Elon Musk, world’s richest man-baby, has no shortage of ego. But this week, he had to eat humble pie, and admit to the world that his AI venture was such a failure that he was renting out its data center to a competitor.

Elon started xAI, his AI company, with great fanfare three years ago. He eschewed normal guardrails for its chatbot, Grok, in the name of “free speech.” Grok quickly became notorious for producing hate speech and nonconsensual deepfake nudes of celebrities and teen girls. This has led to Grok being outright banned in multiple countries, and under investigation in a slew of others. Meanwhile, Grok missed the pivot of AI to coding support, and to semi-autonomous AI agents. Its user base has been eroding, not growing

To fuel his AI ambitions, Elon had built the world’s largest AI supercomputer facility, nicknamed “Colossus”, in Memphis, Tennessee. But with Grok fading and failing, this facility became a money loser, with high-priced chips rapidly aging into obsolescence without producing revenue.

So when Anthropic came calling, looking for more computer power for its industry-leading Claude AI model, whose usage is growing exponentially, Elon had to swallow his pride and listen. He has previously called Anthropic “Misanthropic”, “evil”, “woke”, and “sanctimonious”, with a “deep-seated hatred for Western values.”

No matter. Our boy Musk is positioning his SpaceX-xAI Frankenstein mash-up of a company for the largest Initial Public Offering in history, and he needs to keep a credible AI story alive to get the company’s value where he needs it to be, in order to insulate him from Tesla’s lagging performance and volatile stock.

So Elon met with senior leaders at Anthropic, and - surprise, surprise! - found that everyone was “highly competent and cared a great deal about doing the right thing.”

And the deal was done. Anthropic gets the data center capacity that it needs to service its exponentially growing user base. Elon gets billions in annual revenue, and a reason to avoid having to write down the value of all those expensive chips in Colossus. All this is making SpaceX’s mega-IPO look ever more feasible.

Elon cries uncle, rents data center to Anthropic

Peter Thiel, the mega-billionaire alum of PayPal, early backer of Facebook, and a major investor in Palantir, the defense AI company that helps target US bomb strikes in Iran, is not resting on his laurels. His new company Panthalassa, has just raised $140 million to build AI data centers in the ocean, to be powered by wave energy. I assure you, I am not making this up.

On its face, the only things that makes this plan less fantastical than Elon Musk’s scheme to put data centers into space is that 1) it is located on Earth, where people actually live and can get to the data centers on a conventional boat and not a rocket ship, and 2) they will be surrounded by water, which can actually cool overheating AI chips, rather than the super-insulating hard vacuum of space. (Remember Thermos bottles that were insulated with a sealed vacuum inner chamber?)

Elon has an ulterior motive for pushing his data-centers-in-the-sky pipe dream, which is to give a rationale for his SpaceX-xAI merger, and goose the value of the company’s upcoming IPO. Scratch the surface of the Panthalassa story, and it turns out that Thiel may also have an ulterior motive for raising money for sea-borne AI infrastructure.

As early as 2008, Thiel was advocating for “sea-steading” - building floating city-states in international waters that would be free of all terrestrial laws, regulations, taxes, and social mandates. Although he later turned his attention to building experimental law-free communities on land (he has backed groups that have advocated for takeover of all or part of Greenland, for example), he apparently never lost his fascination with the sea. (Thiel is also obsessed with the AntiChrist, and gave a(n) (in)famous set of lectures on the topic two months ago in the shadow of the Vatican - cheeky!)

Elon Musk. Peter Thiel. Two deeply weird billionaires whose fever dreams are shaping the world for the rest of us.

Peter Thiel’s sea-borne AI fever dream

AI chip designer Cerebras is on track to complete an IPO this year at a share price that would value the company at $48 billion. Cerebras is known for designing AI chips as big as dinner plates, as compared with standard AI chips that are the size of postage stamps. Why so big? Keeping all 4 trillion transistors on one piece of silicon allows all of the 900,000 processing units to communicate with one another at much higher speeds, without the bottleneck of sending data across traditional wires or circuit boards.

Fun fact: both OpenAI CEO Sam Altman and President Greg Brockman were early investors in Cerebras, which was founded in 2016, the year after the founding of OpenAI. In January of this year, the two companies announced a massive multi-year agreement, in which OpenAI would purchase up to 750 megawatts of Cerebras systems, along with the right to acquire up to 10% of Cerebras shares.

Another instance of circular funding, along with a side dish of conflict of interest. par for the course in AI deals these days.

Cerebras, designer of monster AI chips, is poised for $48 billion IPO

Google’s Threat Intelligence Group monitors, analyzes, and disrupts global cybersecurity threats in real time. It focuses on identifying emerging attack patterns, particularly those that use advanced technology like AI.

Recently GTIG disrupted what they describe as the first known zero-day exploit developed with AI assistance. A “zero-day” exploit is an attack on a vulnerability that was previously unknown, so that defenders have had zero days to patch the problem. These exploits are particularly dangerous, as there is initially no known defense. Such exploits can be worth millions on the “gray market” of hackers, if they give access to popular software, because they are so effective for either espionage or theft.

