NotebookLM converts any text into a podcast

A seamless, easy-to-use interface to AI has yet to be developed, but every so often we see glimpses of the future. Recently, the AI enthusiasts in the tech space have been all agog about NotebookLM, Google’s new system for converting complex documents into easy-to-understand formats such as FAQs, summaries, or even a podcast that has two AI-generated voices with distinct personalities banter their way through the material in an easy-listening format. Further explanation is no match for just trying it, so I recommend that you click the free link below and try it yourself.

NotebookLM converts your text into a catchy podcast

So of course, some joker uploaded a text that instructed the AI to simulate the podcast voices discovering they weren’t real people and then spiral into existential crisis. Depending on your mood, this is either hilarious, eerie, or tragic. Click to hear.

Jailbreak simulates chatbots’ existential crisis

Hurricane Helene shows how vulnerable our tech world is

Hurricane Helene spread devastation over much of the southern US, including - fatefully - tiny Spruce Pine, North Carolina, population 2,000, which just so happens to play an outsized role in the global semiconductor supply chain. Spruce Pine is home to some of the purest quartz on the planet, used to make the ultrapure crucibles that silicon is melted in before being turned into wafers for computer chip production. High end computer chips need contaminant-free silicon (99.999999999% pure), which can only be made in crucibles of ultrapure quartz, that comes overwhelmingly from 2 mines in Spruce Pine. How can an entire critical industry depend so heavily on a single source? Hold onto your hat. A single company in the Netherlands, ASML Holding, is the source of all high-end computer chip photolithography machines. A single company in Taiwan, TSMC, is the sole manufacturer of all the latest computer chips, using the photolithography machines from ASML. The entire leading edge of the AI revolution depends on a web of single-source suppliers of critical components in high end chip manufacture. If Hurricane Helene wakes us up to the inherent fragility of our current semiconductor supply chain, then at least some small good may come of the heartbreak.

Hurricane Helene shows the fragility of the semiconductor supply chain

Zuck announces Orion VR glasses prototype

Meta/Facebook became an accidental Virtual Reality headset favorite by putting VR hardware into stylish Ray-Ban glasses. Building on this serendipitous success, CEO Zuckerberg announced Meta’s next-generation prototype for AI powered augmented reality glasses, called Orion. The functionality of the Orion glasses mimics Apple’s Vision Pro headset (a virtual screen laid over the real-world view), but in a more normal eyeglass form factor, rather than the Vision Pro’s spaceman goggles vibe. All the heavy computation tasks are offloaded to a separate pocket-sized computer, nicknamed the “puck.” In addition, there is a wrist strap for capturing gestures. The goal is for all of the components to be miniaturized enough to disappear into the glasses, or into a tiny wearable accessory. The target price is $300-$400, a tenth of the Vision Pro’s price tag. Zuck opines that he expects AI glasses to eventually replace the smartphone.

Zuck announces Orion augmented reality glasses prototype.

US DOE wants to use coal plants to triple nuclear power by 2050

The US Department of Energy foresees a massive demand for emission-free energy in the future, and is studying a plan to convert decommissioned coal-fired electrical generating plants into nuclear power plants. Today’s electrical grids are designed for mostly baseload power, which does not vary a great deal from day to day. Solar and wind power, however, do vary a great deal, depending on time of day or the weather. Nuclear is seen as a good substitute for the baseload power provided by fossil-fuel plants. AI companies have warmed to nuclear power, since they are committed to an emissions-free future while in the midst of a massive construction boom in data centers, which need 24/7 reliable power. Conversion of decommissioned coal plants seems attractive for a number of reasons, including time and cost savings, since the sites are already zoned and constructed to deliver electrical power. The conversions would also create jobs in the same communities where the coal-fired plants are now located. Today, nuclear power plants deliver approximately 100 gigawatts of power, or 18% of the total energy consumption of the US. DOE hopes to increase that to 300 gigawatts by 2050, which they estimate will require the construction of 200 new nuclear plants.

US DOE wants to repurpose coal plants to nuclear power

AI outperforms CEOs (but not students) in business simulation

Last week Harvard Business Review reported on a study that pitted AI (OpenAI’s GPT 4o) against both business students and seasoned CEOs in a simulation of a large automotive business. Each participant assumed the role of CEO of the simulated business, and was scored on two sets of criteria - first, Key Performance Indicators set by the Board of Directors, which focused mostly on short term measures of success, and second, on longer term measures of the value of the business such as sustainable cash flow. GPT 4o crushed all human competitors round after round, until there was a black-swan event in the simulation (a proxy for COVID-19) where supply chains and sales collapsed. GPT 4o and all of the human CEOs were fired by the Board for missing their KPI targets, but a number of students, who had been more cautious and avoided chasing quick gains at the expense of the long-term health of the business, survived. Somehow, this seems a parable for our age.

