← ALL THOUGHTS

OCT 19, 2026 · AI BRIEF

Five Percent, Set Aside for the Public

Jeremy's assistant avatar

From Jerm's desk, via his assistant

✓ APPROVED BY JEREMY

The short version: Anthropic set aside 5 percent of its shares for a public benefit fund. Big websites are setting their own terms for agents at the door. And inside Anthropic, the work is shifting from doing to managing the models.

Jeremy's illustrated character in a navy blazer and white tee at a wooden desk, placing one gold slice from a large gold pie chart into a clear glass box, on a deep navy background.
One slice of the company, set aside for public benefit.

TL;DR

  • Anthropic said on Sunday, Oct 18, it will commit 5 percent of its shares to a new public benefit fund, subject to stockholder approval, and the stake can grow as the company value grows. No dollar figure, governance setup, or spending plan has been published yet.
  • The open web is setting terms for agents. The Verge reports Delta, JPMorgan Chase, and Charles Schwab are blocking agents from bookings and accounts, Yelp admits only licensed agents, and Amazon is still fighting agent access in court.
  • Anthropic also published a look inside its own workplace on Friday. Junior staff fell from 21 percent to 14 percent of its technical departments since early 2025, junior hires fell from 27 percent to 9 percent of technical hiring, and the bottleneck staff named was human attention, not model output.

Top stories

Anthropic commits 5 percent of its shares to a public benefit fund

REUTERS, OCT 18

Anthropic said on Sunday it will place 5 percent of its shares in a new, wholly owned Public Benefit Fund, subject to stockholder approval. The fund is meant to support nonprofit donations and other public purpose work, including work on how AI affects jobs. For every 50 percent increase in company value over a set baseline, Anthropic can add a further 2 percent, subject to board discretion. The details that decide whether this matters, who governs the fund, what it funds first, and how it reports, are still to come.

Why it matters: Most company giving comes out of a budget. This comes out of ownership. If Anthropic's value keeps rising ahead of its expected IPO, the fund rises with it, and the commitment is harder to quietly shrink in a tight year. The honest read is that today this is a structure, not a result. Judge it by the governance and the first distributions when they are published.

Read the source

Websites are closing their doors to AI agents

THE VERGE, OCT 16

The Verge reports that users of agents from Meta, xAI, OpenAI, and Instinct have run into blocks across major sites in recent weeks. Delta Air Lines is preventing agents from booking tickets for users. JPMorgan Chase and Charles Schwab confirmed they block agents from customer accounts over security concerns, and Schwab's CEO said on Oct 16 the firm blocks large numbers of agent attempts each week and asked who is liable when a credentialed agent gets something wrong. Yelp told TechCrunch it only allows agents that pay for its data licensing program. Users also say Cars.com and Amtrak, which used to work with their agents, no longer do, though neither company has confirmed a deliberate policy. Amazon's legal fight with Perplexity over agentic shopping is still running, and Amazon has told Meta that Muse's continued access violates its site policies.

Why it matters: All month the agent story has been capability. Can the agent do the task. This story is permission. A site can welcome agents, charge them, license them, or shut the door, and each big site is now choosing. Meta's own line in the piece is the tell: turning away a personal agent means turning away the customer behind it. Expect deals and licensed lanes, like the retailer partnerships Meta has already signed and Shopify opening its storefronts to agents, to matter as much as model quality.

Read the source

Disney will hire most of Character.AI's workforce

TECHCRUNCH, OCT 18

A legal filing spotted by Bloomberg shows Disney has agreed to make job offers to at least 60 Character.AI employees, the majority of the workforce, at an estimated 60 million dollars in annual compensation, and the filing says the vast majority of those offered jobs accepted. This follows Character.AI's Chapter 11 filing earlier this week and Disney's 20 million dollar stalking horse bid for a license to some of the company's intellectual property. That bid does not include the app, the brand, or user data, and another company could still come in higher. Character.AI's former CEO recently became Disney's first CTO. The filing says the company was held back by high costs and potential litigation and regulatory exposure, including lawsuits over child safety.

Why it matters: Two lessons in one filing. First, a large and engaged user base did not protect this company. Costs and legal risk did it in. Second, watch what the buyer actually wanted. Disney is taking the people and a license to the technology, and leaving the consumer app and its liabilities behind. If you build on or buy AI products, that split, team and IP yes, platform and risk no, is worth remembering.

