New OpenAI 4o image generator is a viral hit “The revolution will be adorable”

OpenAI just keeps on rolling out the hits. This past week, the company announced an upgrade to its workhorse 4o model, which now includes improved image generation from test. Much improved image generation. As compared to its predecessor, DALL-E, 4o can generate images in a wide array of styles, from photorealistic to line drawings, and it will allow you to edit these images fairly precisely with just text instructions. It also renders readable text well, a signal failing of DALL-E. But much to the surprise of the OpenAI team, the thing that absolutely grabbed the attention of the users was its ability to transform photos into cartoons in the style of Studio Ghibli, a much-beloved Japanese animation studio. Social media became flooded with Ghibli-ized family photos, screenshots of popular television series, even a full movie trailer for Lord of the Rings. OpenAI’s CEO tweeted that the company’s computers were “melting” from the massive increase in use. In another tweet, he noted that in the initial wave of adoption of ChatGPT, when it became the fastest growing web app in history, the company had gained 1 million users in 5 days. With 4o image generation, they were gaining 1 million new users per hour. See below for an example of a popular internet meme that has undergone a Ghibli transformation. Then go to ChatGPT, make sure that you choose the 4o model, and ask it to transform a family photo into a Ghibli cartoon. Free users get 3 images a day until OpenAI can scrounge up some more servers.

OpenAI’s rather quiet announcement of their new 4o image generation.

The Revolution Will Be Adorable. (Every real revolution starts as a toy.)

Musk merges Twitter/X with xAI to bail himself out of trouble

In his trademark style of braggadocio, self-promotion, and insane levels of unethical self-dealing, Elon Musk has decided to merge the troubled media company that he personally owns (X, formerly known as Twitter), with the infant AI startup that he personally owns (xAI), and slap a ridiculously high valuation on both companies, which not-so-coincidentally bails himself out of a heap of trouble brewing in his finances due to the $45 billion he borrowed to acquire Twitter. Twitter’s value has plunged under Musk’s alt-Reich regime, which has sent users and advertisers fleeing. That made his lenders nervous. Tesla’s recent stock drop has made Musk’s loan collateral suspect, so he needed to shore up his net worth. Having control of 2 multi-billion-dollar private companies, neither subject to SEC oversight, gave Musk the opportunity to perform some financial legerdemain. By valuing Twitter at $30 billion (recent independent estimates have been around $14 billion) and xAI at $80 billion (its last round of funding in October valued it at $45 billion), Musk created some $60 billion in his pocket out of thin air, enough to stave off the wolves - for now. UCLA Law School Professor Andrew Verstein characterized the deal as “using monopoly money to buy Pokemon cards.“ I couldn’t have said it better moi-self.

Musk merges Twitter and xAI for a big payday

Apple is working on your personalized AI health coach

Apple has been receiving (largely merited) criticism for the glacial pace at which it is rolling out AI features on its iPhones. Siri’s AI upgrade has been pushed back at least a year, possibly two. CEO Tim Cook has recently fired the former head of AI at Apple, and further corporate shakeups are likely. Meanwhile, some green shoots of progress are beginning to show. Apple is reliably reported to be working on an AI revamp of its Health app, with the goal of giving each user their own personalized AI-enabled health coach, which will integrate all the data from the Apple Watch’s fitness tracking with reliable health advice. The Apple Watch is already FDA-approved for detecting both Atrial Fibrillation (a common abnormal heart rhythm) and Sleep Apnea, and indications are that the company will continue to try to shoehorn additional medical-grade sensors into its wearables.

Apple is working on your personalized AI health coach

Anthropic discovers just how AI models actually “think.” (And it’s not how we thought they think.)

To quote a Steve Jobs catchphrase, Anthropic has discovered that its AI model Claude “thinks different.” For manufactured objects that humans invented, we know surprisingly little about the details of how AI models produce their intellectual output. Taking a cue from biological research on animal nervous systems, Anthropic researchers have developed a method for tracing the stages of an AI model’s formulation of its response to a prompt, with a sort of “AI microscope.” What they have found is that AI models have developed their own heuristics, or “rules of thumb”, for determining an answer, that are unlike standard human thought processes. As an example: when asked to add 56 + 39, the AI doesn’t follow grade-school arithmetic rules by adding the 9 and 6, getting 15, putting 5 in the ones place, then carrying the one to make 5+3+1 = 9 in the tens place, for an answer of 95. It arrives at the correct answer by estimating it (is it in the 80s? 90s? over a 100?), then on a completely separate and parallel track, it determines the digit in the ones place. Putting the two tracks together, it comes up with 95, the correct answer. It’s important for us to remember that AI models aren’t actually taught anything. They are exposed to massive amounts of data (basically everything on the internet) and then they are only told when they get a question wrong. At that point they adjust the strengths of their various internal connections so they will get it right the next time. It shouldn’t be a surprise then, that their intellectual evolution doesn’t replicate that of humans, and that they come up with their own, novel ways of solving problems. They are alien intellects, assembling themselves from repeated and massive exposure to our recorded works.

