OpenAI wants to own e-commerce

OpenAI’s current business model appears to be raising billions of dollars from investors, and converting those billions into massive, electricity-guzzling data centers. Some day, they will need to transition to an actual money-making business. CEO Sam Altman has made a number of announcements lately that give a sense of what the company’s plan is for achieving this profit turnaround. OpenAI plans to own the internet - or at least the profitable parts of it.

The company announced yesterday an alliance which will allow ChatGPT users to order anything from Walmart and pay for it without ever leaving their favorite chatbot. This “Instant Checkout” function, which already includes Shopify and Etsy, will be extended to more and more vendors over time. And for every purchase made through their chatbot, OpenAI will, no doubt, collect a fee.

Besides trying to displace Amazon as the “Everything Store”, OpenAI is aiming to disrupt social media as well. The company’s wildly popular Sora 2 text-to-video app has a number of social media features that make sharing videos fun, even addictive. In addition, the company has announced that it is working on a friendlier, more companionable chatbot persona. This will coincide with age-verification controls, now mandated by California (see story below) to prevent teens from getting seduced into unhealthy relationships with a chatbot.

Age verification controls unlock one final frontier of e-commerce - unrestricted content. Yes, this probably means that adults will be allowed to be as sexist and racist as they want. It also opens the door to “Adult Content”, which can be very profitable indeed. It appears that ChatGPT’s innocuous-looking text box may become the doorway to almost everything of commercial value on the internet.

You can now buy things through ChatGPT with its “Instant Checkout” feature

Sam Altman announces plans for “erotica” on ChatGPT, for adults only

California Governor Newsom signs laws regulating AI companions

On Monday, California Governor Gavin Newsom signed into law a batch of bills regulating chatbot interactions with minors. The major bill, SB-243, requires AI companies to implement protocols to detect and deter teen interactions that might lead to suicide or other self-harm. Teens will also be reminded every 3 hours that they are not talking to a person, but to an AI. This bill was passed in reaction to several notorious instances of young people committing suicide after getting involved in an unhealthy relationship with a chatbot. In addition to SB-243, Newsom signed several other companion bills which required age verification of users, warning labels on social media screens for users under 18, and banning deepfakes that are sexually explicit or related to an election. Given that California is home to many of the largest tech companies, and has the largest population of any US state, these laws are likely to have national impact.

California Governor Newsom signs laws regulating AI companions

AI researcher turned down $1 billion, then joined Meta anyway

Just two months ago, Andrew Tulloch, a highly regarded AI researcher, famously turned down a $1 billion+ offer to join Meta/Facebook, choosing to stay at Mira Murati’s Thinking Machines Lab startup. Apparently Meta/Facebook CEO Mark Zuckerberg did not take no for an answer, and last week Tulloch announced that he was leaving for Meta after all, for “personal reasons.” Speculation is rampant over the reasons for Tulloch’s change of heart, but no one who knows is speaking to the press. This story illustrates the insane levels of competition for top AI talent in Silicon Valley, and how fortunes of companies can shift with the addition or departure of just a few highly able employees.

AI researcher turned down $1 billion, then joined Meta anyway

Risky “circular” deals have become the norm among AI companies

Megadeals have become commonplace in AI. What has been less commented upon is the increasing trend for the top AI companies to engage in circular financing. (See chart below.) For example, OpenAI recently struck a deal with chipmaker AMD to deploy a massive 6 gigawatts of AMD’s AI chips. In addition to receiving the chips, OpenAI gets the right to buy up to 160 million AMD shares (10% of the company) for a penny a share. As another example, premiere AI chipmaker Nvidia has agreed to invest $100 billion in OpenAI, all in the form of leading-edge chips rather than cash.

The problem with most of these deals is that they are, in effect, interlocking debt, where, as just one instance, OpenAI does not yet have the money to give to AMD, who does not yet have the chips. These arrangements are based on the supposition that the AI boom will keep on booming. If the music ever stops, or even slows down, these deals may unravel fast, and the fallout could be massive.

Risky circular deals have become the norm among AI companies

Harvard economist warns that almost all US growth is from datacenters

Harvard Professor of Economics Jason Furman has crunched the numbers. For the first half of 2025, a stunning 92% of the entire growth in US GDP was fueled by the massive buildout of AI datacenters. Without the boost from AI infrastructure investments, GDP growth would have been an anemic 0.1%, barely treading water. Datacenter development has even outpaced consumer spending, the usual driver of economic growth in the US economy. (See chart below.)

