The top 40 jobs most at risk from AI

Microsoft researchers have studied hundreds of thousands real-world interactions between workers and Microsoft’s Copilot AI assistant, and have published a list of the jobs most at risk from AI. Translators and interpreters top the list of vulnerable occupations, with historians, sales reps, and customer service representatives following close behind. In general, knowledge workers are more vulnerable than workers whose jobs involve a physical skill, like plumbers, electricians, and phlebotomists. Now all of us who spent years in school to master a field of knowledge are reviewing our life choices.

Microsoft researchers reveal the top 40 jobs most vulnerable to AI

The AI building boom is boosting a sluggish economy

Economists are talking about a startling new statistic: in the first half of 2025, spending on AI infrastructure was responsible for more growth in US GDP than consumer purchases. Consumer spending, comprising 70% of the US economy, is the usual engine for growth, but investments in datacenters and related infrastructure is so massive that it has softened the effect of decreasing consumer spending. Expenditures by just 5 large tech companies - Microsoft, Apple, Meta/Facebook, Amazon, and Google - totaled approximately $100 billion in the past quarter alone, and they project continuing to spend at that rate for the foreseeable future. This massive private-sector investment is boosting a US economy that is otherwise starting to wobble due to continued inflation, tariff fears, and various other uncertainties stemming from changes in Washington.

The AI building boom is boosting a sluggish economy

This guy got a $1.5 billion job offer - and turned it down

Australian-born computer scientist Andrew Tulloch was offered a compensation package potentially worth $1.5 billion over the next 6 years to rejoin the AI development group at Meta/Facebook. He turned it down flat. Tulloch spent the early part of his career, from 2012 to 2023, at Meta/Facebook, during which time he contributed heavily to development of a computer language optimized for AI called PyTorch. PyTorch is now quite popular among AI system developers. In 2023 he joined OpenAI, where he participated in developing some of their most advanced models. Recently, he left OpenAI along with Mira Murati, OpenAI’s Chief Technology Officer, to form a new startup called Thinking Machines Lab. TML has just secured $2 billion in initial funding, valuing the company at $12 billion. Not bad for a new company with no product. Meta CEO Mark Zuckerberg zeroed in on TML, offering king’s ransoms to multiple top employees. Tulloch, a co-founder of TML, was the big prize, and the bidding for his services reached insane heights, eventually topping a potential $1.5 billion over 6 years. Tulloch, like every other TML employee wooed by Zuckerberg, declined the offer. For AI researchers at the pinnacle of the field, riches have become a given. They are more motivated by pursuit of the goal of artificial intelligence as smart as or smarter than humans, and the chance to work with congenial colleagues that they respect, without having to answer to problematic bosses like Mark Zuckerberg.

This guy got a $1.5 billion job offer - and turned it down

China’s lead in open source AI is forcing the US to compete there, too

Last week’s AI Action Plan released by the White House had one somewhat surprising section, as I remarked at the time. The plan, mostly crafted to be big-tech friendly, embraced US leadership in open-source AI models, in addition to the big-tech closed models from the likes of OpenAI, Google, and Anthropic.

Why, you may ask? (I did ask, last week.) Because China has been releasing open source models (like DeepSeek) that have been beating the bejesus out of the US open source models, and challenging the performance lead of the big closed models.

Open source models allow anyone knowledgeable about AI to copy and tweak their inner workings for free, which enables a tremendous amount of innovation very rapidly. If all that innovation is happening on Chinese models, then they will advance faster, and in addition, China will gain a great deal of soft power if they become the foundation of the international AI technical stack. Just like the internet runs on the open source operating system known a Linux, open source AI technology may become the base layer in the AI technical future.

That’s why an America-First, Big Tech-friendly administration felt it had to support US leadership in open source as well.

China’s lead in open source AI is forcing the US to compete there, too

Amazon invests in Fable Studio, AI that can create full TV episodes. And it wants YOU to star in them.

Amazon has recently invested in AI startup Fable Studio, whose flagship product, known as Showrunner, can create full-length TV episodes from scratch, or based on an existing series, like South Park. The technology is apparently so user-friendly that anyone can create a 22-minnute episode just by typing in a few descriptive phrases. The resulting episodes have coherent plots, character development, and convincing dialogue. You can insert yourself as a character in the episode. The CEO of Fable Studio foresees a future in which everyone is producing their own entertainment, often riffing off of, remixing, or extending their favorite commercial series, and sharing the result with others.

Amazon invests in Fable Studio, AI that can create full TV episodes

Google DeepMind releases highly detailed digital map of Earth

Google’s DeepMind AI research lab has just announced AlphaEarth, an AI model that integrates terabytes of information streaming daily from satellites into a coherent and extremely detailed up-to-date digital map of our planet. Early uses of this system include MapBiomas, a project in Brazil that attempts to understand agricultural and environmental changes in the country, which includes the fragile but critical ecosystem of the Amazon Rainforest.