The attack foiled by Google was directed at a popular app, and was designed to foil two-factor authentication to allow unauthorized access. Google disrupted the attack in real time, before damage was done, notified the vendor of the attack, and provided a patch for the vulnerability.

Google emphasized that the era of AI-assisted cyberattacks is now here, and all organizations that maintain significant code bases need to start using AI tools to harden their defenses.

Google foils first known AI-assisted zero-day hacking attempt.

Anthropic’s founders left their jobs at OpenAI to start a new company because they were concerned that the problem of aligning AI behavior to human values wasn’t being taken seriously enough by Sam Altman.

In the years since, Anthropic has made progress on the alignment problem, but has by no means solved it. In a recent paper, they reveal that they have found an approach to instill ethics into AI models that is not only 97% effective, but 28 times cheaper than previous methods.

Up to now, the main way of aligning AI models has been through a method called Reinforcement Learning from Human Feedback, or RLHF. The initial training of the model on the vast, unruly corpus of the internet allows it to develop its emergent properties of sophisticated language use, but also exposes the model to countless examples of unethical behavior (fictional and real). The model is then “fine tuned” by further training it on “gold standard “ answers to specific prompts. It then generates answers to other prompts, and humans score the answers for safety and accuracy. Over time, the model gets better and better at giving safe and accurate responses.

RLHF is slow, expensive, and doesn’t teach the model the “why” of its alignment issues.

Now Anthropic has developed a method in which it explicitly teaches the model ethical reasoning. The model is explicitly trained on ethical principles, and is given examples of sound ethical reasoning about scenarios, in addition to being given examples of ethical behavior. This “show and tell” method is vastly superior to RLHF for ethical alignment. The fact that it is cheaper and easier than the current RLHF method should drive quick adoption of the method by all major AI models.

Anthropic finds a better way to teach AI models ethics

Boston Dynamics, the venerable Boston-area robotic research lab that helped popularize robotics with viral videos of robots dancing and performing gymnastics, is having A Moment.

Last week they were tapped by Time Magazine as one of the 100 most influential companies (of all kinds) in the world. This week, we have news that the South Korean Army is considering using BD robots for surveillance, reconnaissance, and logistics. The intent is to supplement a military that has shrunk by a third due to the baby bust in Korea which has been shrinking the pool of eligible young recruits. South Korean birth rates are among the lowest in the world, far below the level needed to sustain the current size of the population. Since South Korea is not enthusiastic about immigration, they are emphasizing a pivot to robotics to shore up their economy and care for an aging population.

In addition, South Korean auto company Hyundai is an owner of Boston Dynamics. Hyundai is pressuring BD to accelerate the manufacture of their world-class humanoid and quadruped robots to assist with automobile assembly.

Between these two large customers alone, demand for Boston Dynamics robots could soar into the tens of thousands of units per year.

South Korean Army considers Boston Dynamics robots

Hyundai pressuring Boston Dynamics to accelerate robot production

Aurora Innovation, an AI startup that focuses on driverless long-haul trucking, has signed a major agreement with restaurant supply company McLane, to begin driverless hauls in Texas. If this first phase is successful, the plan is to expand this agreement throughout McLane’s footprint in the Sunbelt.

Initially, the driverless trucks will pick up and deliver goods to distribution centers on the outskirts of major cities. McLane will employ local human drivers for “last mile” delivery. This maximizes safety, since the driverless trucks do not have to navigate complex urban environments, and can focus on the much simpler highway transportation leg.

Aurora continues to refine its driverless systems, and is looking to add last mile capability as soon as it is feasible.

Aurora driverless trucks will deliver restaurant supplies in Texas

Health AI startup Bayesian Health has been granted the first FDA clearance for its AI-powered continuous sepsis monitor. Sepsis is one of the major causes of death in hospitalized patients, and it is often difficult to diagnose in its early, most treatable stages. Delay in diagnosis can be deadly, since mortality rates rise sharply with every hour that passes before effective treatment is instituted.

Bayesian Health’s AI system continuously monitors the EHR, including any real time sensors for vital signs, and continuously applies clinical reasoning to the evolving patient status. The system identifies sepsis patients an average of 5 hours prior to unaided clinical diagnosis, and has identified patients as much as 48 hours before clinical signs make the diagnosis clear.

Now that it has been cleared by the FDA, it is expected to be approved for reimbursement by CMS as early as this August.

FDA issues first-ever clearance of AI sepsis monitor

Fitness watches are moving aggressively into the health and wellness space. These devices generate a wealth of continuous physiologic data, which can give valuable insight into the user’s medical condition. In addition, their large, loyal user bases are a ready target for upselling services beyond fitness. Health and wellness are logical extensions of their fitness brands.

Google, who acquired Fitbit some years ago, is upgrading and extending the Fitbit app into Google Health. The new app will function as a hub that will integrate fitness watch data from competitors such as the Apple Watch and the Oura ring, as well as Fitbit. It will also be able to integrate data from US medical records. Google’s AI will then act as a health coach for users.

Boston-based Whoop fitness wearable, which famously has no screen, is moving beyond coaching, with plans to roll out on-demand live video consultations with US licensed clinicians this summer. This feature will require an upgraded Whoop subscription.

Google expands Fitbit to Google Health app

Whoop to roll out on-demand live video consultations to US licensed clinicians