AI models “think different.”

Google rolls out trip planning in Search, Maps, and Gemini

While Apple continues to trip over its own feet in trying to integrate AI into its iPhones, Google is rolling out more and more AI features for its own products. Starting last week, asking a question of Google Search like “plan an itinerary for a 5-day stay in Costa Rica” will trigger an AI Overview that does just that, with attached photos and reviews of places mentioned, with an expandable map to plot locations. Once you have refined your choices, you can save your itinerary in Google Docs or Gmail, or as a custom list in Google Maps. You can also use Google’s powerful Gemini AI model as a trip-planning expert to help pick destinations and give advice on what to pack. Step by step, Google is trying to develop the indispensable AI assistant that will keep you a customer of the company’s products for years to come.

Google rolls out trip planning in Search, Maps, and Gemini

Manus announces paid tiers of $39 and $199/month for its Agent

Manus, the AI Agent that can perform real world tasks for you, such as reserving a table at your favorite restaurant, is now announcing paid subscriptions, even though it is officially still in beta testing. Manus became an internet sensation last month when a small Chinese AI startup called Butterfly Effect pulled together already existing tools with some custom computer code, and shocked the world with a viral video of Manus performing multiple simultaneous real world tasks. Now Butterfly Effect is trying to manage its explosive growth while simultaneously monetizing its Manus Agent with monthly recurring revenues. For $39/month, users will be able to use Manus to perform 2 simultaneous tasks with a small budget of computing time, and for $199/month, users can activate up to 5 simultaneous tasks with a much larger monthly compute budget. Many observers agree that Agents are The Next Big Thing in AI, and Manus will have to compete with well-heeled competitors such as OpenAI, Anthropic, and Google. Still, as OpenAI can attest, being the first app to go viral in a new category is a big leg up on the competition.

Manus rolls out paid tiers for its Agent

Demand surges for humanoid robot rentals in China

China is turning its world-class manufacturing capability toward the task of building a thriving robotics sector. Humanoid robots such as Unitree’s G1 have gone viral on social media for dancing, doing backflips, and practicing kung fu kicks. This has given them a certain charisma, and almost overnight companies have sprung up in China to rent out these robots by the day. Typical rates are $1,400 per day, and business is brisk at these prices. Most rentals are to serve as a crowd-drawing gimmick for trade shows and the like, but one Chinese social media “influencer” rented a humanoid robot for a day to cook, clean, and go out on a date with him. Naturally it was all caught on video, all the better to capture more eyeballs for his social media site.

Waymo ferries San Francisco kids to practice on Saturdays

Waymo, Google’s driverless taxi company, is beginning to spread into more and more cities across the country, as we have reported previously. However, in San Francisco, one of Waymo’s earliest sites, the service has been around long enough that people have begun to alter their daily lives around it. As a case in point, the link below notes an emerging trend: on Saturday mornings, a large percentage of the Waymo robotaxis on the road are ferrying solo kids (age 8-14), presumably to practice (sports or music). Although Waymo’s terms of service specifically forbid unaccompanied minors, many parents are willing to risk having their account shut down because parents are beginning to feel that their children are safer with a robot than with an unknown human driver. The Waymo app allows a parent to track the location of the robotaxi while it is navigating the trip. Some parents use Waymo to take their child to school, avoiding having to fight the morning traffic jam. So far, Waymo seems content to enforce its unaccompanied minor prohibition lightly, if at all, waiting to see if a real market niche might evolve.

Waymo ferries SF kids to practice on Saturday mornings

Some SF parents send their kids to school in a Waymo driverless taxi

Psychotherapy chatbot achieves results comparable to professionals

Last week Dartmouth researchers published a peer-reviewed paper on a clinical trial of an AI-powered psychotherapy chatbot, known as Therabot. The chatbot was accessed via a smartphone app. Over the 8-week trial period, users of Therabot reported a 51% reduction in symptoms of major depressive disorders, a 31% reduction in symptoms of anxiety, and a 19% reduction in symptoms of eating disorders. The Dartmouth research team have noted that these results are “comparable to what we would see for people with access to gold-standard cognitive therapy with outpatient providers.”

Dartmouth’s announcement of the results of the clinical trial of Therabot

NEJM AI paper

AI Scribe benefits seen, but not yet on financials

AI scribes, those chatbots that transcribe clinical conversations between physicians and patients and then summarize the visit for entry into the electronic medical record, are taking US healthcare by storm. They are seeing one of the fastest adoption cycles of any technology in memory. Now Peterson HealthTech Institute has released a report derived from the insights of a panel of experts and early adopters of AI Scribe systems, with assessments of the impact of AI Scribes on 10 important measures of success. The report concludes that AI Scribes have been an unambiguous success in decreasing clinician burnout, decreasing clinician cognitive overload, improving the quality of the clinical note, and improving the patient experience. The impact of AI scribes on financial measures such as increasing number of patient encounters or improving accuracy of coding were seen as either mixed, or too early to tell.

AI Scribe benefits patients and physicians, the impact on financials is less clear