When I started creating a written version of the podcast a few months back, I thought AI Insight made sense… It sounded LIKE AI Inside but offered a different type of insight into the stories.
But the more I think about it, it’s confusing. SO, going forward, it’s named the same. Might not be Earth shattering news, but it keeps things simpler and let me tell you something about me: Sometimes, I NEED simpler.
Alrighty, on with the week’s AI news which you can watch or listen to in podcast form if you wish… OR read it right MEOW!
AGI Predictions and Industry Slowdown
Sam Altman, OpenAI's CEO, recently predicted that AGI could arrive as soon as 2025, citing engineering breakthroughs and current hardware capabilities. However, other industry leaders have varying timelines:
- Geoffrey Hinton and Anthropic CEO Dario Amodei: within 5 years
- Deepmind CEO Demis Hassibis and Ray Kurzweil: at least 10 years
- Andrew Ng and Yann LeCun: decades or potentially not in our lifetime
Interestingly, these predictions come amid reports of slowdowns in LLM development. OpenAI's unannounced Orion model is showing smaller improvements over GPT-4 than expected, while the company grapples with data scarcity. Google's Gemini and Anthropic's 3.5 Opus model are also facing delays and not meeting internal expectations.
The industry is experiencing diminishing returns, data scarcity, cost-to-performance challenges, and limits in scaling. This raises questions about whether AGI will perpetually be just around the corner.
Anthropic Partners with Palantir and AWS for Defense Sector
Anthropic, known as the "safety AI company," has joined a partnership with Palantir and AWS to provide access to Claude AI models in the defense sector. The models will be integrated into Palantir's AI platform, hosted on AWS, with Impact Level 6 accreditation for classified data up to Secret level.
This move follows Meta's recent entry into US intelligence and defense partnerships. While Anthropic assures that government use of its tools is designed to be beneficial and restricted from harmful applications, this partnership raises questions about the potential dark path of AI in defense and intelligence.
AI Advancements in Science
I truly believe in the power of artificial intelligence to pave the way for big progress in the field of science and medicine. Two examples to shine a light on this week:
First, Google Deepmind has released the source code for AlphaFold3, their Nobel Prize-winning protein prediction model, for non-commercial use. The model can predict interactions between proteins and other molecules like DNA, RNA, and potential drug compounds. It has mapped over 200 million protein structures to date.
Second, Japanese researchers have developed an AI model that can screen for medical conditions like high blood pressure and diabetes using a 30-second video of a person's face and hands. The model analyzes blood flow patterns in 30 regions of the face and palm, achieving 94% accuracy for high blood pressure and 75% for diabetes. How cool is that?
Google's "Learn About" Experiment
Google has launched a new experiment called "Learn About," built on its LearnLM AI model. This educational tool provides interactive content, including images, videos, and suggested topics, to help users understand and learn about various subjects. It offers a user-friendly interface that feels like an interactive Wikipedia.
Jeff Jarvis and I did a walkthrough demo of Learn About which you can watch here (at 39m29s)
Particle: AI-Powered News Assistant
Former Twitter engineers have founded Particle, a company aiming to combine AI and news to assist publishers. The platform provides AI-generated summaries while prominently showcasing and linking to original sources. Features include an "explain like I'm 5" option, an "Opposite Side" tool for presenting different viewpoints, and a "Just the Facts" section covering the 5 W's. Particle employs both AI models and human editors for curation.
AI-Generated Art Sells for $1.3 Million
The artwork created by the Ai-Da Robot, titled "AI God" of Alan Turing, sold at Sotheby's auction house for $1.3 million, nearly ten times its estimated value. This marks the first major auction sale of a robot-created piece of art, demonstrating that AI-generated art can indeed have significant value to humans even if it is just speculation.
AI-Assisted Beatles Song Receives Grammy Nominations
As a total Beatles nerd, I can’t pass up stories like this. "Now and Then," the final Beatles song released last year, has received Grammy nominations for Record of the Year and Best Rock Performance. The track relied heavily on AI stem separation technology to clean up John Lennon's original 1978 recording. This nomination marks a significant milestone for AI's role in music production and recognition in the industry. One thing you have to admit about AI in music: It’s getting better all the time.
Here are a few other interesting stories
Microsoft, A16Z AI manifesto: AI for startups - microsoft.com
Randy Travis’s beautiful baritone was lost. AI helped him sing again. - washingtonpost.com
YouTube is testing music remixes made by AI - theverge.com