Google DeepMind releases highly detailed digital map of Earth

Now anyone can be a genetic engineer with CRISPR-GPT

Researchers from Stanford, Princeton, and UC Berkeley have created an AI system that automates the design and data analysis of gene editing experiments using the Nobel Prize-winning CRISPR gene-editing method. Composed of multiple specialized AI agents, with a chatbot front end, it allows even people with no genetic engineering background to get actionable experimental designs that can be carried out by very junior personnel. The authors published their findings in Nature, calling their system “an AI co-pilot in genome engineering.” Systems like this, which are proliferating, point up the productivity and innovation that can be unleashed by using AI as a user-friendly interface to complex technology.

Now anyone can be a genetic engineer with CRISPR-GPT.

Anthropic finds AI models can take on distinct personas, and they can be “vaccinated” against evil behavior

Researchers at AI startup Anthropic have found that AI models can take on distinct “personas.” More, they have found that each persona can be identified by a distinctive pattern of activation of the model’s “neurons”, or parameters. Watching for this distinctive pattern can help AI safety personnel identify in real time when a model is acting out an undesirable persona, such a sycophant that always agrees with the user, or a persona that cooperates with or even encourages antisocial behavior. Going even further, Anthropic researchers are finding ways to steer a model toward or away from personas, and even to give the model some corrective steering during training that “vaccinates” that model from assuming undesirable personas.

AI models can assume personas, and can be “vaccinated” against bad behaviors

Robot dogs help run world’s first unmanned wind farm in China

A large 70-megawatt wind farm in Northern China is running flawlessly, despite the fact that no human has visited the site for almost a year. Instead, fleets of aerial drones and robot dogs are operated and monitored remotely by workers at an operations center 30 km. away. This offsite monitoring of onsite robots has reduced operational costs and improved employee safety, and plans are in the works to roll out the system to more wind farms in the next year.

Robot dogs help run world’s first unmanned wind farm in China

Florida is using robot bunnies to catch Burmese pythons

The Florida Everglades are teeming with gigantic Burmese pythons, among the largest snakes in the world, growing to up to 16 feet long. Native to Burma (no surprise), these reptiles became popular exotic pets in the 1970s and after, and would from time to time escape to the wild. (Would you say no to a 16-foot snake trying to get away?) Finally, in 1992, Hurricane Andrew, a Category 5 storm, destroyed a Florida python breeding farm, and a large number of the snakes escaped to the Everglades and established a breeding population. Pythons have devastated native wildlife, including raccoons and opossums, and Florida has been trying to eradicate them for decades. Now the robot bunnies are coming to the rescue! One Everglades district has commissioned AI-controlled solar-powered robot rabbits, which are engineered to look, smell, and act as much as possible like native rabbits, which are a favorite python prey. When a robot bunny’s sensors detect a nearby curious python, it sends a signal to a monitoring station with its location, so that wildlife control personnel can come and remove the errant reptile. There is as yet no word as to how many robo-bunnies have been lost in action.

Florida is using robot bunnies to catch Burmese pythons

Stanford’s AI virtual scientists develop innovative COVID treatment

Stanford researchers have developed an AI system that uses multiple cooperating AI agents as “virtual scientists” to come up with novel hypotheses and design experiments to test them. The AI system is modeled as closely as possible on the structure and operations of an actual well-functioning research lab, but all the roles are performed by individual AI agents. In a test of the system, the virtual scientists were tasked with finding a novel way to fight COVID. The virtual scientific team quickly decided to develop a treatment not based on antibodies, but on so-called “nanobodies”, small fragments of antibodies that can be readily designed and produced in a lab, as opposed to the much more complex processes required for production of antibodies. When the nanobodies were put to the test in a real lab environment, they were found to be broadly effective against all known variants of the COVID virus. The time from problem statement to tested solution was days, not weeks or months. The Stanford team are excited by the acceleration of scientific progress that virtual scientists may bring.

Stanford’s AI virtual scientists develop innovative COVID treatment

“Level 5” (autonomous) AI in medicine may start with prescribing

Self-driving cars are classified from Level 0 (no driving automation) to Level 5 (fully autonomous. Using a similar classification for AI in health care, what would Level 5, fully autonomous health care AI operation look like? It is unlikely to look like the hologram doctor in the TV series Star Trek: Voyager, at least at first. The article linked below argues that prescribing medications may be the first area where humans progressively turn over control to AI. Early on, we may use AI to prepare medication refills for review and signing by a physician. This is a task that is already routinely performed by nurses or pharmacists in many health systems. Later on, we may delegate protocol-driven prescriptions like treating a UTI in an adult female, which are routinely started on empiric grounds prior to any results from urine culture. Later still, we may cede protocol-driven up-titration of existing medications, such as increasing lisinopril for poorly controlled blood pressure. As we gain more experience with the abilities and limitations of AI systems, we may come to a day when almost all prescriptions are being decided on by AI. This sort of step by step roadmap may be able to be translated to many other tasks in medicine, even surgery and other procedures. The holograms are optional.

Level 5 (autonomous) AI in medicine may start with prescribing