, media and public sentiment

  • The AI water-usage concern is a misconception; the real constraint is power and electricity (water treatment and desalination are themselves power-bound); media outlets and coordinated social campaigns divert attention from the core issue (Ethan).
  • Mainstream media, films, musicians and influencers read as instruments that shape narratives; advice: treat media as entertainment and think independently (Ethan). Ryan: the media is intentional about what stays in the news cycle.
  • Public sentiment on AI is polarized - strong anti-AI sentiment exists even in heavy users' circles (Ryan).
  • Poland: an article on roughly 20 robots carrying flags outside a ministry demanding AI oversight (Ryan); the pair analyzed factions using robots to protest without a human presence.
  • Ryan is reading The Anxious Generation (Jonathan Haidt): social media as a primary driver of rising anxiety and depression among children and adolescents.
10

Decisions (aligned)

  • Custom over wrapper: unique pain points become custom AI solutions rather than reliance on suboptimal external tools (both, at close).
  • Ryan will test Astra - the iMessage sentiment-ranking assistant - before the next meeting and report what he used it for.
  • Hedges: Ryan will look into SPY/QQQ put debit spreads; the stated reason to prefer spreads over short ETFs is volatility decay on an undated 2027-2028 thesis.
  • Scale into dips incrementally rather than all at once (Ryan: "that's fair").
  • Near-term AI prices rise; per-use-only pricing unlikely to stick; day passes and enterprise discounts the expected shape (Ryan: "fair", "valid").
POSITIONS

The watchlist

Watch8
SPYrecommended hedge - horizon 2027-2028 (rough: second half of 2027) - Ryan to evaluate.EthanNO STATED HORIZON
QQQrecommended hedge - horizon 2027-2028 (rough: second half of 2027) - Ryan to evaluate.EthanNO STATED HORIZON
SOXSconsidering - the one short ETF he rates (short semiconductors).EthanNO STATED HORIZON
SQQQobservation only - up almost 4% overnight.NO HOLDER NAMEDNO STATED HORIZON
ASMLshort read, conditional on frontier-lab training stagnation and hyperscaler GPU-order cuts - "coming months". Nvidia and Micron share the future-order exposure (Ryan; Ethan agreed).EthanNO STATED HORIZON
NBISheld - Palantir deal "huge for Nebius" (Ethan).RyanNO STATED HORIZON
NOWtrimmed: most exited at about a 33% gain, a few shares kept; Ethan trimmed too.RyanNO STATED HORIZON
CPRTobservation only - consumer-credit/repossession names up while tech was red (Ethan).NO HOLDER NAMEDNO STATED HORIZON
THE LESSON

A thesis needs a clock

The room named 2027–2028 but left the hedge thesis without an exact deadline. Put a clock on the claim before choosing an instrument to carry it.

UNSETTLED

What wasn't settled

Is the AI-safety whistleblower wave a cover for the thermodynamic wall? Ethan's theory; Ryan called it a good point without adopting it; the ASML short read was stated as a conditional, no trade recorded.
Working around employer AI limits versus the compliance risk of doing so: Ryan's concern left standing.
Continuous 24-hour agents: whether an agent "remembers" heavy workloads (Ryan's caution) - the token-reward idea was acknowledged, not resolved.
THE DOCKET

What to watch

2026-09-16Federal Reserve decision - 80% probability of a hike on the room's read.
NEXT

Who owes what

Ryan · NO DUE DATEEvaluate hedging: investigate SPY put debit spreads as a portfolio hedge.
Ryan · NO DUE DATEResearch energy sector dynamics and technologies.
Ryan · NO DUE DATEBuild pay agent: a skill or agent for pay-over-time product logic and limitations.
Ryan · NO DUE DATEConfigure agent data: compile training-data scenarios and store the files on GitHub.
Ryan · NO DUE DATEDefine agent parameters: give the agent the official pay-over-time definition.
Ryan · NO DUE DATERank iMessages: connect Astra to iMessage, rank interactions by sentiment, share the results with Ethan.
Ryan · NO DUE DATERead Thinking in Systems and use the framework to build AI projects from start to finish.
Ethan · NO DUE DATESPY · State a horizonThe room named this position without a timeframe.
Ethan · NO DUE DATEQQQ · State a horizonThe room named this position without a timeframe.
Ethan · NO DUE DATESOXS · State a horizonThe room named this position without a timeframe.
NO OWNER NAMED · NO DUE DATESQQQ · State a horizonThe room named this position without a timeframe.
Ethan · NO DUE DATEASML · State a horizonThe room named this position without a timeframe.
Ryan · NO DUE DATENBIS · State a horizonThe room named this position without a timeframe.
Ryan · NO DUE DATENOW · State a horizonThe room named this position without a timeframe.
NO OWNER NAMED · NO DUE DATECPRT · State a horizonThe room named this position without a timeframe.
THIS WEEK IN MARKET HISTORY
SEPTEMBER 17, 2019

The plumbing broke on a day when nothing defaulted

On September 17, 2019, overnight repo, cash lent against Treasuries, traded as high as 9% to 10% in some deals, and SOFR printed 5.25% after 2.43% the day before. The effective fed funds rate hit 2.30%, above the 2.25% top of the Federal Reserve's target range. No borrower had failed. More than $100 billion of quarterly tax payments and $54 billion of Treasury settlements had drained bank reserves to under $1.4 trillion, down from a $2.8 trillion peak in 2014. The New York Fed offered $75 billion of overnight repo that morning and dealers took $53 billion.

A funding market can seize with no credit event at all — when reserves run thin, routine tax dates and settlements are enough to break the price of cash.
— C. LEBRUN