Google’s AlphaEvolve is self-improving AI

Google just took a first step toward self-improving artificial intelligence with its new offering, AlphaEvolve. A product of Google’s DeepMind AI research lab, whose previous research in computational biology garnered a Nobel Prize, AlphaEvolve takes an evolutionary approach to algorithm improvement. Given a problem statement, AlphaEvolve uses a collection of AI models to propose computer code to solve the problem, then evaluate the code for how well it solves the problem, propose improvements to the code, test and evaluate those, and on and on until the system decides that no more improvement is readily achievable. The results can be surprisingly powerful. Just in testing the system, Google researchers:

• found a new, more efficient algorithm for matrix multiplication (a mathematical function that is much used in AI models), bettering an existing algorithm that humans had been unable to improve on since 1969

• made substantial progress on solving the “kissing number” problem first proposed by Isaac Newton in 1693 (kissing numbers refer to how many spheres touch a central sphere when closely packed.)

• suggested efficiency improvements in Google’s data center management that have resulted in a 0.7% decrease in Google’s energy use (the savings could power a small town.)

• suggested improvements to Google’s AI chip design that improves efficiency

Once AI can autonomously improve its own constituent algorithms, advancements in the power of AI will likely accelerate significantly, so that improvements come faster and faster.

AlphaEvolve is self-improving AI

Google is putting AI into all of its products

At Google I/O, the company’s flagship annual developers conference being held yesterday and today, Google has made it clear that it is putting AI into every product it makes - Search, Maps, Gmail, Workspace, the whole enchilada. Google’s AI voice assistant, Google Live, is being put forward as the smarter Siri that we’ve all been looking for, but Apple has failed to deliver. Having played catch-up with OpenAI for over 2 years, Google now believes that it has competitive AI products across the board, and is now trying to leverage its enormous customer base, totaling in the billions, to begin capturing more AI market share.

Google is putting AI into all of its products

Microsoft wants to be your AI infrastructure; lays off 6,000 employees

Microsoft is also holding its annual developers conference this week (the better to compete for buzz with Google), and it is laying out a very different vision for the AI future. Google wants to make it easy for you to stay in its product ecosystem, Microsoft wants to allow you to mix and match products from whatever source, but tie it all together with its cloud services. It is even hosting open source AI models, as well as Elon Musk’s Grok, further distancing Microsoft from its once-close alliance with OpenAI. Microsoft apparently wants to be Switzerland, a neutral party trusted by all. It calls its vision the “Open Agentic Web.” Not exactly catchy, but it gets the point across.

In an interesting bit of timing, Microsoft also announced that it was laying off 6,000 employees, approximately 3% of its global workforce. The company’s revenue and profits have been good, but the layoff is described as “clearing the decks” for AI growth. It appears that a disproportionate number of the layoffs were among junior software engineers. CEO Satya Nadella has been quoted saying that up to 30% of Microsoft code is now being generated by AI. A touching post from one laid-off employee has gone viral. She was, as it happens, a Director in the AI section.

Microsoft wants to be your AI infrastructure

One employee’s reaction to being laid off

OpenAI releases Codex, its AI -powered software coder

It is no surprise that OpenAI, the scrappiest attention hog in the top tier of the AI race, released a blockbuster announcement last week, in anticipation of Google’s and Microsoft’s dueling developer conferences this week. The company released Codex, a next-generation AI coding tool that automates a significant amount of the busywork associated with software development, allowing software developers to focus on the core logic and important features. Early reviews have been highly favorable. AI coding assistants are becoming more and more capable, and there seems to be no end in sight as to how much of the coding task AI will eventually be able to shoulder - many informed insiders think it could eventually reach 100%, and humans would become managers of the AI assistants.

OpenAI releases Codex, its AI-powered coding assistant

From chalkboards to chatbots: AI improves learning in World Bank pilot

We now have the official report of the World Bank experimental study on the impact of AI chatbots as individual tutors of English for high school students in Nigeria. In only 6 weeks, students made average gains of 1.5 to 2 years of “business as usual” classroom-based schooling. We have known since at least 1984, the year that Benjamin Bloom’s seminal “The Two Sigma Problem” was published, that one to one tutoring produces outcomes vastly superior to usual classroom instruction. Until now, there have never been the resources to implement that finding widely. Now, with always available, cheap-to-use, highly capable AI models like ChatGPT, it is feasible to provide a one to one AI tutor for every student. Our challenge is to retool the educational system so that AI becomes a help to learning, rather than a means to avoid it.

