Tech, in brief

France's National Assembly voted 11 to 10 for a tax on internet traffic on 30 October. [1]

Firms that send data to French users would pay €0.055, about 6 US cents, per gigabyte. [1]

The vote adds the tax to next year's budget bill, but more votes are to come. [1]

If it passes, the biggest bills land on the few firms sending most of the data. [1]

Economists call it a per-unit tax: the bill grows with gigabytes sent, not with profit. [1]

France's budget would gain about €10 billion a year, if the author's estimate holds. [1]

Nobody yet knows whether users would end up paying: the two sides disagree. [1]

The amendment must still survive further votes in France's parliament before it becomes law. [1]

France's lower house voted 11 to 10 to tax Google, Netflix, Meta and others for the data they send to French users, though France's government opposes it and more votes remain. [1]

A big round number, like €10 billion a year, makes a first vote by a single seat sound like a law already passed.

Would be billed €0.055 for each gigabyte sent to French users, up to 15% of their French sales. [1]

Whether France's traffic tax survives the rest of the budget process.

The vote was 11 to 10, France's government opposes it, and further parliamentary votes are still to come. [1]

Whether the cost would reach users.

The finance committee's lead member says firms would pass it on; its author says free apps cannot. [1]

How the traffic would be counted without watching users closely.

Minister David Amiel doubted it can be traced to each firm without complicated monitoring. [1]

OpenAI, which makes ChatGPT, now aims to raise up to $20 billion in its new round. [3]

It first sought $30 billion; the Financial Times reported the $10 billion cut on Friday. [2]

So far it has secured about $11 billion, The Wall Street Journal reported. [4]

Interest from some large investors was tepid, the Journal said. [4]

Venture rounds can close in stages: money signed by a deadline, then more later. [2]

Investors from the United Arab Emirates, including the state-backed fund MGX, should be the largest backers. [2]

OpenAI will next ask family offices, which manage one rich family's money, and wealthy individuals. [2]

The round values OpenAI at $1.25 trillion, down from the $1.4 trillion it first pitched. [2]

AI shares fell all week; Alphabet's dropped about 12% on Thursday, its worst day on record. [5]

OpenAI has told investors it plans about $600 billion of computing spending by 2030. [6]

Anthropic, its main rival, plans to sell its shares to the public next month. [2]

A later offer will let wealthy individuals buy at the same $1.25 trillion value. [2]

The round is still open, so the final size could change, CNBC's source said. [3]

Elsewhere, OpenAI has secured only about $11 billion of the $20 billion it now wants, as investors ask whether AI spending will pay back. [2][3]

How much OpenAI's round finally raises.

The round is still open, the first close is expected below $20 billion, and OpenAI did not immediately respond to requests for comment. [3][4]

Pope Leo XIV called on Friday for artificial intelligence to be "disarmed". [7]

He urged states meeting in Geneva to take steps toward banning lethal autonomous weapons. [8]

He means weapons that choose targets and use force with no person deciding. [7]

Decisions on life, freedom and the use of force must stay under human oversight, he said. [7]

He named two choices never to hand to machines: a target list and a border database. [7]

"Machines are tools, not persons," he said, rejecting legal rights for AI systems. [8]

In 2024 Pope Francis told the G7 that no machine should ever choose to kill. [7]

States meet under a UN weapons treaty in Geneva from 16 to 20 November. [7]

Arms-control experts call the demand human control: a person answers for each use of force. [8]

He asked for shared standards and independent oversight, not more statements of principle. [7]

Not all of the AI industry agrees: Mistral's boss, Arthur Mensch, has rebutted him. [7]

Why a machine can be handed part of a decision, not all of it

A decision has steps: gather facts, list options, choose, act

Each step can be done by a person or by a machine

Engineers rank the split in levels, from a person doing it all to a machine acting alone

In between: the machine suggests, or acts unless stopped, or acts and reports

The higher the level, the less time a person has to catch a wrong move

So designers keep a person at the steps that cannot be undone

The choice is rarely all human or all machine.

A simple five-step version of the levels of automation. The steps are a ranking we made up for this picture, not a measurement.

The agent chose the therapy and the brain area, but every step that could not be undone waited for a vet's yes.

He asks that the choice to use force stay at the lowest level, where a person decides every time.

Dazzle's agent acts on its own until it flags a task it cannot do, then a person takes over.

Its fear is the top level: a model that copies itself onto devices with no person at any step.

A person who only approves can stop checking. If the machine is right a hundred times, the hundred-and-first wrong move may get approved too.

The shopper, the vet, the soldier and the teenager on ChatGPT each sit at a different level, often set by someone they never met.

