Runway releases new video AI, rakes in $

Breakout text-to-video AI startup Runway is making waves with its release this week of its newest and most powerful model, Generation 3 Alpha. Social media is filling up with demos of its ability to translate brief text prompts into stunning videos. On the strength of its AI video prowess, Runway is extending its already considerable lead in VC funding by raising $450 million at a valuation of $4 billion.

Runway releases new video AI

Runway raising $450 million at a $4 billion valuation

Amazon to spend $100 billion on data centers over next 10 years

AI needs a lot of “compute.” Cloud services giant Amazon aims to continue garnering the lion’s share of that compute by building massive data centers wherever there is cheap land and power. This is turning rural counties all over the US into hubs of critical AI infrastructure, with all the attendant local benefits (jobs, tax revenue) and challenges (increased strain on the local electrical grid, industrialization of once-bucolic communities.) The scale and rapidity of development is staggering, and is unlikely to slow any time soon.

Amazon plans to spend $100 billion on data centers

Microsoft AI chief sparks backlash with remarks on “fair use”

Microsoft’s new AI boss Mustafa Suleyman is in hot water again, this time for implying in an interview on CNBC that all content on the public web (i.e., not your emails) should be available free for AI companies to train their models. This unleashed a firestorm of criticism by content owners (newspapers, magazines, record companies, artists) who are arguing that such use is a violation of copyright. In general, the AI companies that can afford to do so are buying off the opposition with licensing deals for access. Suleyman is no stranger to controversy. Microsoft “acqui-hired” him just 3 months ago, by importing him and the bulk of his team at AI startup InflectionAI, despite the fact that Suleyman had raised $1.5 billion for that venture. Investors in Inflection were appeased with a “services agreement” between the remnant of Inflection and Microsoft, that in effect guarantees that investors will get their money back. Amazon has just copied this ploy to avoid regulatory roadblocks to pure-play acquisitions, by acqui-hiring the team at AdeptAI, a once-promising AI startup with a talented team that could not raise money fast enough to stay in the brutally expensive AI model-building game.

Bad-boy AI boss at Microsoft implies the web should be free for AI

Anthropic will fund 3rd-party evaluators

OpenAI competitor Anthropic’s new AI model, Claude Sonnet 3.5, is acing all the benchmark tests. And that’s becoming a problem. AI models are improving so fast, that soon the current set of benchmarks will be obsolete, unable to capture that improvement in any measurable way. So Anthropic has decided to share some of that sweet, sweet VC money it raked in, to fund outside groups to develop the next generation of benchmark tests, so Anthropic can work on acing them.

Anthropic will fund development of new AI benchmarks by 3rd parties

AI is transforming farming with driverless tractors and drones

Farming is a capital-intensive business, with uncertain returns that depend on the weather. So farm technology has been relatively conservative, not changing fundamentally over the last 100 years. AI is now in the process of changing all that in a hurry. Driverless tractors are plowing fields, guided by GPS and an iPhone, while Roomba-like robots weed the crops, and drones spray for pests. Meanwhile, hyper-localized and ever more accurate AI-based weather models are helping farmers optimize decisions about planting, cultivation, and harvesting. In the new world that AI is ushering in, even farming is becoming a high-tech occupation.

AI will deliver customized Olympic recaps with voice clone

Amid all the angst about political deepfakes and Morgan Freeman’s unauthorized voice clone, NBC somehow thought that this was the perfect time to have AI bring us the latest sports news. NBC has just announced that they will be streaming a nightly recap of the upcoming Paris Olympics that is voiced by an AI-cloned replica of the iconic voice of sportscaster Al Michaels. The recap will use generative AI to tailor the recap to the interests of each individual subscriber to NBC’s online Peacock online streaming service. The 79-year-old Michaels says that he is fully on board, likely welcoming the chance to view his own customized Olympics recap in his jammies, while collecting sweet licensing fees for use of his voice clone.

NBC to deliver AI-customized Olympic recaps via voice clone

MIT’s “soft-hands” robot can bag groceries without damage

Robots are getting ever more capable, but they still have a “stiff hands” problem, which makes them likely to crush or break anything fragile. MIT’s Computer Science and AI Laboratory (CSAIL) is developing a system called RoboGrocery, to advance understanding of how a robot can pack a mixture of hard items (e.g. cans) and delicate items (e.g.grapes or breakable glass) in a way that avoids damage. This is a research project, so you are unlikely to see a robot bag packer at your neighborhood grocery any time soon, but the lessons learned will be publicly available, and so may start influencing the design of warehouse robots in the not-so-distant future.

MIT’s “soft hands” robot can bag groceries without damage

AI without MatMul?

AI is an energy hog, and a lot of that is due to the complex calculations current AI models require, known as “matrix multiplications,” or MatMul for short. Nvidia chips are optimized for MatMul, which goes a long way toward explaining that company’s dominance of the AI chip market. Now researchers from UC Santa Cruz and elsewhere have developed an AI algorithm which avoids MatMul altogether, with striking decreases in energy costs. If validated, this method may help us avoid the looming AI energy apocalypse, while making Nvidia stock that much less of a sure bet. (But don’t ever underestimate Nvidia CEO Jensen Huang, an immigrant who rose from cleaning toilets for a living to head of one of the most valuable companies on the planet.)

Researchers discover new energy-efficient AI algorithm

AI analysis of brain scans identifies 6 types of depression

Finding the right antidepressant medication for a given patient has long been a notoriously hit-or-miss process of trial and error. Now Stanford researchers have developed an AI model that analyzed the functional MRI (fMRI) scans of the brains of depressed patients, which show the activation level of different areas of the brain, and discovered that there are 6 distinct patterns of brain activity that are associated with depressive symptoms. The hope is that identifying the “biotype” of the patient’s depression can lead to more precise therapies.

AI analysis of brain scans identifies 6 types of depression

Nature Medicine paper

AlphaFold may unlock predictive medicine

Google DeepMind’s groundbreaking AlphaFold AI model that accurately predicts protein folding based on genetic sequence information is poised to spark a revolution in predictive medicine, according to an article in AI in Precision Oncology from 2 weeks ago. The review article “describes a shift toward predictive medicine, in which AI, integrated with genomic data, revolutionizes our understanding of diseases, facilitates drug design, and enables personalized therapies.”

AI integrated with genomic data will shift us toward predictive medicine

AI in Precision Oncology paper

Old-school ML predicts response to ICB chemotherapy

It has become commonplace to equate AI with neural network models like ChatGPT. However, artificial intelligence also encompasses old-school Machine Learning Models, which are simpler, but can be amazingly powerful for the right problem. A paper in last month’s Nature Cancer describes how a logistical regression model called LORIS was developed to predict the response of cancer patients to the new Immune Checkpoint Blockade chemotherapy drugs, using just 6 common clinical, pathologic, and genomic features, and outperforming all previous scoring systems.

Old-school machine learning model predicts response to new ICB chemo

AI detects early Alzheimers from speech

Boston University researchers have developed an AI system that analyzes speech patterns of patients with Mild Cognitive Impairment, and can predict progression to Alzheimer’s disease within 6 years with an accuracy of 78.5%. This is another in a burgeoning list of AI models that use analysis of everyday behaviors to make sophisticated and highly accurate diagnoses.

AI detects early Alzheimer’s from speech