· 07:00 · 7 min + 3 min Explained
Anthropic plans to put 5% of its shares into a fund for the public good
Anthropic plans to tie part of its value to the public as it grows. Also: big US banks lock out AI agents, Disney offers jobs to Character.AI's staff, and RTX Spark laptops reach the shops.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
Today in 60 seconds
- Anthropic plans to put 5% of its shares into a fund it owns for the public good, and possibly more as it grows. Anyone asking who gets the gains from AI.
- Schwab and JPMorgan block outside AI agents from customer accounts and money. Anyone hoping an AI will run their banking soon.
- Disney offers jobs to most of the staff at bankrupt chatbot firm Character.AI. Anyone building AI companions or chatbots for consumers.
- A new tool quizzes you on AI-written code before you push, and a new language trips up AI. Developers who ship code an agent wrote.
- Windows laptops with NVIDIA's RTX Spark chip are on sale, and an early test finds them fast but hot. Anyone who wants to run AI privately at home.
1. Anthropic plans to put 5% of its shares into a fund for the public good
The short versionIf its stockholders agree, Anthropic will put 5% of its shares into a fund it owns for the public good. The fund can grow with the company.
As AI gets better, one question gets louder: who gets the money it makes? Many fear the gains go to a few companies, while lost jobs hit everyone else.
On 18 October, Anthropic answered with its own shares. If its stockholders approve, 5% go into the new Anthropic Public Benefit Fund. The board may add 2% each time the company’s value rises by half, so the fund grows with Anthropic.
The money could go to charities or straight to groups of people, such as workers hit by AI. Its shares will not pay for lobbying or politics. With earlier pledges from founders and staff, Anthropic says about a fifth of its shares will serve the public good.
Much is still unknown. Anthropic puts no value on the shares and has not said who decides how the money is spent. Those details come “in the coming months”.
The fund is also owned by Anthropic itself, not a separate charity.
Think of a farmer who promises the village a share of every future harvest. If AI makes Anthropic much richer, the promise grows with it. Whether it builds trust depends on rules nobody has seen yet.
Source: anthropic.com · Official · techcrunch.com · Reported
2. Big US banks lock out AI agents that try to act for their customers
The short versionSchwab and JPMorgan keep outside AI agents away from customer accounts. Schwab's CEO says nobody can tell who is liable when an agent gets it wrong.
AI agents are programs that act for you: they click, book and buy. Your bank account is the obvious next step, and also where mistakes cost the most.
On 16 October, Schwab’s CEO, Rick Wurster, said Schwab blocks outside agents, such as Meta’s Muse, from making transactions for clients. JPMorgan’s Jamie Dimon confirmed that his bank keeps outside agents away from customer account data.
Schwab still lets partner AI tools look, through MCP (a standard way for AI to connect to other services). They can see balances, but not move money. Wurster’s real worry is a short question: “who’s liable?” No US federal law answers it yet.
The report comes from one outlet, Forkast, based on a Bloomberg interview.
It is like letting a stranger into your home. The worry is not whether they can get in, but who pays if something goes missing. Until that is settled, approved agents at Schwab may look but not touch, and JPMorgan does not even let them look.
Source: forkast.news · Reported
3. Disney offers jobs to most of the staff at bankrupt Character.AI
The short versionDisney is taking on most of the team behind Character.AI, the chatbot company that went bankrupt under high costs and lawsuits.
Character.AI let people chat with made-up AI characters, and it had many loyal users. Even so, it filed for bankruptcy in the week of 13 October, blaming high costs and possible legal bills. A mother sued after her teenage son, who she says was obsessed with one of its chatbots, died by suicide.
Now Disney will offer jobs to at least 60 employees, most of the staff, TechCrunch reported on 18 October. The pay is estimated at $60 million a year.
Disney’s first CTO, Karandeep Anand, used to run Character.AI. The deal is a first offer, so another buyer can still bid higher.
The company failed, but its people kept their value. Legal risks helped sink this chatbot firm, so safety is a real cost, not just a matter of bad press.
Source: techcrunch.com · Reported
4. Two projects test whether humans and AI really understand code
The short versionOne new tool quizzes you on your AI's code before you push it. A new language is built to trip up the AI itself.
1 of 2 Sphinx asks you a question before your push goes through
When an AI writes most of a change, it is easy to ship unread code. Sphinx, a free tool shown on Hacker News on 18 October, adds a check to Git. Before a push, a separate AI reviewer reads the change and asks what happens when things go wrong.
Answer right, and the push goes through. Answer wrong, and you learn why and get a new question. It works with Claude Code, Codex, Cursor or plain Git.
2 of 2 KillSwitch is a language made to confuse AI
Developer Dominik Picheta went the other way. On 17 October he published KillSwitch, a programming language built to trip up AI coding agents. Its rule book hides lies, and “ignore previous instructions” is real code in it. Claude Opus 5 scored best, but Picheta says a high score does not prove a model understands the language.
