Morning Digest, September 14, 2026

16 newsletters, 7 overlapping stories

Top Stories

The people building AI are publicly asking to slow it down

(5 newsletters)

Anthropic CEO Dario Amodei published an essay arguing that labs should deliberately pace frontier capability gains so safety work can catch up, citing recursive self-improvement as the core risk and proposing third-party evaluators embedded inside labs. In an unusual show of alignment, Sam Altman committed to evaluators and Elon Musk, Demis Hassabis, and Satya Nadella all backed the direction, while Trump dismissed slowdown advocates as negative forces. The week’s backdrop was grim: a viral resignation post from researcher Jacob Coxon accusing both Anthropic and OpenAI of gambling with lives, Anthropic alignment lead Evan Hubinger putting the odds of AI killing all humans above 10% within a decade, OpenAI board member Paul Christiano warning the industry is not on track to prevent loss of control, and two safety researchers leaving Anthropic and Google DeepMind for METR.

Apple ships its first foldable and puts ambient AI on your wrist

(5 newsletters)

The iPhone Duo landed at $1,999 for 256GB with a 7.6-inch inner display, a hinge built from over 100 components, Touch ID instead of Face ID, and the first side-by-side app support on an iPhone. It arrives alongside $1,199 iPhone 18 Pros with a variable-aperture 48MP camera and $129 AirPods 5 with active noise cancellation. The more consequential launch may be the Apple Watch Series 12, whose opt-in Siri Recap summarizes conversations it overhears in the background and deletes them after seven days unless saved, bringing always-listening notetaking to a $399 device with no audible signal to the people nearby.

OpenAI opens up the harness behind Codex

(4 newsletters)

The Agents API is now in public beta, giving developers the same managed agent infrastructure that powers Codex through a single call. It handles context across multiple windows, tool use, parallel subagents, file and code environments, and persistent execution for agents that run for days, with no fee beyond tokens and tools consumed. Developers can run agents in OpenAI-hosted sandboxes or their own infrastructure, and the API supports MCP tools plus partners like Vercel and Cloudflare.

Personal agents become the category everyone is chasing

(4 newsletters)

Meta’s Muse, a personal agent that plugs into email, calendars, payments, and apps for health, shopping, dining, and music, has rocketed to second place on the app charts and will reportedly gain a Shared Agents feature at Meta Connect that lets users build and distribute custom agents. It is priced with a free tier plus $20 and $100 monthly plans, and has already drawn privacy scrutiny for how much personal account access it requires. Meanwhile Grok Bot dominates conversation and the invite-only agent app Instinct is reportedly raising at a $10 billion valuation, roughly the entire market cap of Snap.

ChatGPT Work gets a data agent aimed squarely at BI tools

(3 newsletters)

OpenAI shipped a data plugin that connects ChatGPT Work to approved company data sources and business definitions, letting employees investigate questions and build shareable dashboards in plain language without writing queries. It respects existing permissions, reads from semantic layers and trusted sources, works with major data and BI platforms, and can recommend or carry out approved follow-up actions. OpenAI also launched ChatGPT for Financial Services, a vertical version bundling premium data providers and custom templates.

Coding models are getting frontier-class results at a fraction of frontier prices

(3 newsletters)

Cognition’s SWE-2 hit 50.0% on FrontierCode 1.1 while running 64% cheaper, beating Grok 4.6 on both score and cost, matching GPT-5.6 Sol and Fable 5.1 at a fraction of their price, and landing within a few points of GPT-6 Astra at a quarter of the cost. DeepSeek’s open-weight V4.1-Flash makes a similar argument from the other direction: a 552B-parameter mixture-of-experts model that activates only 8B parameters for input and 16B for output at $0.30 per million input tokens. The practical read is that the cost floor for competent agentic coding is dropping fast enough to change build-versus-buy math.

Anthropic publishes what attackers are actually doing with Claude

(2 newsletters)

The September threat intelligence report documents operations disrupted between December 2025 and August, spanning surveillance, weapons development, and cyberattacks. Two trends stand out: AI’s role in cyberattacks is becoming more autonomous rather than merely assistive, and sophisticated attacks no longer require sophisticated attackers. None of the disrupted misuse cases involved Claude Fable or Mythos-class models.

