How Google's new AI could reduce the aviation industry's carbon footprint

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, guides, technical analysis and exciting developments in open source.

In today’s Dispatch:

• NPR correspondent Shannon Bond gives us a glimpse into the DEF CON 31 competition in Las Vegas last weekend, where thousands of hackers attempted to ‘break’ the most popular AI chatbots. Full results and insights from DEF CON 31 won’t be made available until February.

• A blog post from Everypixel suggests that AI has already generated over 15 billion images using text-to-image algorithms – a number it took photographers 150 years to achieve, spanning from 1826 to 1975.

• OpenAI has announced GPT-4 for content moderation. Mods can label a set of queries as safe or not, then GPT-4 learns from the set, trying to match the experts' labels by iteration. This iterative process yields content policies that are translated into classifiers, enabling the deployment of the policy and content moderation at scale. OpenAI claims their model will cut current users find this method quicker and more dependable than Anthropic’s ‘Constitutional AI’.

Plus: US Democrats launching an AI work force, an open source ChatGPT-powered village simulator, trending tools and more.

(Goofs: yesterday, we included an incorrect link in our social media/videos section for some cutting edge AI-powered drones picking fruit in Israel; here’s the correct link)

Google is using AI to help tackle the aviation industry’s hidden ‘dirty secret’

From Fast Company: Google's new artificial intelligence technology, tested by American Airlines, can guide pilots to adjust their flight paths to reduce contrails by 54%. Contrails are thin clouds formed when the plane's exhaust produces ice crystals at high altitudes. These trails can persist for up to 10 hours under certain conditions, acting as a barrier that traps atmospheric heat.

More details:

• Contrails are responsible for over a third of global warming caused by airplanes.

• Contrail prevention isn't novel; military planes have sought to bypass them for stealth reasons. The issue is predicting the exact altitude to achieve this. Juliet Rothenberg, heading product management for Google's Climate AI projects, highlights the challenge in forecasting "weather in the sky", especially in relation to relative humidity, due to limited sensors at those elevations.

• Google's solution involved manually labeling numerous satellite images of contrails, a tricky endeavor since they closely resemble natural cirrus clouds. These labeled images were then fed to AI models which, when integrated with vast amounts of weather and flight data, could predict contrail formation with enhanced accuracy. Open-source contrail data from Breakthrough Energy further informed their predictions.

• During the six-month evaluation with American Airlines, it was revealed that minor alterations in flight elevation at precise spots could significantly reduce contrail formation. Current software that aids pilots in navigating turbulence can be tweaked to incorporate contrail data.

Takeaways: By merely making advised flight path adjustments, airlines have a straightforward means of lessening the current environmental impact while working on the bigger challenge of curbing CO2 emissions from fuel. Google's intent to freely distribute its data and tools could be a big step for adoption. While the current data says these flight paths increase fuel consumption by 2%, Google claims it could get as low as 0.3%.

AI models are powerful - are they biologically plausible? AI transformers might help us understand the role of astrocytes

From MIT News: Researchers from MIT, the MIT-IBM Watson AI Lab, and Harvard Medical School have proposed a hypothesis that may explain how a transformer (an artificial neural network which powers AI systems like ChatGPT) could be built using biological elements from the human brain. They suggest that a biological network composed of neurons and very common brain cells called astrocytes could perform similarly to a transformer.

More details:

• Astrocytes are non-neuronal brain cells that communicate with neurons. They might play a key computational role for humans; though abundant, the exact functioning of astrocytes remains a subject of intrigue.

• Transformers are the dominant architecture for modern AI systems. It's been unclear so far how these transformer models could be mirrored in biological systems due to their unique design, but bridging neuroscience and AI could provide deep insight into both fields.

• The researchers developed a mathematical model of a neuron-astrocyte network aligned with a transformer’s architecture, essentially modeling a transformer's self-attention mechanism through biology.

Takeaways: The relationship between neuroscience and AI is just now beginning to be explored. Not only can better understanding the brain inform AI development, but AI can also help illuminate the deep mysteries of our own neural processing. Merging insights from neuroscience with AI could provide pathways to creating AI that is not only more efficient and sustainable, but also intelligent in a manner more visibly reminiscent of human cognition than current AI systems. Understanding the biological feasibility of AI models will help pave the way for advanced AI integrations in areas like brain-computer interfaces and advanced prosthetics.

From our sponsors:

Get smarter in 5 minutes with Morning Brew (it's free)

There's a reason over 4 million people start their day with Morning Brew - the daily email that delivers the latest news from Wall Street to Silicon Valley. Business news doesn't have to be boring...make your mornings more enjoyable, for free.

Check it out!

‘Smallville’: an open source ChatGPT-powered simulation of social life

The story: In Monday’s Did you know? section, we highlighted an “AI village” simulator called Smallville that had been developed by a research team from Stanford and Google. The goal of this experiment was to create AI capable of believable, human-like behavior in a simulation setting. The research team has made the code available on GitHub - and there’s a helpful installation guide video available from YouTube creator Matthew Berman.

More details:

• Characters in the simulation form habits, relationships, and coordinate behaviors in surprisingly human-like ways.

• Simulations and agent data are saved locally so you can replay and analyze them.

• You can customize simulations by modifying agent configuration files with their personas, traits, memories, etc.

• The agents are powered by ChatGPT, so running longer simulations can get costly with the OpenAI API. Support for local AI models is coming.

Takeaways: There is a palpable excitement in the gaming community about leveraging AI to create more compelling non-player characters. The Sims, a human life simulator, has become one of the most popular video game franchises of all time by attempting to portray variations on day to day life.

With this simulator, you can go into the code and see not only the AI characters’ innate personality traits, but learned personality traits and spatial memory (what the character has seen/witnessed in the simulation) as well. It’s all pretty incredible - and a very exciting prospect for the future of gaming.

Caveats: some complete agents “remember” details they have not experienced. Others show erratic behavior, like not recognizing that a one-person bathroom is occupied or that a business is closed. Some agents use oddly formal language in intimate conversation.

Social media/videos/podcasts:

• ChatGPT Tests Into Top 1% for Original Creative Thinking [Reddit]

• Machine learning and LLM’s in Healthcare [Podcast]

• A ‘general purpose’ humanoid robot for sale from Unitree [X]

• It’s not just Smallville: come on over to AI-town [X]

• Proof that AI Understands? DeepLearning.AI founder Andrew Ng & ‘AI Godfather’ Geoffrey Hinton [YouTube]

Did you know?

Futureverse has been developing a text-to-music AI model called JEN-1 for a few years, and they’ve just demoed their project for the first time. Not even Meta’s open source AudioCraft/MusicGen approaches the sound quality (48khz stereo) for text-to-music.

Futureverse was formed from the merger of 11 different companies late last year and has raised over $54 million in funding.

Trending AI Tools & Services:

• Motion: increases productivity with automation and AI that intelligently plan your day, schedule meetings, build to-do lists, etc.

• Replicate: API now supports server-sent event streams for language models. This lets you update your app live, as the model is running

• Clay: combine data from any provider to find your ideal customer and send personalized messages powered by AI

• Pictory: automatically create short, highly-sharable branded videos from your long form content

• Faye: automate web3 customer queries with AI and blockchain data.