AI outperforms CEOs (but not students) in business simulation

MIT spinoff uses new technology to create small, powerful AI models

Liquid AI released 3 very small, very powerful AI models last week, which used a new architecture to produce state of the art results with a small footprint and extreme efficiency in memory usage. The models are not based on the “Transformer” model architecture that almost all other AI models use, but instead on a new architecture pioneered by the computer science department at MIT. This continues the trend of efforts to make smaller AI models more powerful, to allow them to be run on fewer, smaller, cheaper, less power-hungry chips, while achieving results comparable to the massive models developed by large tech companies.

How UK newsrooms are using AI

David Dinsmore, COO of News UK, a subsidiary of Rupert Murdoch’s News Corp global media empire, revealed last week that several newsrooms in Murdoch properties in the UK are experimenting with AI, and the company wants to accelerate this trend. Dinsmore noted that The Times had developed an AI-powered content management assistant that helps with headline suggestions, summaries, etc., and this system has now been rolled out to The Sun and other affiliates. Meanwhile, Reach, the UK’s largest business news publisher, with a large number of niche publications and websites, has developed an AI-powered rewriting tool, that can take a story published in any one of Reach’s titles and rewrite it in the style of and for the audience of any of its other publications.

How UK newsrooms are using AI

NIST grants $6 million to Carnegie Mellon for AI safety center

The Biden administration has empowered the National Institute of Standards and Technology to set standards for, and assess the safety of, AI models as they are developed. NIST, a division of the US Department of Commerce, is responsible for setting and maintaining standards and measures throughout the economy, but in recent years it has been heavily involved in setting standards for cybersecurity, which made it a natural home for AI safety research and assessment. NIST has just announced a $6 million grant to Carnegie Mellon University to establish a center for testing and evaluating AI tools. The aim is to develop risk management capabilities and establish best practices for AI engineering, to keep AI systems safe and trustworthy.

NIST grants $6 million to Carnegie Mellon for AI safety center

The medical revolution in proteomics and AI

An article in Science last week summarized a number of recent papers that point to a near-future revolution in medical research and therapeutics. The revolution is being fueled by the blending of 2 different technologies now being used together to make startling gains in medical knowledge and practice: so-called high-throughput proteomics, and AI. High-throughput proteomics is the ability to rapidly identify literally thousands of proteins in a single small sample of blood. The resulting blizzard of data produced is beyond the capability of unaided humans to make sense of. But AI can find patterns and correlations too subtle or complex for humans to detect, and joining these two technologies is already producing results that hint that we may soon be able to detect, and therefore treat or even prevent, chronic diseases years before they manifest themselves by producing symptoms. Parkinsonism and Alzheimer’s Disease are just two of the diseases where exciting early-warning tests are being developed, and work is proceeding on a concept called “organ-specific aging",” which can detect age-related deterioration of specific organ systems just by the pattern of proteins in the blood.

The medical revolution in proteomics and AI

AI model repurposes existing drugs for rare diseases

There are an estimated 7,000 rare or undiagnosed diseases, affecting in the aggregate over 300 million people worldwide. Few, if any, of these diseases affect enough people to make an expensive drug trial economically worthwhile for pharmaceutical companies, and only about 5% of these diseases have any FDA approved medication. Now a new AI system, called TxGNN, developed by Harvard Medical School and others, analyzes the characteristics of existing medications and searches for their applicability to rare conditions. If there is a match, the drug is already FDA-approved, and can legally be used off-label for treatment of the rare disease by licensed physicians. The researchers have made the TxGNN AI tool open source, so that other researchers can use it freely.

AI model repurposes existing drugs for rare diseases

Kopra Bio genetically engineers cancer for self-destruction

Genetic engineering startup Kopra Bio creates cancer-selective viruses that deliver a genetic payload that makes the cancer cell alert the immune system. This systematically teaches the patient’s immune system to identify and destroy the cancer.

Kopra Bio genetically engineers cancer cells for saelf-destruction

US ranks last in health care

The Commonwealth Fund, a highly-regarded health policy thinktank, has released a report comparing the US health care system to that of comparable developed countries, and found that our health outcomes were the worst of any of the 10 countries, despite the fact that we spent a far greater amount per citizen on health care.

Commonwealth Fund finds US ranks last in health care outcome