Christmas gifts from OpenAI and Google: MUCH smarter AI models

December brought a cornucopia of AI goodies from arch-rivals OpenAI and Google. Most of them have been discussed in the previous 2 issues of this newsletter. But perhaps the most important releases from each of these companies is a new type of AI model that is designed to be much better at reasoning and planning than prior models. Google’s Gemini 2.0 and OpenAI’s o1 and o3 models appear to herald a new direction for AI, with less emphasis on massive training sets, and more focus on spending more computation at question-answering time (known as inference time) to carefully analyze the question and develop step-by-step answers to it. The early tests of these “reasoning” models hint at a step change in intelligence. As one example among many, OpenAI’s o1 model used advanced reasoning to correctly diagnose 80% of a set of challenging real-world medical cases; a panel of human physicians were only able to diagnose 30% of these cases. (See the write-up below in the top story in the “AI in Medicine” section of this newsletter.) Impressive as this is, OpenAI’s just-announced next-generation o3 model is reported to be even smarter. If true, the implications are likely to be massive.

OpenAI and Google release AI models with advanced reasoning skills

Salesforce desperately pushes AI, an existential threat

Hard-charging CEO Marc Benioff is making AI a top priority for his giant CRM (Customer Relationship Management) software company, Salesforce. He has pushed his software developers to create AI-powered virtual employees, known as Agents, for Salesforce customers. He has insisted that all sales to new customers include AI services. And now, he is hiring 2,000 sales reps after downsizing the company by 7,000 positions just 2 years ago. For Benioff, and other entrenched incumbents in SaaS (Software as a Service, i.e. a subscription model of enterprise software delivery), AI is a 2-edged sword. On the one hand, it allows the company to upsell existing customers to buy the shiny new AI capabilities. On the other hand, AI may soon make all of Salesforce’s existing software redundant and irrelevant. FinTech giant Klarna has already terminated their Salesforce subscription, substituting cheaper SaaS alternatives, and layering inhouse-developed AI projects over them to make up the difference in functionality. If more companies follow Klarna’s lead in this area, Salesforce’s AI dream may well become its living nightmare.

Salesforce desperately pushes AI, an existential threat

Sam Altman-backed nuclear power company gets massive contract from data center operator

OpenAI CEO Sam Altman has his finger in almost every AI-adjacent pie, from models, to chips, to (now) nuclear power. AI’s massive appetite for electricity is causing cloud services providers like Amazon, Google, and Microsoft to turn increasingly to nuclear power for stable, 24/7 baseload power with low carbon emissions. Now a Sam Altman-backed nuclear power startup, Oklo, has landed a massive contract to supply electricity to Switch, a company that operates data centers. Large regulatory hurdles remain for Oklo to get certification for its proposed mini-nukes, but the incoming Trump administration is widely expected to be much more open to cutting red tape for projects like these.

Altman-backed nuclear power startup lands contract with data center operator

Google spinoff develops quantum-powered quantitative AI

SandboxAQ is a Google spinoff that is trying to combine AI (the “A” in its name) with quantum computing and quantitative models (the “Q”) to make AI better at math, coding, and logic, a current weak spot. Although Google’s recently-announced Willow chip seems to be advancing ever nearer toward making quantum computing a practical reality, SandboxAQ is mostly focusing on innovative architectures of normal-physics computing to increase AI’s quantitative power. Early applications are seen in biomedicine, cybersecurity, and finance. SandboxAQ has just raised $300 million in VC funds, valuing the company at $5.6 billion.

SandboxAQ raises $300 million for quantitative AI (and maybe quantum?)

Range Rover makes AI-enabled TV ad

Production company Tool uses a hybrid AI-plus-live camerawork process to make television ads in a day that would otherwise take weeks, while also slashing the number of people involved in the process. A case in point is the ad in the link below, in which a Range Rover in the suburbs dreams of all the exotic off-road places it would love to travel to. Live shots of the car are interspersed with AI-generated shots depicting locations around the world, creating a globe-spanning story with no travel required.