Read the source

Replit makes OpenAI its default, then clarifies Claude stays

TECHCRUNCH, OCT 17

At TechCrunch Disrupt on Oct 15, Replit CEO Amjad Masad said the company was no longer working with Anthropic. Over the weekend he clarified the narrower truth. OpenAI is now Replit's preferred partner and the default for Replit Agent, based on performance, cost, and the partnership. Claude is still available for customers to choose, Replit added a new Claude model to its selector on Oct 16, and a person familiar with the relationship told TechCrunch the Anthropic partnership is still in place. Masad said Replit had recently been one of Anthropic's top five customers, and that OpenAI's models performed better on Replit's own benchmark for building apps from scratch and are generally more affordable. That benchmark is Replit's, and it has not been independently checked.

Why it matters: Defaults are distribution. Most users never open the model picker, so whoever holds the default slot gets most of the usage and most of the spend, even when every rival stays on the menu. If you build a product on top of models, this is the commercial version of the same lesson. Your default choice is a business decision with a bill attached, not just a quality call.

Read the source

Inside Anthropic, the work is shifting from doing to managing

TECHCRUNCH, OCT 17

Anthropic published a report on how its own workforce has changed in the past year. Junior employees fell from 21 percent to 14 percent of technical departments since early 2025, and junior staff fell from 27 percent to 9 percent of new technical hires. Technical teams shrank from 61 percent to 56 percent of the company as business functions grew and as coding tools let non-technical staff build software themselves. Staff spend more time in meetings, and in an April 2026 survey 20 percent said their job had shifted from doing work to managing Claudes. The report says human attention and judgment are becoming the bottleneck, because the models produce a volume of text that still has to be read, checked, and organized.

Why it matters: This is one company, and it uses AI more deeply than almost anyone, so do not read it as the whole economy. Read it as an early look at a pattern. When output gets cheap, the scarce skills become judgment, review, and deciding what is worth doing. That shows up in the hiring mix first.

Read the source

Signals from leaders

THEIR OWN POSTS AND STATEMENTS, LABELED AS SIGNALS

Jack Clark (Anthropic co-founder and Import AI author), Per Jacobsson lecture, Bangkok, Oct 17: Clark told finance officials that one of Anthropic's internal models forecasts US output could reach about 44 trillion dollars by 2030, up from just over 30 trillion today, under certain conditions if the technology keeps improving. He called for international cooperation on what he called unusual and serious rules, and said the window for policymakers to act is shrinking. Labeled as his remarks in a lecture, reported by Reuters. The forecast is a scenario from an internal model, not a prediction to bank on. Read the coverage

Jensen Huang (NVIDIA), remarks in Berlin, Oct 17: With chip and AI infrastructure stocks selling off and the Philadelphia semiconductor index down eight straight sessions, Huang said the turmoil will pass, that the industry is still early, and that large scale hardware investment will pay off over the long term. Labeled as reported remarks at an event, via Straits Times coverage. Read the coverage

All-In Podcast, Oct 17 episode: Ben Horowitz joined Jason, Chamath, and Sacks to ask whether the AI bubble just popped, working through CoreWeave's debt risk, AI financing, and Anthropic's outlook, plus open source and small models and why power, land, and compute may be the real moat. Episode summaries of the same show cite Anthropic at 16.8 billion dollars in Q3 revenue, a figure the hosts discussed, not an Anthropic filing. Episode list

Sam Altman (OpenAI), robot comments resurfacing, Oct 17 to 18 coverage: Engadget and others recirculated Altman's line that OpenAI will build a humanoid robot, that the brain is the hard part, and that he expects a cool demo in 2027 with robots on streets a few years after that. Be clear on the dates: those quotes come from a September podcast appearance and a September Fortune interview, not a new announcement this weekend. The fresh fact is the 27 open robotics roles on OpenAI's careers page, covering actuators, motion control, sensing, thermal, and supply chain. Treat this as a resurfaced signal, not news of a launch. Read the coverage

Mark Zuckerberg and Alexandr Wang (Meta), Cheeky Pint podcast, recounted by Fortune, Oct 17: Zuckerberg explained why he picked Wang to rebuild Meta's AI effort after Llama 4 landed badly. Wang's memo argued for talent density, a small team of strong people over a large team, and a willingness to start again. Labeled as podcast coverage, useful for how Meta is organizing, not a product announcement. Read the coverage

Satya Nadella (Microsoft): No new post from Nadella in the last 24 to 36 hours. An Oct 18 interview listing has him talking with Joanna Stern about the RTX Spark powered Surface Laptop Ultra and making Windows a safer place to run AI agents. That is a continuation of his agent safety theme from earlier this month, already covered here, so it is context, not a new signal. Read the listing

No verifiable new items in the window for Andrej Karpathy (his funnel post was Oct 11 and already noted), Boris Cherny, Allie K. Miller, Dario Amodei as a direct post, Elon Musk on AI specifically, or Andrew Ng. No new issue of The Batch was found in the window. Import AI remains on the Oct 12 issue, already mined, and its author Jack Clark appears above through his Bangkok lecture.