AI improves learning in World Bank pilot

NotebookLM is now an app

NotebookLM, Google’s viral hit of an AI study assistant, is now an app on iPhone and Android. The app will allow you to connect multiple sources of information to a “notebook”, and then will summarize, analyze, create flash cards, pose review questions, or even create an engaging podcast with 2 AI voices discussing the major points of the linked information.

NotebookLM is now an app

Nature article proves the persuasiveness of chatbots

An international team of researchers ran an experiment in which subjects debated either a human or an AI chatbot (GPT-4), and the effectiveness of each in influencing the subject’s opinion was measured. It found that the chatbot was at least equally persuasive to humans when no personal information was known about the subject, but when personal details were given to the debater (human or chatbot), the AI was significantly more effective than humans at marshaling that personal information for persuasion. In fact, the chatbot was more persuasive than a human in 2 out of 3 cases. Dealing with highly persuasive chatbots spreading misinformation is a problem that society will need to grapple with in the coming years.

Nature article proves the persuasiveness of chatbots

NetFlix and YouTube experiment with intrusive AI ads

Two of our major sources of streaming video content, Netflix and YouTube, are experimenting with AI to create and position ads to be more effective at motivating viewers to buy advertised products. Netflix is seeking to use AI to seamlessly blend ads into the movie you are watching, so that they seem part of the story. YouTube is maintaining the division between story and ads, but wants to get away from pre-programmed ad slots, and have the AI place the ad just after a moment of high emotion, on the theory that viewers will be most engaged at that time. Both initiatives seem horrible to moi, examples of ham-handed manipulation that will likely decrease the enjoyment of the viewer.

Netflix wants AI to blend ads seamlessly into the video

YouTube wants AI to place ads immediately after an emotional high point in the video

Singapore’s ubiquitous robot dogs

Four-legged, doglike robots are becoming a common form for a variety of commercial uses, including surveillance and inspection. Agile, stable, and able to traverse stairs or rough terrain, these robo-dogs are versatile workhorses. Few places have embraced the use of robot dogs as widely as Singapore. Here the robots are being used as guide dogs for the visually impaired, watch dogs for construction sites and transportation stations, and inspection assistants for hazardous or difficult to reach environments. The robodogs can be equipped with cameras and a variety of other sensors to allow human supervisors to assess situations remotely.

Singapore’s ubiquitous robot dogs

Giant robot bugs are working on farms

Although robotics and AI are spreading rapidly into agriculture, most of the effort to date has been directed toward row crops, with separate plants in neat rows, which lend themselves easily to various types of automation. Little has been done to bring these technologies to bushy, perennial crops, like strawberries, blueberries, or grapes. Now commercial robots for these applications are beginning to hit the market, and they look nothing like John Deere’s semi-autonomous tractors. One solution for bush crops is a giant robotic insect, whose segmented body and multiple legs are ideal for dealing with unplowed terrain, weeding under a canopy of leaves, and getting close to the stalk or stem in order to test soil moisture.

Giant robot bugs are working on farms

Baby cured of deadly genetic disease by gene editing

Researchers and clinicians at the Childrens Hospital of Philadelphia achieved the first-ever cure of a deadly genetic disease by editing an infant’s genes through infusion of corrective base-editing nanoparticles into its bloodstream. Baby KJ Muldoon was born with an extremely rare genetic defect (affecting 1 in every1,300.000 live births a year in the US) that causes a deficiency of an enzyme in the liver which processes protein, a condition that has a 50% mortality rate in infancy. The team identified the exact DNA defect that was causing the problem, and designed a precise replacement method for that stretch of DNA through a technique called base editing. Baby KJ accepted the treatment well, and almost immediately showed improved ability to process protein and significantly lowered toxic byproducts such as ammonia in the blood. KJ is now 15 months old and doing well. This feat points the way for similar treatments of other genetic defects.

Baby cured of deadly genetic disease by gene editing

Cleveland Clinic, Oracle, G42 partner for health AI platform

Cleveland Clinic has partnered with Oracle (whose CEO Larry Ellison was last seen with President Trump and OpenAI CEO Sam Altman announcing a $500 billion investment in AI data centers) and the UAE AI startup G42 (which was recently given permission by President Trump to acquire top-level AI chips from Nvidia) in order to create a health AI platform. The idea here is to combine Cleveland Clinics world-class clinical expertise with Oracle’s cloud infrastructure and G42’s AI expertise to create AI tools and systems to improve medical care and patient outcomes. The partnership will operate in both the US and the UAE. Expect more of these clinical - cloud - AI alliances as medical systems work to incorporate AI, while cloud providers and AI companies try to get a piece of the $5 trillion the US spends annually on health care.

Cleveland Clinic, Oracle, G42 partner for health AI platform