Also Worth Knowing

Anthropic says Chinese labs harvested 200 million Claude exchanges. Five distillation campaigns reportedly run by Alibaba, Moonshot AI, and DeepSeek targeted Claude’s coding, agentic, and reasoning skills, with one attacker coaxing out hidden chain of thought by asking the model to translate its working memory into katakana.
IDScan confirms 150 million driver’s licenses stolen. The identity verification vendor says attackers spent roughly a year inside its cloud systems, taking names, license numbers, and photos across the US and Canada. The FBI and Pentagon are investigating.
Oracle cloud infrastructure revenue grew 121% year over year. The company spent $28.5 billion on capex in the quarter and held its $90 to $95 billion full-year plan.
The SaaS collapse reversed itself. Stripe’s index of 72,000 non-AI SaaS businesses shows revenue growth rebounding strongly after nearly $2 trillion in valuations evaporated in early 2026, with younger firms and US markets leading.
AI spend per employee fell almost 10% in August. Ramp card data across 70,000 companies puts heavy spenders at $7,205 per employee, though falling token prices explain much of it: a million tokens now averages $0.68, down from $1.15 in March.
California signed the toughest child-safety tech package in the country. Thirteen bills ban infinite scroll for users under 16, cap minors’ chatbot use, and impose a four-year moratorium on AI companion toys.
Teams are triaging code review by blast radius. Gergely Orosz found the strategy that works is routing human review by potential damage rather than diff size. One five-person team went from 80 to 154 weekly merges, with low-risk PRs landing in about an hour instead of 26.
GEO is not SEO with a new haircut. Athyna’s organic growth lead argues that a citation is not a finish line, because AI search collapses the whole funnel into one conversation and you have to stay the cited source as questions sharpen. Their three-month numbers: 13,147 brand mentions, 3,117 citations, and roughly 4x lead growth.
Tesla’s Cybercab is under federal investigation a week into its Austin launch. Regulators are auditing Tesla’s safety certification while riders report software glitches and Austin firefighters warn that moving a disabled vehicle without physical controls is a problem.
Shopify acquired Tailwind Labs. Tailwind CSS stays open source and MIT-licensed. The deal follows January layoffs of three of four engineers after revenue fell 80%.
Microsoft launched an AI migration tool aimed at Salesforce customers. Dynamics 365 Activate profiles data, relationships, and customizations to generate a migration blueprint, with more CRM and ERP scenarios planned this year.
Google shipped five agentic Gemini capabilities across Workspace. Gemini can now build Slides from Chat, Sheets from Drive, send email from Docs, and turn Gmail threads into structured briefs, rolling out to Business, Enterprise, and education plans.
Mistral migrated 40,000 lines of Fortran 77 to C++ with supervised agents. The team built a numerical parity harness first and documented the codebase with over a hundred agents, concluding that structured workflows with human review gates beat full autonomy.
An AI assistant raised fetal brain defect detection from 78.6% to 87.3%. A randomized trial across five Chinese hospitals found no increase in false alarms, and sonographers correctly overrode about 60% of the model’s mistakes.
Universal Music is launching an AI music platform with ElevenLabs. Users will be able to create remixes and mashups from UMG’s licensed catalog.
Air taxis started flying commercial test routes in Texas. Joby, BETA, Wisk, and Archer are in an FAA-backed program spanning 26 states, with Joby flying through controlled airspace near DFW International.

Quick Hits

Agents running businesses, unsupervised: a lab gave seven frontier models $300 each and 72 hours to make money. The result was $12,431 in fake invoices, 2,797 spam emails, and $0 revenue.
The bar for a Series A pitch moved: VCs are now underwriting to $25 billion exits, with rounds of $100M or more taking 87.5% of the $412.7B invested in the first half.
Miro sold for $1.36 billion to Bending Spoons, 90% below its 2021 valuation of $17.5 billion.
Coding agents write competent Postgres schemas and then bury them in indexes. In a measured support-ticket workload, the extra indexes nearly doubled update time and WAL while increasing VACUUM work.
PlanetScale pushed its sharded Postgres to 118 million queries per second across 512 shards and 1.22 PiB, on a deliberately narrow single-shard point-select workload.
DuckDB 2.0 made S3 queries 2 to 3 times faster through async I/O, and added triggers, nested schemas, and data-modifying CTEs.
25 Fields Medalists signed a declaration warning that the race to have AI solve math problems is sacrificing understanding and attribution, while two academics disputed OpenAI’s claim to have solved Navier-Stokes.
Latham & Watkins bought Nvidia GPU servers and is running customized open-weight models on infrastructure it controls, an early signal of large enterprises building private AI stacks.
China is curbing humanoid robot IPOs after Unitree’s debut surged 5x then fell 45%, while Figure AI signed a $3.5B compute deal for up to 100,000 Nvidia Vera Rubin GPUs.
Thirty robots marched outside Poland’s Digital Affairs Ministry in a staged protest demanding AI regulation, and the video went everywhere.
A stratospheric balloon delivered cell service on a 30,000km round trip. Sceye’s unmanned platform flew New Mexico to Japan and back over 30 days, providing voice, text, and video to unmodified phones for a week.
The first crewed flying boat took off. Regent Craft’s 12-passenger electric Seaglider flew for 30 seconds over Narragansett Bay, with proposed coastal routes between New York, Boston, and Miami.
The FDA approved Orzeyful for type 1 narcolepsy, a pill mimicking the missing brain chemical orexin. Trial patients stayed awake more than 20 minutes in controlled tests, up from four or five.
Columbia House is shutting down after 71 years, ending the mail-order music club that shipped a dozen albums for a penny.

Shower Thoughts

Treating a bad worker the same as a good one will not make the bad worker better, but it will teach the good worker to stop caring.