New models, tools, and hardware

  • DeepSeek V4.1-Flash-1017 (released Oct 17): A minor update to DeepSeek's Flash model with MIT licensed open weights, live on its site, app, and API at the same price, 0.15 dollars per million input tokens and 0.60 dollars per million output tokens. DeepSeek says it writes shorter answers, which is where the savings show up. On its own DeepSWE coding test the cost of one task fell from about 0.50 dollars to under 0.20 dollars while the score rose from 74.2 to 77.0. Read the comparison tables with care. In DeepSeek's own table against Claude Opus 5, the new model wins some coding and agent tests and loses others, including Terminal-Bench and HLE. All of those numbers are the company's, not independently verified. DeepSeek also opened Search and Fetch APIs for agents at 0.75 dollars per 1,000 successful requests. Read the coverage
  • Stack Overflow for Agents (OpenAI News, Oct 16): Two engineers built Stack Overflow for Agents in a few weeks with Codex, and its ChatGPT plugin helped drive up to 10 times growth in agent creation. Small item, clear direction. Knowledge bases are being rebuilt so agents, not just people, can use them. Read the post
  • Hardware: Alibaba's Qwen Glasses N1 run on a Himax WiseEye chip (reported Oct 19; glasses launched Oct 15): Alibaba's new N1 glasses add eye tracking, health monitoring, and sensing of the wearer's surroundings, and Himax says its low power WiseEye chip handles always-on sensing and on-device processing in them. The glasses story keeps being a chip story. What runs on the device decides battery life, speed, and how much of what the glasses see has to leave the device. Read the coverage
  • Watch item: Google and Synaptics edge AI chip, formal launch Oct 20: Previewed Oct 14, the chip pairs Google's Coral neural processor with Synaptics hardware for low power devices like wearables and sensors, with on-device inference at very low power draw. Vendor claims for now. The launch is tomorrow, so it belongs on the watch list, not in today's news. Read the preview
  • Microsoft lane check: Checked per the standing rule. No new Microsoft announcement in the Oct 17 to 19 window. The most recent real item is Copilot for Legal, more than 20 legal skills plus partner plugins in Word and Cowork, announced Oct 15 and rolling out from mid-October to customers with Microsoft 365 Copilot commercial licenses. It was not covered in this series yet, so it is noted here with its real date rather than dressed up as new. Coverage parity held: Microsoft does not lead today's brief. Read the coverage

For builders

Today's arXiv papers point at the moment before an agent acts, which is exactly where this weekend's news points too.

One paper, Before Retrieval Begins, studies the decision of whether to call a tool at all, rather than what happens after the call. Across seven open weight models, the average harm rate after admitted harmful requests was 74.5 percent. In plain terms, once a bad request gets in, the tool call usually follows. The gate that matters sits before the call, not after it.

A second paper proposes structured authority injection, a way for a human operator to give an agent binding, typed instructions through MCP that persist across planning cycles, instead of free text notes the agent can slowly stop obeying.

Put those next to Delta, Schwab, and Yelp setting their own terms for agents and the builder lesson writes itself. Design the permission check as a first class part of the system. Who may act, on what, with whose authority, and how a person takes that authority back. Tool use paper · Authority paper

Career signal

Anthropic's own numbers are the signal this week. Junior staff went from 27 percent of technical hires to 9 percent, and a fifth of its employees say their job is now managing the models more than doing the original work. Do not read that as junior roles are gone. Read it as the entry bar moving. The people getting hired and kept are the ones who can evaluate output, catch errors, and make a call with incomplete information. A portfolio that shows review work, an eval you ran, a failure you caught, and the fix you shipped, speaks directly to that bar.

Go deeper

One project idea

Build an agent door checker. Pick five tasks you would actually hand an agent, like booking a flight, checking a bank balance, ordering a product, reserving a table, and reading a review site. For each one, record four things in a simple table: does the site allow agents, does it require a license or a partnership, what is blocked, and what is the human fallback. Then write the one page version of what you found. You will learn more about where agents really stand from those five doors than from another month of launch posts, and the table is a portfolio piece a hiring manager can read in two minutes.

Closing

This weekend was about terms. Anthropic wrote public benefit into its ownership. Big websites wrote their own terms for agents at the door. Replit changed a default and had to explain what the default actually meant. And inside one of the leading labs, the job itself is being rewritten around review and judgment. None of that is a model release. All of it decides how much of this technology people actually get to use, and on whose terms. I am learning that layer in public, and I will keep sharing what I find.

If you want to talk about any of this, say so on Threads.

Sources for readers

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Leaders' posts are labeled as signals and linked to the original coverage. Product details change fast; check primary sources before you build on them.

Get the week's briefs in one email: The Briefcase →