In partnership with

AI Is Already Wreaking Havoc On Global Power Systems

The rapid expansion of artificial intelligence (AI) is dramatically increasing the demand for data centers, which in turn is straining energy grids worldwide. In places like Loudoun County, Virginia, data centers consume massive amounts of power—sometimes as much as 30,000 homes. This surge in electricity demand is outpacing supply, leading to long waits for grid access, potential outages, and rising energy costs. Goldman Sachs estimates that data centers could account for 8% of the US's total power use by 2030, up from 3% in 2022. As data centers grow in size and number, they are starting to consume more electricity than entire countries, complicating global energy transition efforts.

Tech giants like Amazon, Microsoft, and Google aim to power their data centers with renewable energy, but the energy needs of AI, particularly for training and inferencing, are immense. For example, Nvidia’s H100 chips, essential for AI, draw significant power, and the demand is only increasing. The anticipated growth in AI and data centers is leading to infrastructure challenges, with companies like Dominion Energy in Virginia building new substations and power lines to cope with the demand. However, even with these efforts, the energy requirements for AI are pushing companies and nations to reconsider their energy strategies, including turning to nuclear power.

Globally, regions like Malaysia and Texas are emerging as key data center hubs due to their capacity to meet power demands quickly. However, these areas face their own challenges, such as reliance on coal in Malaysia and potential water supply issues in Texas. As new data centers continue to be developed, the energy demands are expected to outstrip the renewable energy supply in many regions, leading to higher energy prices and potential strain on local grids. This rapid growth highlights the need for significant advancements in energy infrastructure and policy to support the burgeoning AI industry while managing environmental and societal impacts.

Two Forbes 30U30 Founders Transforming Mental Wellness

Forbes 30 Under 30 winners founded Aura to solve the $100B problem - mental wellbeing. 

Aura has quickly grown to 8 million users & 100k+ paying subscribers, and attracted investments from legendary Silicon Valley VCs & executives from Spotify, Facebook, and Apple. 

The best part? You can invest now before the opportunity ends in a few days.

Top AI Companies Really Want to Make Their Chatbots Funnier

Leading AI companies, including Anthropic and xAI, are striving to make their chatbots funnier. Despite advancements in AI, humor remains a significant challenge. DeepMind researchers, who enlisted comedians to evaluate AI-generated jokes, found the results lacking in originality and overly cautious. Anthropic's new model, Claude 3.5 Sonnet, claims improvements in grasping nuance and humor, while xAI's Grok aims to be a more humorous alternative. However, the consensus is that current AI humor is limited and far from matching human wit.

Humor in AI is seen as essential for creating engaging conversational agents. Companies like Anthropic and DeepMind are focused on developing chatbots that can handle complex queries and maintain user interest through pleasant interactions. Anthropic co-founder Daniela Amodei acknowledges that humor in AI is challenging but believes improvements are being made. In experiments, responses from various AI models were found to be bland and unoriginal, indicating that achieving genuinely funny AI remains a work in progress.

The pursuit of humor in AI also involves navigating the fine line between funny and offensive. DeepMind's Juliette Love highlights the risk of polarizing humor and emphasizes the need for a balanced approach to minimize potential harm. As tech companies continue to innovate, the dream of a witty AI co-worker may be distant, and for now, users may have to settle for simple, dad-joke-level humor.

Francois Chollet - LLMs won’t lead to AGI - $1,000,000 Prize to find true solution

Francois Chollet, an AI researcher at Google and creator of Keras, argues that large language models (LLMs) like those developed by OpenAI will not lead to artificial general intelligence (AGI) because they rely heavily on memorization rather than true understanding. To address this, Chollet, in collaboration with Mike Knoop, co-founder of Zapier, has launched a $1 million prize to solve the ARC benchmark, a set of puzzles designed to test machine intelligence's ability to reason and learn in a way resistant to memorization. Chollet explains that while LLMs perform well on tasks that can be memorized, they struggle with novel tasks that require on-the-fly problem-solving and reasoning, a critical component of AGI.

The ARC benchmark is intended to serve as an IQ test for machines, requiring only core knowledge, such as basic physics and counting, that any young child possesses. Each ARC puzzle is unique and novel, making it difficult for models to succeed by merely drawing on memorized patterns. Chollet emphasizes that true intelligence involves synthesizing new solutions and adapting to new situations, a capability that current LLMs lack. He highlights the importance of developing new AI paradigms that combine the strengths of deep learning with discrete program synthesis to achieve broader generalization and adaptability.

Chollet and Knoop hope the prize will inspire researchers to explore innovative approaches beyond the current focus on LLMs, which have dominated AI research and funding. They believe that the open competition and the sharing of results will accelerate progress toward AGI by encouraging diverse research directions and collaboration. The prize aims to reveal whether scaling up LLMs alone can achieve AGI or if fundamentally new ideas are necessary to overcome the limitations of existing AI technologies.

Other stuff

• The encyclopedia project, or How to Know in the Age of AI

• How to run an LLM on your PC, not in the cloud, in less than 10 minutes

• llama.ttf is a font file which is also a large language model and an inference engine for that model. What!?

• I am using AI to automatically drop hats outside my window onto New Yorkers

• I found a copy of my work labeled as « impressive AI generation » and without any attribution…

• A peek inside San Francisco’s AI boom

• Researchers describe how to tell if ChatGPT is confabulating

• Testing Generative AI for Circuit Board Design

• How to stop Perplexity and save the web from bad AI

Hope you enjoyed today's newsletter

⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.

OR