Range Rover makes AI-enabled TV ad

Search is morphing into AI answers

Search Engine Optimization (SEO) consultancy BrightEdge released a report last week detailing the shift in online information-seeking away from traditional Search, and toward the newer AI-powered question answering. The report reveals that in November 2024 alone, OpenAI’s AI-powered search service, which was released in August, grew 44% month over month. Meanwhile, AI search engine Perplexity experienced 71% growth in the same timeframe. Under this pressure, Google has added “AI overviews” to its search results, to mimic the upstarts while retaining its cash cow of link-based search. Users so far are showing a preference for answers to spammy ad-driven links, meaning the shift to answers threatens Google’s core business of selling ads for link results.

SEO consultancy BrightEdge sees transition from Search to AI answers

Co-creator of the Roomba vacbot now heads startup for emotional support robots

Colin Angle, co-founder of Roomba-maker iRobot robotics company, has a new startup, this one seeking to make emotional support robots. Angle’s new venture, known as Familiar Machines, aims to develop a new kind of home robot focused on health and wellness. The company apparently is focusing on robot-human interactions, and their robots may function as companions, or even as furry pets. The market for AI-powered companions, either online like websites Replika and CharacterAI, or in robot form, is estimated to be in the tens of billions of dollars per year.

Co-creator of Roomba now heads startup for robot companions

Arizona approves online charter school where all instruction is by AI

The Arizona State Board for Charter Schools recently approved an online charter school in which all instruction is to be delivered by AI, in just 2 hours per day. The school, called Unbound Academy, will use existing AI-powered self-paced learning platforms like Khan Academy. There will be humans involved in the instruction, but not very many of them, and they may not be accredited teachers. After the 2-hour AI instructional block, the rest of the school day will be devoted to “life skills workshops” on topics such as financial literacy, public speaking, and entrepreneurship.

Arizona approves online charter school where all instruction is by AI

New AI shows “superhuman” performance on medical diagnosis

OpenAI’s new(ish) o1 AI model was pitted against physicians on a test of clinical reasoning, and came up with the correct diagnosis in 80% of the cases presented, while the human physicians were correct in only 30% of cases. The New England Journal of Medicine, arguably the premiere medical journal in the world, has a long-running regular feature, called Clinicopathologic Case Conferences, in which an expert physician is presented with challenging real-world cases, and asked to come up with a diagnosis. These case studies were presented to the o1 model for diagnosis, and o1 not only outperformed all previous AI models on this test, but also far outperformed a panel of human physicians who had been presented the cases previously. The day may come when failure to consult an AI on diagnosis will be considered medical malpractice.

Twitter/X summary

arXiv paper

UK approves 4 AI tools to avoid missed fractures in the ER

Failure to diagnose bone fractures are among the most common diagnostic errors in urgent care, accounting for 3% to 10% of all recorded errors. Now the UK’s National Institute for Health and Care Excellence (NICE), which sets standards for technology in health care, among other things, has published a draft standard endorsing 4 different commercial AI radiology technologies for assisting human physicians to detect bone fractures in ER patients. AI has long been known to be excellent at image analysis and categorization, and we can expect to see more of these human-in-the-loop applications of AI as assistive technologies for human physicians.

UK’s NICE approves 4 AI tools to avoid missed fractures in the ER

AI is enabling new blood tests for cancers

Two new technologies are coming together to produce much more powerful blood tests to detect cancers than we have known before. In the past 20 years, research found that DNA-coated carbon nanotubes were uniquely suited to sense complex mixtures of molecules in the blood. The output of this sensing was flashes of light from fluorescence of the nanotubes. Unfortunately, the complex patterns of fluorescence were beyond the capability of humans to analyze unaided. An AI model has been trained to make sense of the fluorescence patterns, and now ovarian cancer, a notoriously difficult disease to detect in its earliest, curable stages, can be detected long before the patient shows symptoms. Generalizing this technique to other diseases is reasonably straightforward, so screening blood tests for many other hard-to-detect diseases are likely to follow.

AI is enabling new blood tests for cancers

Dexcom offers AI analysis of OTC blood sugar sensor data

Medical device maker Dexcom recently made its continuous blood glucose monitoring system available over the counter, without a prescription. Now anyone can get a real time readout of their blood sugar beamed to their smartphone from a sensor attached to their arm. Following up on this initiative, Dexcom is now emailing AI-generated weekly summaries of their data to each subscriber, analyzing how meals, sleep, ands activity affect their blood sugar. Giving users AI-generated insights into their data from wearables is likely a growth industry open to many device makers.