Harvard economist warns that almost all US growth is from datacenters

Nvidia is selling $4000 desktop supercomputers

Premiere AI chipmaker Nvidia got its start selling graphics cards to computer gamers. More recently, it has been selling the world’s most advanced AI chips to megacompanies like OpenAI, Amazon, and Google. Today, Nvidia gets back to its consumer technology roots, selling the DGX Spark, a personal AI supercomputer for $3,999. Although priced comparably to a Macbook Pro laptop, the Spark sports a high end Grace Blackwell superchip, 128 gigabytes of memory, up to 4 terabytes of storage, and produces a staggering 1 petaFlop of computing power. (A Flop is a floating point calculation, a measure of computer power. A petaFlop is a million billion Flops.) This is 25 times the power of the Macbook Pro, at a comparable price. With the DGX Spark, you can run highly capable AI models on your desktop, and Nvidia CEO Jensen Huang very much hopes that you will.

Nvidia’s DGX Spark is a desktop AI supercomputer

US residents are open to AI automation, but not for all jobs

Harvard Business School researchers used a large-scale survey of US residents to asses popular attitudes toward allowing AI to automate various occupations. They found surprisingly little resistance to automating most jobs if the AI could perform as well or better than humans and at a lower cost. However, there was a small subset of occupations for which there was near-universal repugnance at the idea of AI automation. These were primarily in the fields of caregiving, therapy, and spiritual leadership. (See chart below for examples.)

US residents are open to AI automation, but not for all jobs

Samsung proves that with AI models, less can be more

Samsung researchers have attacked a central dogma of current AI theories, called scaling laws, which posits that to get a model to behave more intelligently, it has to have more parameters. Current state of the art AI models have approximately a trillion parameters. Now Samsung’s AI scientists have released a paper demonstrating that a tiny AI model of only 7 million parameters can outperform models 10,000 times larger on reasoning tasks. The secret is recursion - the tiny model makes a stab at the answer, then checks its work, considers alternative approaches, and refines the answer. The model goes through this refining process up to 16 times for each answer. The end result is superior reasoning results with much smaller and and more efficient models. If this method holds up under scrutiny, expect a significant amount of future AI computing to be done on computers owned by individuals and small-to-medium size companies, rather than in the cloud on massive datacenters.

How a tiny AI model can beat much larger ones at reasoning

arXiv paper

Ukraine and Russia race to develop AI war drones

The future of warfare is being forged in Ukraine. Faced by an invading force from a country with more than 3 times its population, scrappy Ukraine began to develop cheap drones to use for reconnaissance, and later to carry grenades and other explosives. Soon Russia began jamming the radio frequencies that the drone operators used to fly them. This has sparked a major race on both sides to develop AI-operated drones that can fly to a target autonomously and perform their mission, without the danger of jamming from the enemy. Ukraine’s famous “Spider Web” drone attack last June on Russian airbases as far as 4000 km from its border used over 100 drones which were partially AI-enabled. Ukraine’s President Volodymyr Zelensky has warned that AI drones have the potential to be as or more destructive than nuclear weapons, and has called on the international community to set up global rules for their use.

Ukraine and Russia race to develop AI war drones

Autonomous underwater bot to circumnavigate the globe

Last week, the Redwing, an autonomous sub, was launched from the docks of the Woods Hole Oceanographic Institution in Massachusetts, on a scientific mission to travel around the world underwater, a journey that will take more than 5 years. To travel using as little energy as possible, the bot will take advantage of prevailing currents, and will also manipulate its buoyancy by taking in or expelling small amounts of seawater. As the sub slowly bobs up and down underwater, small winglets on the side convert the rising and sinking motion to a forward glide. Along the way, it will continuously collect data on water salinity, depth, and temperature, giving scientists a 3-dimensional map of the oceans. Every 8 to 12 hours, the sub will surface to beam back data via satellite link. In total, the autonomous vehicle is expected to travel some 45,00 miles underwater.

Autonomous underwater bot to circumnavigate the globe

Twin Health improves diabetes outcomes with digital twins

Twin Health, a Silicon Valley digital health startup, has recently received third party verification of the efficacy of its “digital twin” approach to health coaching of diabetics. Type 2 diabetics using the Twin Health AI-enabled health coaching app achieved significantly lower average blood sugar levels, as well as more rapid decrease in need for insulin and/or GLP-1 agents, greater weight loss, and greater improvements in blood pressure and cholesterol, than competing digital health solutions. The company uses data from wearable sensors and user input to create a “digital twin” of the patient, and employs AI to make hyper-personalized real-time suggestions on diet, and exercise.

Twin Health announcement

New England Journal of Medicine Catalyst article

Robots are revolutionizing IVF

Worldwide demand for in vitro fertilization is growing, due to shifting societal norms that encourage delayed reproduction, plus improved technology leading to better success rates. As the number of IVF procedures rises, multiple companies in medical tourism countries such as Turkey and Mexico are developing robots that automate various aspects of the IVF process, maintaining high success rates while slashing costs. The decreased costs are opening the possibility of IVF to families who cannot afford the tens of thousands of dollars required for traditional procedures.

Robots are revolutionizing IVF