Personalized AI health coach announced by Sam Altman, Arianna Huffington

OpenAI CEO Sam Altman, who has fingers in many pies, has teamed up with Arianna Huffington, former publisher of the liberal e-news site Huffington Post, in a venture to bring hyper-personalized AI health coaching to everyone. Huffington is currently CEO of Thrive Global, a wellness company that markets itself to corporations as a way to enhance productivity and avoid employee burnout. The new venture plans to use AI to learn everything about your habits and medical history (Danger Will Robinson! Privacy alert!) in order to coach you to healthier habits. A personalized AI health coach is a good, fairly obvious idea, but Big Tech’s track record on privacy is spotty at best, and AI has been known to leak sensitive data, or even make up purported “facts” out of thin air. So, creation of a safe and reliable health coach will be a boon, but success is by no means assured.

AI health coach announcement on Twitter

AI health coach article in Time magazine

Microsoft releases upgrade to Phi-3 open source small AI model

In addition to backing OpenAI’s monster models to the tune of $13 billion, Microsoft is doing some of the best work in the world on making small AI models highly capable. Best of all, they are open-sourcing their work, so that the AI community can learn along with them, Now Microsoft has released a significant upgrade to its latest Phi-3 model, which makes its performance with 3 billion parameters even closer to that of the big models like ChatGPT (with 175 billion parameters.) Microsoft, one of the largest companies in the world, thinks small can be beautiful, too.

Microsoft releases upgrade to Phi-3 open source small AI model

Google discovers method to slash AI power consumption

AI is an electricity hog. Data centers are projected to consume 8% of US electricity in 2030, up from 3% in 2022, the last “pre-AI” year. This is 3 times the total electricity consumed by the Netherlands, and about 1/3 of the total electricity consumption of Japan.

Luckily, people can get very clever about avoiding staggering costs like this, and some of the clever people at Google’s DeepMind AI research division have come up with a new method for slashing the power demands of training AI models, one of the most compute-intense phases of an AI model’s lifespan. The new method, known as JEST, reduces the compute needed for training an AI model by 13x, and reduces electrical power consumption by 90%.

Google discovers method to slash AI power consumption

Original research report on arXiv

Japan introduces huge humanoid robot to maintain rail lines

Japan is all in on robots. Japan, like Korea, has a low birth rate and a low immigration rate, which makes for an aging population and fewer young workers to perform manual labor. Robots are a major part of their plan to fill the gap.

West Japan Railways is deploying a 40-foot tall maintenance robot with coke bottle eyes on a Wall-E style head, and enormous arms that can be fitted with paintbrushes or lansdscaping tools to maintain the company’s extensive rail system. The robot is mounted on a truck that can ride the rails and put the mechanical railworker in position to perform its duties. An operator sits in a small cockpit on the truck, viewing the world through the robot’s camera-eyes and manipulating the lengthy arms.

The company says that the robot will initially be used to cut down overhanging branches from nearby trees, and to paint the metal frames that hold the electrical cables above the trains. According to the company, not only is the robot an effective response to labor shortages, it helps prevent accidents from workers falling from the metal frames or receiving electrical shocks.

Japan introduces huge humanoid robot to maintain rail lines

Label behind K-pop stars BTS launches AI virtual girl band

Music label HYBE, which has backed a number of K-pop mega-bands such as BTS, NewJeans, and SEVENTEEN, has now launched an all-AI girl band known as SYNDI8 (pronounced “Cindy-8”). The cartoon avatars are voiced by AI voice synthesis, created by HYBE subsidiary Supertone, a voice technology company acquired by HYBE in 2022 for $32 billion. The virtual bandmembers are given an elaborate backstory , which has them originating in a magical fantasy land. Backlash from human artists has been immediate, who see AI-generated music as an existential threat. We reported last April on an open letter signed by 200 artists including Stevie Wonder, Billie Eilish, and Nicki Minaj decrying the use of AI to displace human artists. The trend seems to be against the humans though - YouTube is reportedly in talks with major music labels about AI music deals.

Korean music label launches all-AI virtual girl band

Morehouse will use virtual teaching assistants this fall

Venerable HBCU (historically black college or university) Morehouse College is implementing a program of AI-assisted virtual teaching assistants this fall. Students will have 24/7 access to virtual avatars powered by OpenAI’s chatbot models. The avatars are trained on the curriculum for each course, and can act as tutors, communicating in the student’s native language.

Morehouse College initiates AI teaching assistants

Try a demo of the technology here (Warning- the initial download takes a minute or two)

Figure 1 humanoid robot begins trial at BMW assembly plant

VC-darling robot startup Figure has struck a deal with BMW to try out their humanoid robot Figure 1 in BMW’s Spartanburg, SC assembly plant. The link below leads to some video of initial tests. To my eye, the robot looks a bit slow, but hey, it doesn’t take coffee breaks or need to pee, so maybe it evens out.

Video of Figure 1 humanoid robot at BMWS assembly plant

AI helps design superefficient iron magnets for smaller MRIs

MRI machines are notoriously large, noisy, and expensive. A lot of that stems from the need for superstrong magnets, which currently use large coils of supercooled niobium-tin alloy. Now scientists have used machine learning, a subsegment of AI, to design much more efficient, smaller, and cheaper magnets from iron. This is projected to reduce the size and cost of MRI machines enough to make it practical to place them in doctor’s offices, making them much more accessible. The research also has possible other applications that require strong magnetic fields, such as fusion reactors.

AI helps design superefficient iron magnets for smaller MRIs

AI helps detect more breast cancer

Two recent studies have highlighted the ability of AI analysis of mammographic images to increase detection of early and hard-to-see breast caners. In both studies, the AI analysis was used in addition to standard human readings by radiologists. In the first study, AI was made available to radiologists, and the before and after rates of detection of breast cancers showed an increase in tumor finding. In the second study, a formal AI “second read” process was implemented, wherein a radiologist and an AI model both read the same image, and this process showed that the AI was finding cancers missed by the human. Long before we trust AI to read mammograms unsupervised, this sort of “human in the loop” augmentation of human expertise will improve medical care.

“…Impact of using AI in Mammography…” in Radiology

“AI for breast cancer detection…” in Pubmed from Lancet Digital Health

AI speeds detection of cancer by mass spectroscopy

Mass spectroscopy has been a promising tool for cancer detection for years, as it can detect changes in the cell-surface polysaccharides, or glycans, that are associated with cancers. Unfortunately, analyzing glycan mass spectroscopy signatures can require hours of time from highly trained professionals, limiting its usefulness. Now an international group of researchers led by the University of Gothenburg in Sweden have developed an AI model that can automate and speed up analysis of mass spectroscopy sufficiently to make it a useful tool in cancer diagnostics.

AI speeds detection of cancer by mass spectroscopy

Using AI to develop more effective antibodies

The field of protein engineering has been transformed by AI, but predicting function from sequence has been a weak spot. Now researchers at Stanford University have developed a “structure-informed language model” to engineer synthetic antibodies against a virus, and noted a 25-fold improvement in neutralization of the virus, and a 37-fold increase in affinity for resistant strains of the COVID virus.

AI designs more effective antibodies against resistant strains of COVID