How much energy does AI use? It turns out - a lot. Carnegie Mellon University and AI startup HuggingFace looked into the energy consumption associated with AI in a recent report.

According to their tests, image generation consumes between 0.06 to 2.9 kWh for every 1,000 inferences, with the average image generator using about 1.35 kWh per 1,000 inferences.

To put it into perspective, charging an average smartphone uses 0.012 kWh of energy. Therefore, creating nine AI-generated images uses roughly the same amount of energy as fully charging your phone. In terms of environmental impact, each image results in the production of about 0.1 to 0.5 grams of carbon dioxide.

We highlighted Mark Zuckerberg’s appearance on the Dwarkesh Patel podcast in a previous AI Logs, and how he saw energy usage as the main bottleneck for the growth of AI.

This is rapidly becoming apparent with Bloomberg reporting that data center developers in Northern Virginia are asking Dominion Energy for ‘several gigawatts’ of electricity, as much power as several nuclear reactors can generate.

So far, AI’s energy use has been relatively under the radar, but as usage increases and AI is incorporated into everything from social to search, it could come under significant scrutiny.

Apple’s ‘outrageously powerful chip for AI’

On Tuesday, Apple introduced new iPads powered by the M4 chip, described as “an outrageously powerful chip for AI”.

The M4 chip enhances capabilities in CPU, GPU, and dedicated AI through its "neural engine". This is part of a broader industry trend where tech giants, including Microsoft, Amazon, and Google, have also released AI-specific chips.

Apple have been slow to fully incorporate AI, but the latest chips give it an opportunity to bring many AI features that run in the cloud on to its customers’ own devices, making apps run faster and offering greater privacy.

search.chatgpt

Is OpenAI about to go head-to-head with Google? Speculation has been rife since OpenAI’s SSL certificate logs showed they had created search.chatgpt.com.

Bloomberg have now followed up citing ‘a person familiar with the matter’, that OpenAI is developing a feature for ChatGPT that can search the web and cite sources in its results.

This could be coming sooner than we think with Tom’s Guide predicting it may be with us within a week.

gpt2-chatbot is back

In last week’s AI Logs we talked about the strange appearance, then disappearance, of gpt2-chatbot. Now it seems to be back, with the emergence of two models: 'im-a-good-gpt2-chatbot' and 'im-also-a-good-gpt2-chatbot'.

Though not admitted, it does now appear this is an OpenAI project. Prior to the models being released, Sam Altman tweeted:

The models are available at chat.lmsys.org in the Arena battle mode. You are not guaranteed to get them first try, so it may take a few refreshes. However, it looks to be worth the effort, showing capabilities and training data beyond GPT4.

Did a friend forward this e-mail to you?

IE₊ SUPPORT INTERESTING ENGINEERING
_{Invest In Science And Engineering}

Enjoy exclusive access to the forefront of AI content, highlighting trends and news that shape the future. Join a community passionate about AI, delve into the latest AI breakthroughs, and be informed with our AI-focused weekly premium newsletters. With IE+, AI reporting goes beyond the ordinary - and it is Ad-Free.

NEWS

• Anduril unveils AI-powered electromagnetic warfare tech to hunt drones

  Pulsar’s architecture is modular and open, ensuring that the system can be rapidly updated over time.
  

• AI priest, Fr Justin, okayed Gatorade baptisms and incest — gets demoted

  Ex-Google engineer’s AI church under fire as AI priest allows catholic offenses.
  

• Pro music video powered by OpenAI’s Sora released in a world-first

  Trillo generated 700 clips with Sora, 55 of which were used in the final video.

• Matter vs. antimatter: CERN cracks particle physics puzzle with AI

  CERN researchers find first evidence of CP violation in the decay of the strange beauty meson.

MUST READ

Med-Gemini beats GPT-4, achieves 91.1% accuracy in medical diagnostics

Med-Gemini, an advanced AI model specialized in medicine, is more factually accurate, reliable, and produces nuanced results for complex clinical tasks than OpenAI’s GPT-4.

DeepMind, Google’s AI research lab and Google Research released a recent paper on their upcoming AI tool for use in the healthcare space.

Still in the research phase, Med-Gemini has been equipped with the latest technology that can even surpass popular industry standards, according to Google researchers.

AI PICTURE OF THE WEEK

TUTORIAL

Level Up Your Coding with Gemini

Gemini's capabilities can be incredibly useful when working with code in several ways, such as learn a new coding concept through interactive code examples.

Choose a new concept: Pick a programming concept you've been wanting to learn, like recursion, object-oriented programming, or working with APIs.

Find a simple code example: Look for a basic code example online that demonstrates the concept you want to learn. Tutorials or code snippets from documentation are good places to start.

Ask Gemini to explain and elaborate: Paste the code snippet into the Gemini interface. Ask it to explain the code line by line, breaking down the functionality of each part.

Challenge Gemini to modify the code: Once you understand the basic structure, ask Gemini to modify the code in specific ways. You could ask it to:

• Add functionality to the code (e.g., implement error handling)

• Rewrite the code in a different style (e.g., using a loop instead of recursion)

• Extend the code to solve a slightly more complex problem

Test and analyze the changes: Run the modified code and see if it produces the expected results. Analyze the changes Gemini made and see if you understand the reasoning behind them.

Example Prompt:

I pasted a simple Python function that calculates the factorial of a number. Can you explain how each line of code works? Once I understand it, can you modify the function to also calculate the sum of all the factorials from 1 to a given input number?

Beyond the basics: As you gain confidence, use Gemini to tackle more complex coding challenges. Ask it to generate code snippets that solve specific problems based on your requirements.

Tips: The more specific you are with your instructions, the more helpful Gemini's response will be. Break down larger problems into smaller, more manageable steps. Don't be afraid to experiment and ask Gemini "what if" questions to see the different ways code can be manipulated.

TOOLS OF THE WEEK

Nova AI: Nova are launching a video content analyzer, using AI to categorize and organize footage easily. AI recognizes and tags different elements of your videos enabling you to manage and retrieve content.

Perplexity: With OpenAI potentially entering into search, it’s a good time to try Perplexity. Perplexity is an alternative to traditional search engines, where you can directly pose your questions and receive concise, accurate answers backed up by a curated set of sources. It has a conversational interface, contextual awareness and personalization to learn your interests and preferences over time.

Palette FM: Palette FM is a free AI-driven tool that colorizes black and white photos. Upload your photo, pick your filter, and watch your photos transform.