- AI Weekly Wrap-Up
- Posts
- New Post 1-24-2024
New Post 1-24-2024
Top Story
BMW to deploy Figure’s Robots in SC plant
Elite carmaker BMW prides itself on the company’s German engineering. Now it’s adding American robots to the mix. Sunnyvale, California-based FigureAI is developing AI-powered humanoid robots for industry, particularly warehousing and manufacturing. Figure’s insight is that we’ve built our world for humans, so humanoid robots are the best form-factor for automating general purpose work. Many disagree, but BMW is giving Figure a chance to prove their robots in a very high profile test at its flagship manufacturing plant in Spartanburg, South Carolina.

Clash of the Titans
Tech giant Microsoft thinks “small is beautiful” for LLMs
Microsoft CEO Satya Nadella thinks big. He poured over $11 billion into OpenAI for access to its groundbreaking ChatGPT AI technology. Lately, though, he’s been thinking small - as in smaller AI models that are highly capable but cheaper and easier to run on a customer’s own computers. Microsoft recently released an open source version of its Phi-2 LLM, which has only 2.7 billion parameters (tiny for a Large Language Model) but performs as well as models 10x or more its size. The future of AI running on your laptop (or maybe even your Apple Watch?) may be nearer than we thought.
Want to participate in the tiny revolution and run an LLM locally on your own computer? Follow the links below. I did, and now have my own local LLM, which just produced a rap version of the Gettysburg Address.

Pentagon is developing AI-directed autonomous weapons
Autonomous weapons are coming - and early versions are already here. Despite ethical dilemmas, the Pentagon is investing heavily in AI-powered semi-autonomous weapons systems, such as drone swarms and unmanned attack aircraft. They are also using AI for reconnaissance and battlefield intelligence. It is not too much to say that a full AI makeover of weapons, tactics, and ultimately, strategy is underway for the US military - and militaries around the world.

Fake Biden robocalls urge New Hampshire voters not to vote
Just prior to the high-profile New Hampshire presidential primary election yesterday, robocalls using a voice clone of President Joe Biden were sent out by an unknown source to urge New Hampshire voters not to vote. This violates several federal laws, including laws against election interference, and safe to say, somebody is likely to be in a whole heap of trouble real soon.

Fun News
AI girlfriends increase sociability and decrease suicidal thoughts
AI companions may actually do the socially isolated some good. A recent article in Nature surveyed 1006 users of Replika, an online digital social companion. Users of the service expressed more loneliness than age-matched peers, and typically demonstrated extremely high engagement with the app. Strikingly, 3% of users reported that the service halted their suicidal ideation.

Positivity of outcome increases from groups 1 to 4. Group 4 is “It stopped my thoughts of suicide.”
Barcelona bets on “digital twin” as the future of urban planning
Barcelona has created a data-based replica of itself - its “digital twin” - on a supercomputer housed in a deconsecrated chapel. Its purpose is to simulate proposed city planning projects and evaluate the projected outcomes. In the past, city officials say, urban planning initiatives have been driven by catchy ideas and politics. The digital twin project aims to bring rigor and accountability to the process. Digital twin technology was invented by NASA in 2010 to improve simulations of spacecraft missions - now it is being used widely, wherever there are complex interactive systems, from cardiac ICUs to electrical power plants.

Barcelona’s MareNostrum supercomputer crunches the data for its digital twin
Harvard uses AI tutors for intro Computer Science course
Harvard CS profs believe in eating your own cooking. Last semester (Fall 2023), professors running the legendary Introduction to Computer Science course introduced GPT4-based AI “tutors” to help students learn the course material. 88% of students liked the AI assistance, and used it regularly. Yet, despite being designed by some of the smartest pedagogues in the world, using state of the art AI, the tutorbots sometimes hallucinated and got things wrong. As they say in academia, “More research is needed.”

AI in Medicine
FDA clears new AI scanner to detect skin cancer in primary care
Patients often ask their PCP if a certain skin growth is cancer, and the truth is, PCP’s are not trained to make that judgement reliably. The default has been for the PCP to send anything vaguely suspicious to the Dermatologist, which is time consuming for the patient, and often a waste of the Dermatologist’s time. Many companies are working on AI scanners to help make these judgements in the PCP’s office more reliable through AI. This particular scanner, from Miami-based (where there’s lots of sun, lots of skin, and lots of cancer) DermaSensor, scans the lesion with different wavelengths of light, and uses AI to analyze the reflection patterns and identify the 3 major kinds of skin cancer. Look for there to be an increasing number of point-of-care scanners for various diseases, so that more conditions can be diagnosed on-the-spot in the PCP’s office.

AI model suggests that brains naturally develop musical ability
We have long used models of our brain to inform our intuitions about how digital neural networks function. Now neuroscientists are using neural networks to understands how the brain functions. Researchers at the Korea Advanced Institute of Science and Technology (KAIST) used an artificial neural network to demonstrate that music-selective neurons develop spontaneously during the process of learning to process a variety of natural and artificial sounds. This suggests that the ability to detect music is an innate epiphenomenon characteristic of neural networks trained on sound detection.

That's a wrap! More news next week.