For most of the past 3 years, the AI model race has been relatively evenly matched between OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude. However, OpenAI was the acknowledged category leader, the company that put AI on the map with its breakout product used by hundreds of millions of users each month.

Google, who invented the modern AI architecture, caught up in model quality, but never in mindshare. Anthropic, which was founded by disaffected ex-OpenAI employees, was seen as the feisty kid brother, always competing just out of its league. Until now.

In the past 2 months, Anthropic has surged ahead in every important dimension.

Meanwhile, Anthropic keeps doing what it has been doing - using its category-leading Claude Code software development system on its own code, and releasing approximately one new feature per day.

Anthropic hits $44 billion annual run rate, up $14 billion in just 30 days

As he will be the first to tell you, Elon Musk, the world’s richest man-baby, was an original funder of OpenAI. He promised a billion dollars: he ended up giving $38 million (promises, promises), or 3.8 percent of his pledge, which is 0.0002 per cent (2 ten-thousandths) of the $180 billion that OpenAI has raised from investors to date.

Somehow, Elon thinks his 2 ten-thousandths of a per cent stake should allow him to dictate the fate of one of the most significant companies in the world right now. His argument is that OpenAI defrauded him and illegally broke its not-for-profit charter. He wants OpenAI to be forced back into the not-for-profit organization, and for CEO Sam Altman and President Greg Brockman to be fired.

Both sides are playing to the jury, trying to make the other side out as the baddies. (News flash - they are all baddies.) However the jury decides, the “equitable remedy” imposed is up to the judge, who is unlikely to give either side a clear victory.

Elon fights OpenAI in court

Meta/Facebook is no stranger to having to justify its actions with legally suspect and morally questionable rationales - just look at the numerous lawsuits it has had to defend against its social media algorithms.

Most recently, Meta thought it had found a brilliant workaround that would allow it to acquire Manus, a Chinese company that is a world leader in developing AI Agents that can perform tasks for you on your computer or on the internet. China rarely allows Chinese firms with advanced technology to be acquired by foreign companies. So Manus moved its operations and its people to Singapore last summer, safely beyond the grasp of the Chinese government. Or so they thought.

Meta announced that it was acquiring Manus for a reported $2 billion in December, and key personnel from the company began plans to move to California right away. However, last week the Chinese government announced that the Manus acquisition was NOT approved, and gave Meta mere weeks to unwind the deal. Unwinding includes full restoration of Chinese personnel and Chinese technology, and includes stripping out any features or code that Meta may have added to its products that depend in any way on the intellectual property of Manus.

China has made an unambiguous statement that it considers its AI technology and its associated highly skilled personnel to be national strategic assets. No amount of legal sleight-of-hand is going to deter the People’s Republic from defending its hold on both AI products and the people that build them.

China demands that Meta unwind its $2 billion Manus acquisition

Fintech juggernaut Stripe has just come up with a protocol to allow your AI Agent to shop for you, safely and securely.

Currently, if you want your OpenClaw or Claude Code agent to shop for you online, you have to give the agent access to your actual credit or debit card. This creates 2 vulnerabilities to your bank account. First, the agent could hallucinate or otherwise go off the rails, and buy stuff you don’t want, or don’t need, or can’t afford. Second, the agent could be hacked or misled into revealing your card details to a bad actor who then loots your account. Stripe has now come up with a protocol in which your agent shops with no knowledge of your card details. Once it finds a potential purchase, it must get your explicit approval. Only then will a Stripe subsidiary, known as Link, which already keeps details of millions of cards secure for commerce, issue a one-time card for the exact amount of the purchase. The one-time card value is deducted from your actual account behind the scenes.

Stripe gives AI agents secure access to your card

In the Western US Wildfire Belt. including the West Coast, the Desert Southwest, and the Mountain West, power companies are increasingly placing AI-monitored videocameras on their towers. The videocams are placed in areas susceptible to wildfires, and the real-time images that are captured are analyzed immediately by AI vision models that have been trained to distinguish a wisp of smoke from fog or dust. Experts estimate that the cameras can cut reaction time to a nascent wildfire by 45 minutes, sufficiently fast that lives and property can be saved.

One recent example was a beginning blaze called the Diamond Fire that was contained and extinguished within a small 7-acre plot of land.

Wildfire cameras monitored by AI speed up response.

Uber is investing big into self-driving cars, which the company expects will eventually supplant most or even all of its current 10 million drivers.

Uber has given up on its goal to develop its own self-driving car, and has pivoted to trying to be the data and infrastructure provider for the rest of the self-driving industry. To further this goal, it is paying its “fleet partners”, companies that own 100-plus vehicles such as local livery and taxi companies, or 1000-plus vehicles such as car rental and leasing companies, to outfit their vehicles with a package of sophisticated sensors. These sensors include high-resolution cameras, radar, and lidar. The cars in the fleets will accumulate millions of miles of actual driving data, which can be used as training data for self-driving systems.

Uber pays its “fleet partners” to carry sensors for self driving data

Last week Time Magazine released its list of the 100 most influential companies in the world. Many of the names are familiar, but to my eye, one company stood out for its relative obscurity despite it outsized impact. That company is Boston Dynamics, a company that made robots before they were cool. In fact, Boston Dynamics is the company who arguably did the most to make robots cool.

Its Atlas humanoid robot danced and jumped its way into our imagination beginning in 2013. Its iconic Spot dog-bot set the form factor for many of the utilitarian inspection and security robots being deployed today.

Its current owner, Hyundai, is now ramping up use of the Boston Dynamics robots in car assembly plants.

Boston Dynamics named one of Time’s 100 most influential companies

Last week, a team of medical researchers from Harvard and Stanford reported in the journal Science on a study in which an AI chatbot (a now-outmoded OpenAI “o1” “reasoning” model from a year-and-a-half ago) was pitted against expert physicians to test clinical reasoning and decision-making in challenging cases. Out of 6 different experiments, the chatbot’s performance was consistently rated as higher than that of the physicians by independent raters who were blinded to whether the source of the diagnosis was a physician or AI.

All of the experiments were based on text-only inputs, so no actual physical exam or conversation with the patient was involved.

AI models are rapidly improving, and health care is a major focus for all of the top AI model makers. Although it is likely that current and near-future models will be insufficiently reliable for unsupervised real-world diagnosis, attention to creating a usable AI clinical decision support system would be a boon for physicians and patients alike.

Mayo Clinic has developed an AI model that can analyze routine abdominal CT scans to detect pancreatic cancer months or even years before clinical diagnosis. Pancreatic cancer rarely causes symptoms before it has spread, so the 5-year survival rate is 15% or less.

The REDMOD system was trained on thousands of abdominal CTs, including a number from patients who were later found to have developed pancreatic cancer. The system was able to detect subtle abnormalities invisible to humans, allowing it to detect 73% of these occult cancers, twice the performance of expert human radiologists. These cancers were detected at a median of 16 months prior to clinical diagnosis.

This type of deep image analysis is becoming ever more common, and the challenge now is to assemble the different single-disease detection models into a comprehensive, easy to use clinical assistant that fits into the radiologist’s larger workflows.

Mayo’s REDMOD AI detects pancreatic cancer up to 3 years earlier