Exin Therapeutics, a biotech start-up, says an AI agent did brain surgery on a mouse. [9]

The goal was to reduce seizures in a mouse bred to have epilepsy. [9]

The agent chose one gene therapy from 19 options listed in a spreadsheet. [9]

It picked the brain area and worked out where to inject from 3 research papers. [9]

Then it drove robot arms that injected 250 nanolitres into the mouse's brain. [9]

A licensed vet had to approve every step that could not be undone. [9]

The vet also held a switch that cut power to every motor. [9]

The team used Anthropic's Sonnet model after its newer Opus 5.5 refused the job. [9]

Gabriel Ocana-Santero, the firm's chief executive, says he knows of no agent doing this before. [9]

So far the account comes from the company's own thread on X. [9]

The mouse recovered well and was filmed moving around its cage. [9]

The agent chose the therapy and the brain area, but every step that could not be undone waited for a vet's yes.

Some AI helper start-ups send the tasks their agents fumble to human workers, CNBC reports. [10]

AI agents are programs that act for a person online, such as booking or buying. [10]

They still struggle with phone calls, doctor's appointments and booking local services. [10]

Marissa Mayer, a former Yahoo boss, runs Dazzle AI, which passes some jobs to contractors. [10]

One Dazzle user's agent mixed up window glass with glass for a tabletop. [10]

Dazzle's system flagged the mix-up and handed the task to a person. [10]

Duckbill, another start-up, sells agent makers a way to hire people for tasks. [10]

Its boss, Meghan Joyce, says the model is "not dissimilar from Uber": small, simple tasks. [10]

Agents like Meta's Muse and OpenAI's Dots still hit roadblocks that need a person. [10]

When an agent calls a company, the voice on the line may be a hired person. [10]

Dazzle's agent acts on its own until it flags a task it cannot do, then a person takes over.

Get paid work finishing the tasks agents fumble, sold to agent makers through firms like Duckbill. [10]

Executives at about 50 companies told investors and Wired that AI agents bring little business yet. [17]

T-Mobile, a US phone network, said "We haven't seen anything of significance" from agents. [17]

eBay, the online marketplace, called agent activity very, very small. [17]

Hilton, the hotel chain, called bookings from agents insignificant. [17]

Amazon's boss, Andy Jassy, said few customers shop through outside agents yet. [16]

He expects Amazon's own helper, Alexa for Shopping, to be the main way people shop there. [16]

Amazon says more than 400 million customers used it in the past year. [16]

Amazon blocked Meta's Muse agent from its site about a month ago. [16]

Meta says more than 10 million people in the US and Canada have set up Muse. [16]

Investors had feared agents would weaken shoppers' loyalty to brands and start price wars. [17]

Marketers call it brand loyalty: shoppers return to a shop they trust out of habit. [17]

Amazon gains by keeping shoppers in its own helper, which knows their address and payment details. [16]

Samsung hosted OpenAI's boss, Sam Altman, for talks with about 50 of its presidents on Saturday. [13]

They met at Samsung's training centre in Yongin, south of Seoul. [13]

Bloomberg says the talks aimed to deepen ties on memory chips as shortages grow. [14]

DRAM, the memory chip Samsung makes, is the fast memory a computer works in. [13]

Two days earlier, Altman said it was "quite terrible" for memory to sit in home computers. [13]

He said home computers use that memory about 5% of the time. [13]

Engadget notes OpenAI has tied up about 40% of the world's memory output through deals. [15]

Samsung has worked to supply memory for Stargate, OpenAI's data centre project. [13]

About 2,300 Samsung executives have taken AI training so far. [13]

Memory bought for AI data centres has pushed up prices of phones and PCs, Engadget says. [15]

Economists call this allocation: deciding who gets a good that is in short supply. [13]

Samsung published nothing Altman said and named none of its executives in the room. [13]

Are courted by OpenAI while memory is short, which Altman says will leave each with about $100 billion of free cash a year. [13]

Pay more as memory goes to AI data centres, Engadget says, with OpenAI holding about 40% of output through deals. [15]

What, if anything, Samsung and Sam Altman agreed.

Samsung published nothing Altman said and named none of its executives. [13]

OpenAI said on Friday ChatGPT will stop giving teenagers practical information about weapons. [12]

The limit holds even when a teen's question looks harmless. [12]

For all users, ChatGPT will give less detail about past mass attacks and their attackers. [12]

It will not reconstruct an attack, such as which weapons were used. [12]

ChatGPT can now end a chat that shows someone fixing on harming another person. [12]

If it ends several such chats, OpenAI may pause the account for a time. [12]

OpenAI says it made more than 200 referrals to police in the third quarter. [12]

Eight US states have sued OpenAI over harm to children. [12]

Its first Community Safety Report comes out in early 2027. [12]

An age-guessing system decides who gets the teen version, using signals such as behaviour. [12]

