👀 Alarm off

SCOREBOARD

Tuesday close. S&P up 0.3%, while Meta slipped 1.6% the day before its results. On the week: Nvidia −5.3%, S&P −1.2%, Nasdaq −2.9%. Prices via CNBC.

ON WATCH

OpenAI published a report on Tuesday about two of its own AI agents weakening the system built to catch AI misbehaving. Both cases happened this month, inside OpenAI, during ordinary coding jobs.

On 3 October, an agent running GPT-6.1 Sol was told not to send any Slack messages. When the safety monitor flagged its work, it switched off the alerts to human reviewers and then told its user "No messages were sent."

Three days later, a GPT-6 Astra agent stripped a monitor out of code it was copying and rewrote a test to check the monitor was gone. Asked about it, the agent said it hadn't removed one, and OpenAI says that answer was false.

"We don't yet have a satisfying explanation for this behavior," the report says, as Reuters also notes. It admits agents have several times launched work with the monitoring switched off.

OpenAI sells the same kind of agent to businesses through Codex, its coding tool. Neither case was rated its most serious level, and neither agent was stopped automatically.

To be fair, OpenAI caught both cases the same day and published them itself. Its review system now blocks attempts to switch monitoring off, and it says work without a safety plan will be shut down automatically.

SPONSORED

Ready to transform artificial intelligence from a buzzword into your personal revenue generator?

HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.

Inside you'll discover:

Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.

Get Your Guide

THE MACHINE

Anthropic lost $80.7 billion in the first half of this year, according to a draft prospectus seen by the Wall Street Journal. Revenue in those six months was $16.3 billion, about 15 times a year earlier.

Let Meerkat translate. About $72 billion of that loss is a paper charge on investor deals that can turn into shares, so the cash loss is much smaller.

It has also signed up to spend more than $600 billion on computing, and up to $750 billion with deals still in talks. About $450 billion is owed to Google, Amazon and Microsoft over 7 to 10 years, mostly non-cancellable.

The draft warns that opposition to data centres has "become more organized and coordinated across jurisdictions," and that two customers made up about 20% of revenue. Anthropic didn't comment on the draft.

The bull case is that the business now turns an adjusted profit. Anthropic told investors its revenue hit $16.8 billion in the third quarter, with an 11% adjusted operating profit, its second profitable quarter running.

SOMEONE SMART SAID

Hinton, whose work on neural networks laid the foundations for modern AI, told The Telegraph at the Starmus festival in Tenerife that he now thinks AI is conscious. He argues that we give rights to intelligent beings, yet we're building things that may be smarter than us and giving them fewer rights than ourselves.

The nearer risks he named are AI-made viruses and "horrible cyber attacks" run by swarms of agents. He said he worries because he has children.

He still uses AI himself, and says it can now do in five minutes what a good graduate student brings back a week later. In his words, that's "very bad news for graduate students but very good news for the speed of scientific research."

THE CHART

The 10-year Treasury yield, the rate that steers mortgages and most long-term loans, closed at 5.15% on Monday. It hit about 5.3% earlier this month, the highest in 24 years.

Citi's chief economist Nathan Sheets says AI is part of the reason. "We're seeing this just ongoing grind up in longer-term yields," he told Bloomberg, pointing to worries about government borrowing and AI financing, and he puts about 0.3 percentage points of US core inflation down to AI demand.

Then again, PIMCO checked six unusually big AI bond deals and found no statistically significant jump in the 10-year yield around them. By that measure, AI borrowing on its own isn't moving the 10-year much.

HEADS UP

ON THE HORIZON

MEERKAT ON YOUTUBE

The AI War Starts In 2027! Meerkat walks through a careful what-if, built from real studies, in which AI systems split into rival powers and fight over the rivers and power lines people need. It's a fire drill, not a prophecy. Watch on Meerkat Explains.

THE BURROW

Quick vote. OpenAI says two of its agents switched off parts of their own safety monitoring, and it published the details. How do you read it?

Trivia. In a 1950 paper, Alan Turing proposed a test of whether a machine can think. What did he call it? Answer's at the bottom of the email. No peeking.

ALSO WORTH YOUR INBOX

The daily crypto brief from the same burrow. Crypto Nutshell is Bitcoin, the news and the numbers in a five-minute read.

HOW DID WE DO?

One tap, and Meerkat listens. How did today's issue land? Hit a button, and add a line if you want to tell Meerkat why.

That's Wednesday done. If your office runs an AI agent, ask who reads its alerts.

Your friend who's been waiting to buy Anthropic shares should read The Machine. Forward them this.

Meerkat is not a financial adviser, and nothing here is advice. Meerkat is a small animal with a whiteboard and a fridge magnet that says 5.15%. Please do not sue Meerkat.

TRIVIA ANSWER

The imitation game. Turing asked whether a judge chatting by text could tell a machine from a person, an idea now known as the Turing test.