AI’s Rise in Energy Consumption Prompts Concern For Environmental Protection
This week in AI: AI’s Energy Need Prompts Environmental Questions, Apple and Microsoft launch SLMs, New Federal AI Safety Board Includes OpenAI, Microsoft and Google, and Synthesia Now Has Emotions.
Watch this week’s episode from Alamo Park in San Francisco
Advancement #1: AI’s Rise in Energy Consumption Spurs AI Emission Regulations for Environmental Protection
The International Energy Agency reports that AI data centers use 1.5% of global electricity, with GPUs requiring nearly four times more power than typical cloud servers. Research from the University of California indicates that every 10 to 50 prompts using ChatGPT-3 require a half-liter of water to cool the system. Energy breakthroughs are becoming more crucial for the future of artificial intelligence.
Because of this, this week’s attention from major companies has turned to the energy consumption risks of Generative AI. Two key developments emerged.
Salesforce and Arms pushed for increased environmental regulation of AI, focusing on energy consumption and emissions transparency in tech,while Sam Altman (and Andersen-Horowitz) invested $20 million in Exowatt, a solar-powered energy startup aimed at supporting AI data centers.
Salesforce is urging companies using AI to publicly disclose metrics regarding their AI's energy carbon footprint to improve efficiencies and inform user choices. Salesforce also advocated for the environmental impact of AI systems to be recognized as a risk factor in legislative assessments of high-risk models, stressing the need for corporate-government collaboration to meet climate goals. Arm's CEO supported this stance, noting in an interview that AI's high energy demands are unsustainable.
In Washington, Senator Ed Markey and three other congressmen introduced the Artificial Intelligence Environmental Impacts Act of 2024, hoping to establish standards for reporting AI's environmental effects and create a voluntary reporting framework for developers.
My Initial Thoughts: Before this week, I had not even considered the environmental impacts of the load required to handle all this GPU. The conversation had always been more focused on how AI would help protect the environment (and solve complex problems), but in reality, it seems like it may be the opposite. I am curious, though, why companies as large as SalesForce or Arm is pushing for regulation, since it clearly would impact them at a significant cost. I wonder if they’re publishing their usage…?
While pinning down the precise environmental impact of AI technologies remains challenging, the necessity for frameworks like the one proposed in the AIEIA is undeniable.
Advancement #2: Apple and Microsoft Launch On-Device SLMs
This week's focus has shifted towards on-device model launches, with both Apple and Microsoft unveiling open-sourced compact models. This hints at a significant transition from large language models (LLMs) to a growing interest in developing more specialized, smaller language models (SLMs).
Apple Open-ELM
Apple's new OpenELM, optimized for on-device usage (such as iPhone and Mac), allegedly outperforms other language models by using a layer-wise scaling strategy in its transformer architecture for increased efficiency and accuracy. It features eight models across four sizes—270M, 450M, 1.1B, and 3B—all trained on public data.
Apple is taking an atypical approach in transparency by sharing both the CoreNet code and the detailed "recipes" for its OpenELM models, ultimately allowing users to inspect how Apple built them. This includes the training logs, pre-training configurations, and the code required to convert models for use with the MLX library for inference and fine-tuning on Apple devices.
This transparency into a company’s building process is unusual for model launches, but Apple’s move hopes to empower the open resource community. The announcement comes weeks ahead of WWDC in June, where Apple is likely to debut its new iOS with AI features.
Microsoft Phi-3
Microsoft's Phi-3 is a compact language model that runs locally, capable of managing 128K tokens of context. This performance, comparable to GPT-4 and superior to Llama-3 and Mistral Large in terms of token capacity, is notable given its 3.8 billion parameter size.
Microsoft utilized a unique "curriculum" training approach, drawing inspiration from children learning methods. They drew inspiration from how children learn through books with simpler words, bedtime stories, and sentence structures that talk about broader topics. Because there aren’t enough children books to supplement the learning dataset, Microsoft used a list over 3,000 words and had an LLM create 'children's books' to train Phi.
My Initial Thoughts: The larger read-between-the-lines headline is AI is making data to train AI. Imagine if your video game started creating its own levels because it knew you were about to run out of new ones to play. That's kind of what's happening with AI right now.
There have already been concerns voiced that AI is going to run out of data to train on in the next few years, so, the next natural step is ai will make data to train on for when that data runs out. Microsoft has been one of the first ones to hint that they're already working on it by creating more children books to train Phi because of that shortage.
Separately, the growing trend towards smaller language models (SLMs) is changing how we build AI, hoping to be more efficient and fitting for different needs. SLMs are smaller, making it quicker to set up, able to run on local devices without needing to connect to the internet, and keeps data private.
We’re moving towards having many types of AI models to choose from, depending on what a job requires. Some companies might need big models for complex tasks, while others might do just fine with smaller ones, especially for tasks that need quick answers. This will enable companies to pick the best AI model for their specific situation, making AI more useful, and cost-accessible.
Advancement #3: The New Federal AI Safety Board Includes OpenAI, Microsoft and Google CEOs
The US Department of Homeland Security announced on Friday the formation of a federal AI safety board, set to include CEOs from major global companies and industries. Aiming to harness expert advice to safeguard airlines, utilities, cyber-security and other essential infrastructure against AI-powered threats, the board will feature influential figures such as Google, Microsoft, and OpenAI CEOs.
Additionally, the panel includes executives from Delta Airlines, Northrup Grumman, Amazon Web Services, IBM, chipmakers like AMD, Anthropic, various government officials, and Stanford’s Human-centered Artificial Intelligence Institute.
The 22-member AI Safety and Security Board, established under a 2023 executive order by President Joe Biden, reflects the government's effort to work closely with the private sector in order to address the dual aspects of AI risks and benefits in the absence of specific national AI legislation.
My Initial Thoughts: Having engaged with AI task forces at state, local, and private levels, I've already observed a range of efforts like the AI Safety Board in action. This new board AND higher level of government collaboration, however, is definitely the strongest alliance to date, and it offers a comprehensive approach to overseeing and securing AI's integration into critical infrastructure.
I simply hope the focus remains on the betterment and protection of humanity, and personal interests (and lobbying) do not get in the way. Getting such powerful people all in the same room to determine the fate of the world… well, that’s a bit scary.
Advancement #4: Synthesia’s New Update Can Do Emotions
Synthesia, an AI video avatar startup, has introduced a new update this week that helps to better portray emotions. Called Expressive Avatars, they used footage of real humans from their studio to enhance the performance of generated videos, allowing the incorporation of more emotion, improved lip tracking, and natural human expressions into their training data.
Unlike many existing AI video tools that stitch together video snippets to simulate conversation, Synthesia's avatars are generated instantly, aiming to offer a more authentic and lifelike interaction that mirrors real human reactions and speech. With this release, Synthesia claims to have developed "the world's first avatars fully generated with AI”.
My Initial Thoughts: I definitely saw about a 20-30% improvement in vocal cadence and tone between the two versions, but it’s still pretty obvious to me it’s AI generated. We still got some work to do, team.
Housekeeping:
I made a GenAI uDemy crash course, and it’s on sale this week. If you haven’t gotten it yet, check it out here: https://www.udemy.com/course/gen-ai-crash-course/?referralCode=CF12E2E9C12C217B0514