Noahpinion
01:47 PT
The second Trump presidency is a world-historic failure
So what
Price a Democratic House and plausible Senate into 2027 policy assumptions, and stop assuming the AI deregulation posture survives the midterms.
Noah Smith argues the second Trump term has already failed on the only scoreboard that matters near-term — public opinion — and that the failure was self-inflicted rather than structural.
- Approval at historic lows going into the midterms; Republicans now disapprove of Trump on cost of living, his signature 2024 issue
- Prediction markets now favor Democrats to take the House and give better-than-even odds on the Senate despite a brutal map
- A $5,000-per-citizen check promised contingent on a GOP midterm win is polling as a net negative and actively driving voters away
- Party trust has flipped: the GOP led the Dems a year ago; Democrats now lead on the economy and immigration
- Peter Thiel and other Tech Right figures are visibly distancing; Smith concedes the one hedge is that stalling AI regulation could retroactively redeem the term
Claude Code
19:19 PT
This week in Claude Code: build your own mods, a wrap-up allowance, and more
So what
Re-run prompt-audit across your repos — two model launches in ten days means your CLAUDE.md files are now tuned for models you no longer run.
Anthropic shipped mods — user-written TypeScript hooks into Claude Code itself — alongside Opus 5.5 and Sonnet 5.5, and opened a plugin directory with a developer portal and review process.
- Mods can hold or rewrite a tool call, draw UI, add commands, or replace built-ins; Claude will write them for you from a prompt. CLI and desktop v2.1.287+
- Opus 5.5 is the new default: roughly Fable 5.1 level, 30%+ faster than Opus 5, 40% cheaper, shorter replies, fewer em dashes
- Cloud sessions out of research preview with one-time bonus credit of $100 on Pro and $250 on Max, claimable by October 7; five-hour limits up 20% plus a one-time reset
- Wrap-up allowance draws from weekly limits to finish a task when you hit the five-hour wall — once a week on Pro, every time on Max and Team Premium
- Claude Code now reads AGENTS.md when no CLAUDE.md exists; /checkup prompt-audit flags instructions written for older models; claude-api can build evals and hillclimb them
McKinsey & Company
13:48 PT
Economic conditions outlook, September 2026
So what
Thin signal: executive sentiment is drifting up off the June trough, which argues against timing deals to a sentiment inflection that hasn't happened.
McKinsey's September 2026 economic conditions survey finds executive pessimism easing from June but still net negative, with six-month expectations mixed.
- Negative views on the global economy remain more common than positive ones among surveyed executives
- The improvement is directional only — caution, not confidence, is the posture
- Companion pieces cover managing geopolitical value at stake and the August Global Economics Intelligence summary
Every
12:41 PT
How to Get Better at AI by Asking AI
So what
Make your AI grade your own usage against a rubric weekly — the self-assessment loop is cheaper than any training program you'd buy.
Arielle Shipper's account of going from single-thread chats to orchestrating subagents, using the AI itself as the teacher — five prompts, scored against Mike Taylor and Laura Entis's Eight Levels of AI Adoption.
- Asked Codex to read her session history and place her on the framework: it came back 5.5, with orchestration named as the missing step; four months later she scored 7.9
- Set up a recurring Friday task that scans the week's sessions, scores her, cites supporting sessions, and proposes experiments for the following week
- Replayed a finished Stripe sales-tax project as a multi-agent design: API extractor, verifier, workbook builder — seeing a past project redone is what changed behavior, not the advice itself
- Orchestrator thread holds goal, constraints and definition of done; each specialist gets a context packet rather than the whole project
- Built a Self Improve skill: after a Ramp-card task produced off-voice Slack messages, she had Codex critique its own draft and write durable rules into the skill file
The Information
12:01 PT
Exclusive: OpenAI Hires Top Trump AI Official to Work on National Security
So what
OpenAI is buying both sides of the Washington bench; assume its national security and federal procurement position hardens regardless of the midterm outcome.
OpenAI hired Thomas Lind, former AI policy lead at the White House Office of the National Cyber Director, to run cyber and strategic risk on its national security policy team.
