Is AI about to become self improving?

The electric chair was invented by a dentist named Alfred Southwick, who designed it by modifying his own dental chair. He watched the first execution in 1890 and announced to the room “this is the culmination of ten years of work and study.” He was thrilled.

When he needed design advice he approached Thomas Edison, who told him to use Westinghouse’s technology instead, specifically to make people associate Westinghouse with death. A dentist invented it, Edison turned it into a corporate weapon, and somehow the thing that makes everyone most uncomfortable is still the chair.

What’s in store:

Anthropic’s AI might be getting too big for its boots.
AI hackers are coming for the national grid.
Mindstream Picks: Meta is winding down support for Quest 2 and Quest Pro.
The only AI Agent starter pack is here!

Read Time: 5 minutes

ANTHROPIC

A very expensive game of Frankenstein

Anthropic has warned that its AI models could soon become capable of helping build more powerful versions of themselves, potentially speeding up AI development.

In an updated financial filing, the company highlighted the risks of recursive self-improvement (RSI), where AI helps make itself smarter.

Supporters believe this could lead to major scientific discoveries and better productivity, while critics worry it could make AI harder to control.

In brief:

Anthropic uses its models to help develop more advanced systems.
Rapid progress could make it harder to test new models and spot potential risks.
Anthropic says Claude won't be trusted to build future AI models entirely on its own without human oversight.

A horrible day to have reading comprehension

Anthropic admitted that it has previously missed major improvements in its models until after they happened.

It now believes future models could soon reach a point where AI speeds up its own development, making it harder for safety checks to keep up.

However, none of its current models have reached that point yet.

The company is now considering changes to its safety rules to make sure humans stay in control as AI becomes more advanced.

Claude might be the first nepo baby to give birth to itself! -MG

Anthropic says AI could soon help build smarter AI. What do you think?

Excellent, I've seen this film

Please unplug the entire building

Vote for live results and see results + opinions from yesterday at the bottom of the email.

AI AGENTS

Build Your First AI Agent Today With This Starter Kit

Everyone's building AI agents. Most people are stuck on step one: which tool do I even use? This kit skips the rabbit hole and hands you a ranked breakdown of 7 agent builders — scored by cost, ease, and power — plus real data from 300+ business leaders on how they're actually deploying agents right now.

Get instant access to:

Survey data from 300+ biz leaders on real-world agent use
7 ranked builders scored by cost, ease, and power
Budget and skill-based guidance for choosing your builder
A no-code revenue agent template ready to deploy

Three resources. One afternoon. One agent working for you.

Get the Starter Kit →

Riddle Me This

What is the only word that can contain hundreds of letters?

Find the answer at the bottom of the email!

How I AI

I am using it to do virtual setup staging and segmentation in Aligners biomechanics of Orthodontics in dentistry.

Dr Lalita, India

Tell us how you’re using AI! →

AI CYBER SECURITY

The hackers have upgraded their evil empire

The UK’s National Cyber Security Centre (NCSC) has warned that AI-powered malware is highly likely to be used in a cyberattack on essential services, including hospitals, airports and power stations, by 2028.

The agency said AI could make hacking easier by helping attackers find targets, spot security weaknesses and create harmful software.

The warning focuses on ransomware attacks, where hackers lock computer systems and demand payment to restore access.

Previous UK victims include Royal Mail, Capita and several local councils.

Here's what else matters:

AI could help hackers carry out ransomware attacks against essential UK services.
British military networks face daily cyberattacks, including attacks linked to Russia.
Experts say governments and businesses need advanced AI tools to find and fix security weaknesses.

The national grid would like to unsubscribe

The NCSC also warned that Russia and China are expected to use AI regularly in cyberattacks.

Its chief executive, Richard Horne, said the UK could see a sharp rise in attacks in the coming months as AI makes it easier to find weaknesses in computer systems.

The warning follows recent government advice for households to keep three days' worth of food and water in case of emergencies, including cyberattacks.

However, AI could also help organisations find and fix security problems before hackers exploit them.

Recent cyberattacks on supermarket refrigeration systems across Europe have also raised concerns about how easily everyday services could be disrupted.

Cyber security - it’s the hot new industry everyone is talking about. - MV

Mindstream Picks

Space: Betelgeuse, one of the brightest stars in the night sky, has been growing dimmer, with astronomers investigating how its natural pulsations, dust clouds and possible companion star contribute to its changing brightness.

Business: Virgin Media O2 is planning thousands of job cuts and pausing fibre upgrades as part of an £800 million cost-saving drive by 2029, with reports suggesting up to 4,000 roles could be affected.

Music: Sinéad O’Connor’s daughter Róisín Waters has hinted that previously unreleased music could be coming soon, as the singer’s family works with Chrysalis Records to celebrate her legacy.

Don’t Miss: Meta is ending major software updates for its Quest 2 and Quest Pro VR headsets, with the final OS update expected by March 2027. Security patches and critical fixes will continue until the end of 2027, as Meta shifts its focus to the Quest 3, Quest 3S, and upcoming VR Glasses.