Drift Notes: What’s happening in AI this week?

It was a rough few days for AI stocks. On Thursday, the Nasdaq dropped 5.5% in a single day, and Nvidia’s shares fell more than 9%. The trend spread overseas on Friday, with Japan’s and South Korea’s main indexes each closing down about 6% (South Korea’s market even paused trading briefly to calm things down). Investors say they’re nervous about two things: possible tougher rules on AI and data centers after the US midterm elections, and whether all the money going into AI will actually pay off. Things steadied a little on Friday, with the Nasdaq up 1.1%, and it’s still up about 4.6% for the year.

A lot of the AI boom runs on investors believing the spending will pay off, and this week showed how quickly that confidence can wobble. One rollercoaster of a week doesn’t tell us whether this was a passing scare or the start of something bigger. For AI trainers, the thing to keep an eye on is whether companies keep investing in AI, since projects tend to follow demand.

Prosecutors say three people used AI to pretend to be company executives and lawyers, and walked away with about $129 million. They would email a company’s accounting team posing as the CEO, then set up a call where AI copied the CEO’s voice (and sometimes face) before handing the employee off to a fake attorney. That “lawyer” allegedly read from a handwritten script titled “Lawyer Call.” Targets included hospitals, a bank, and a pet food company, and the group’s nicknames include “Rocky Balboa” and “Superman Cash.” The three have been charged but not convicted, and extradition requests are pending.

It’s a sign of how convincing AI voices and faces have become. A familiar voice on a call is no longer proof of who’s really there, so double-checking payment requests through a separate channel matters more than ever. For AI trainers, it’s a reminder of why careful testing for misuse matters: the same technology behind useful tools can be turned against people, and those who know what to look for help make the safeguards better.

A new UK research project called BRAINIAC, led by the University of Exeter, aims to use AI to spot dementia earlier. The AI will combine brain scans, genetics, blood tests, and digital memory-and-thinking tests into one clear picture of a person’s brain health. Half of the £20 million comes from a UK public research funder and half from industry, charity, and National Health Service (NHS) partners, and the project will involve 11 universities and 10 NHS sites over six years. The team wants to make diagnosis 30% to 50% faster and flag more than 40% of the people it assesses as at risk before symptoms appear, with the first clinics opening in 2029.

About 982,000 people in the UK live with dementia, a number expected to reach 1.4 million by 2040, and today’s treatments only help when it’s caught early. That’s why early detection is the focus and this project is one of the first to test procedures in real life. For AI trainers, medical AI like this usually depends on expert human review to check what the model gets right and wrong, so clinical knowledge is part of what makes it trustworthy.

Researchers at MIT and NYU have built a system called FLEX that helps robots read between the lines of what we tell them. In a simulation, a robot arm carrying a cup of coffee was told, “Don’t spill on the laptop!” and it also steered clear of a nearby phone without being asked. It seems the AI language model behind it quickly learned that liquids and electronics don’t mix. The team tested it on three simulated tasks, serving coffee, moving a candle, and carrying a bowling ball, and the robot learned to avoid electronics, flammable objects, and fragile things from a sentence or two. The process takes about five seconds, and the study will be presented at the Conference on Robot Learning this month.

Everyone’s home is different. No robot arrives knowing what matters to you, and the researchers want robots that can pick that up from a quick conversation instead of a long list of rules. It’s still early to tell whether this system extends to real life, since these were virtual tests, not robots in real kitchens. For AI trainers, it’s a neat example of why human language matters: the robot’s common sense comes from the input and background knowledge it was trained on.

Kew Gardens’ latest State of the World’s Plants and Fungi report — yes, that’s a real and wonderful read — put together by more than 400 scientists from 40 countries, says technology is changing how we protect nature. AI is helping identify tricky plants and even spot species that are new to science. Specimens can now be analyzed on a much bigger scale, and digital tools are speeding up how they’re cataloged and shared. Kew alone has digitized 7.4 million plant and fungi specimens, which would stretch almost 3,000 km if laid end to end.

The report is upfront about the limits: AI identifications still need to be checked by scientists, and the technology doesn’t fix what’s causing biodiversity loss in the first place. It also has its own footprint, since the data centers behind AI can use a lot of electricity and, in some places, water for cooling. For AI trainers, it’s a good example of why expert review matters, because a botanist’s knowledge is what confirms whether an AI’s plant identification is right.