- ONCD played a leading role in assembling the June AI executive order; Lind was its AI policy lead
- He reports into Sasha Baker, former deputy Under Secretary of Defense for Policy under Biden
- Continues a pattern of OpenAI stocking its leadership bench with former Trump administration officials
AI Market Fit
10:30 PT
Brian Chesky’s Playbook for the Agent Era
So what
The main-agent seat is priced as if it's won; Chesky's hands-on failure test says multi-step consumer commerce is still unbookable through any of them.
Brian Chesky shipped Airbnb's largest AI release while publicly saying his own product breaks inside the consumer agents everyone is funding — a deliberate bet against the chatbot interface.
- Chesky tried booking an Airbnb through Instinct and Meta's Muse; both struggled with hotels and failed harder on homes
- Instinct raised $1B at a $10B valuation, 4x its August price; Muse has millions of downloads and sits atop the app stores
- Airbnb's fall release makes the whole app the agent — AI search with no chat box, on the thesis that multi-step travel (browse, compare, message host, decide with friends) isn't a single transaction
- Chesky told Sam Altman the ChatGPT app store isn't the answer and argues Apple and Google will decide the agent OS
- He claims 80% more features shipped with AI, calls token leaderboards a mistake, and expects 50 to 100 more Anthropics
Ben Thompson (Stratechery)
10:01 PT
Dots and Question Marks (This Week in Stratechery)
So what
Two of the three plausible agent aggregators spent the week misallocating — Meta toward enterprise, OpenAI toward branding chaos — which widens the opening for whoever stays focused.
Thompson's week argues agents are the ultimate Aggregators — the biggest prize in consumer tech — and that both Meta and OpenAI are fumbling the moment in opposite ways.
- Apps, Agents, and Aggregation: agents reveal apps as a means not an end; something like Muse could aggregate the Aggregators and become the most valuable product in the world
- A day after that bullish case, Meta announced Meta Enterprise Platform — Thompson can't identify the customer and calls it a distraction from the biggest opportunity in consumer tech
- OpenAI's Dev Day brought back the chaotic branding of a year ago: new pricing tiers, overlapping products, and dots pitched at mass market but gated to pro subscribers
- Sign In With ChatGPT and the product transition are the coherent thread Thompson digs out of the mess
- Sharp China reads Xi's D.C. visit as Mao's strategic stalemate phase, possibly a prologue to counter-offensive; the weekly video is on Frontier Overhangs
The Information
09:54 PT
Sean Parker Spreads the Gospel of AI Music
So what
Stability is no longer a distressed asset story; watch whether label and producer relationships become the moat generative audio startups can't buy.
Sean Parker breaks a two-year silence on Stability AI, which he quietly took control of after it nearly collapsed, and frames it as a vehicle for getting recording artists and producers onto AI tools.
- Parker's arc with music runs Napster, then early Spotify investor via Founders Fund, then long-serving Spotify board member
- Stability was close to cratering out of existence before he assumed control two years ago
- His pitch is adoption from inside the industry rather than disruption from outside — he calls himself one of the few masochists willing to go back
The Information
09:33 PT
AI Data Center Debt Is Showing Up Everywhere
So what
Screen your fixed-income exposure for data center SPE paper; broad high-yield ETFs now carry AI build-out credit risk you never underwrote.
AI data center debt has become its own credit category, and the high-yield end of it is wobbling — with the biggest AI names sitting one or two leases away from obscure special purpose entities.
- At least 20 high-yield bond deals financed data center projects in the past 12 months, via issuers like CleanSpark and TeraWulf through SPEs
- Spreads widened on tech giants' investment-grade debt, investors demanded bigger concessions on new AI data center high-yield bonds, and bank loan markets for AI project finance turned skittish
- Google backs certain Fluidstack lease obligations; Fluidstack builds for Anthropic, which is dramatically expanding TPU use. Two Nevada SPEs (SV RNO, PR RNO) lease to an unnamed $3T+ tenant Bloomberg identifies as Nvidia
- Nvidia disclosed $25B of not-yet-commenced data center leases plus $20B it plans to reassign to third parties; a SoftBank affiliate guarantees the tenant on SB Energy's Cosmos campus
- State Street and Charles Schwab high-yield ETFs already hold Fluidstack and Fleet Nevada project debt among top holdings — under 1% each, but the exposure is no longer avoidable