Safety designers call these guardrails: limits that hold whatever reason a user gives. [12]

Adults wrongly placed with teens must prove their age through Persona, an outside checking firm. [12]

Anthropic's Frontier Red Team, which tests its AI for dangers, published new research. [11]

It says small AI models may be months away from copying themselves across poorly protected devices. [11]

That could create botnets: networks of hijacked devices that are hard to kill. [11]

In 2016 the Mirai botnet, built from routers and cameras, knocked big websites offline. [11]

Chris Canal, whose firm EquiStamp checks AI models, said it came earlier than he expected. [11]

Michael Bailey, a Georgia Tech professor who studied Mirai, doubts it was shown convincingly. [11]

He says Anthropic's test ran in a simulated network where breaking in was trivial. [11]

Bailey says AI could cut the human work of adapting an attack to each new device. [11]

Security researchers call a self-spreading program a worm: it copies itself with nobody clicking. [11]

AI companies have played up dangers before as a form of marketing, Engadget notes. [11]

Its fear is the top level: a model that copies itself onto devices with no person at any step.

Whether AI can really spread itself across devices.

Anthropic says small models may be months away, but a Georgia Tech expert says its test used a simulated network and a trivial break-in. [11]

Developers say Anthropic's Claude now refuses to rebuild specific existing video games. [22]

The rule users shared says: "Don't build a recreation of a specific existing game." [22]

Complaints on Claude Code's GitHub page report refusals partway through ordinary game work. [22]

Users also report refusals of reverse-engineering: studying how finished software works inside. [22]

Some game makers have faced scams threatening to clone their demos into full games. [22]

On 6 October Anthropic limited its models' advanced hacking skills to approved security groups. [22]

Game makers who feared clones gain protection, but honest developers lose working time. [22]

Anthropic has not confirmed the rule, so exactly what is blocked is unclear. [22]

Hit refusals partway through projects under a rule Anthropic has not confirmed. [22]

What exactly Claude now refuses to build.

The rule text users shared has not been confirmed by Anthropic. [22]

Apple's chief executive, John Ternus, says Apple does not want to ship AI for AI's sake. [18]

Ternus became Apple's chief executive at the start of September. [18]

Apple has been called behind on AI after it delayed a stronger Siri, its voice helper. [18]

That stronger Siri finally launched this summer. [18]

"I really like where we're at," Ternus told Wired's Steven Levy. [18]

He said people should not have to think whether AI or older software did it. [18]

Levy asked whether AI agents could make apps unnecessary; Ternus said making products is deeply human. [18]

AI data centres have become an issue in the US midterm elections, TechCrunch reports. [20]

They could sway some races in the vote on 3 November, it says. [20]

Data Center Watch counted almost $200 billion of projects delayed, stalled or blocked by opposition. [20]

That was in the first half of 2026 alone. [20]

Five states, Pennsylvania, Georgia, Texas, New York and Virginia, weigh stricter data-centre rules. [20]

Hundreds of new data centres are planned in Virginia alone. [20]

In Brooklyn, a crowd booed the question "Who here is pro-AI data center?" [20]

Zoning rules, which decide what is built where, can hide a site's purpose until too late. [20]

Residents fear deals struck with local officials before anyone can object. [20]

Nobody yet knows how many races data centres will actually swing. [20]

Amazon has switched off voice commands for the AUX input on several older Echo speakers. [19]

The AUX input is a socket for plugging in sound from a phone or CD player. [19]

Commands like "Alexa, play from line in" no longer work. [19]

Four models are hit: the third and fourth Echo, the Echo Plus and the first Studio. [19]

Amazon told The Verge it had told the very small number of people who used it. [19]

The cable still works, and the Alexa app can still switch the input. [19]

Speakers bought as early as 2018 lost a feature their owners had used. [19]

Amazon did not say why it removed the commands. [19]

Lost voice control of the AUX socket on speakers bought as early as 2018, with no reason given. [19]

Ford, Rivian and GM are using 3D printing for factory tools and some car parts. [21]

Engineers call it additive manufacturing: a printer adds material layer by layer from a design. [21]

Rivian has spent over $5 million on printers, and about $1.7 million a year on materials. [21]

Its head of 3D printing says the machines usually pay for themselves within a year. [21]

Printing cuts the time to make a part from weeks to days, or even hours. [21]

Ford put a vending machine of printed parts beside its F-150 pickup line near Dearborn. [21]

Printing in-house cuts what carmakers pay outside suppliers for tooling, Rivian says. [21]

Rivian puts its savings only "well in the millions", giving no exact figure. [21]

Ford workers have designed some of the printed hand tools themselves. [21]

Rivian's 3D printers cut the time to make a factory part from weeks to days or hours, and usually pay for themselves within a year.