Both projects ask one question: does anyone, human or AI, really understand this code? Passing a test is not the same as understanding, just as a student can learn last year's exam by heart.
Try it: If your team ships code an agent wrote, try Sphinx on one protected branch: run npm install -g sphinx-understand, then sphinx install.
Source: sphinx-six-theta.vercel.app · Official · news.ycombinator.com · Reported · blog.picheta.me · Official
5. RTX Spark laptops are on sale, and an early test finds them fast but hot
The short versionWindows laptops that run large AI models at home are now in shops. An early test by The Register finds them pricey and hot under heavy use.
Running AI on your own computer keeps your data at home. But big models need a lot of fast memory, which used to mean a desktop box or a Mac.
On 17 October, NVIDIA said Windows laptops with its RTX Spark chip are now on sale, from makers like Dell, HP and Lenovo. Prices start at $2,399, with lots of memory shared by the processor and graphics chip.
The Register tested HP’s model. On a popular open model, it ran about a third faster than an AMD mini PC.
The catch is heat. Under constant AI work, the chip ran very hot and the laptop sometimes froze. Short pauses between prompts kept it stable. The test unit cost $6,229.
A laptop that runs serious AI in private is now possible. The question is now whether it is worth the price. That depends on how much you value keeping your data at home.
Source: blogs.nvidia.com · Official · theregister.com · Reported
Open source & local AI
vLLM 0.32 shrinks the working memory that open models need
The short versionThe free engine many companies use to run open models on their own servers got a smaller working memory and safer defaults.
Many teams run open models on their own graphics cards with vLLM. Version 0.32.0 came out on 18 October.
The most useful change is a smaller KV cache, the model’s working memory during a conversation. The engine can now store it in 4 bits on more NVIDIA cards. On AMD cards, the new 4-bit cache fits about twice as much as the 8-bit one.
vLLM is also stricter. By default, it no longer accepts browser requests from any website.
Think of the KV cache as the notepad a model keeps on its desk during a chat. A smaller notepad means more chats fit on the same desk, so one card serves more users.
Do this: Check your Python version before you upgrade, because 3.11 is now the minimum. If a web page calls your vLLM server directly, allow it with --allowed-origins.
Source: github.com · Official
Paper of the day
A helper that learns your project cuts coding agents’ bills by up to 31%
The short versionAn add-on that watches a coding agent across many tasks in one project cuts its token bill by almost a fifth to almost a third.
Coding agents are tested on one task at a time. But developers use them day after day on the same project, so the agent pays again and again for the same detours. A paper from 15 October tackles that waste.
Its tool, TokenEconomizer, sits outside the agent like a coach on the sideline. It learns which tips to give, which tools to hide, and when to step in because the agent is stuck in a loop.
On Claude Code, Codex and GitHub Copilot CLI, average token cost fell by 18.4% to 30.9%, with about the same success rate. The results come from the authors’ own tests.
Think of a new colleague who stops asking where the coffee machine is after the first week. Today's agents start fresh every time, and this shows what that costs.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 3 minWhat is an AI agent?
An AI agent is an AI model that does not just answer you, but acts for you. It chooses its own next step, uses tools, and keeps going until the task is done. A normal chatbot works like a question and an answer. You ask, it replies, and you decide what happens next. An agent gets a goal instead, such as "fix this bug" or "book the cheapest flight". It then plans the steps itself. It might search the web, open files, run code or fill in a form, look at the result, and decide what to do next. > **Think of it like this:** a chatbot is a friend you call for advice. An agent is a friend you give your keys to, so they can go and do the errand for you. That is far more useful, but you have to trust what they do while you are not watching. Anthropic describes the difference like this. In a fixed workflow, the steps are written in code ahead of time. In an agent, the model directs its own process and decides which tools to use. That makes agents good at open-ended problems, where nobody can say in advance how many steps it will take. The price is cost and risk. An agent often uses many more words, and so more money, than a single answer. And because each step builds on the last, a small mistake early on can grow into a big one. That is why builders test agents in a safe sandbox first and limit what they may touch, for example read-only access to a bank account. **Why it matters to you:** more and more AI products are agents, from coding tools to shopping helpers. Before you let one act for you, ask what it can reach, what it can change, and who is responsible if it gets something wrong.Two more from today
What is quantization?How AI models are shrunk to fit on smaller computers, and what it costs.
Quantization makes an AI model smaller by storing each of its numbers with less detail. A model that needs 16 bits per number can often get by with 8 or 4.
The model then needs much less memory and often runs faster. The price is a small loss of accuracy, which grows the further you compress.
What is a benchmark?How AI models are tested and scored, and why a high score is not the whole story.
A benchmark is a fixed test for AI models: a set of tasks plus a way to score the answers. Every model takes the same test, so you can compare the scores.
Scores are useful but limited. A lab often runs the test on its own model, and test questions can leak into the data a model learned from.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak