AI’s data dependency & the rising cost of electricity

As AI models lean heavily on publisher content, rising demands from data centers impact U.S. electricity bills. Plus, check out five new AI hacks for developers.

November 05, 2024

Juggling tasks, deadlines, and ideas from all directions? Atlassian’s got your back. With tools like Jira for planning, Trello for organization, and Confluence for seamless collaboration, it’s like having a project command center that everyone can get on board with. Ideal for remote and hybrid teams, Atlassian keeps everything—from brainstorms to big projects—in one place, so your team can stay focused and in sync.

Try it out for free and see how much easier work can feel.

Latest headlines 👇

A Ziff Davis study reveals that AI giants like OpenAI and Google rely heavily on premium publisher content to train their models, more than they disclose publicly.

  • Key Findings: Analysis showed that 10% of data in some training datasets came from top publishers, which could strengthen publishers' case for copyright protection and compensation.

  • Industry Shift: Many media companies now focus on longer-term content deals for AI chatbots rather than one-time training licenses.

  • Legal Impact: An ongoing lawsuit by The New York Times against OpenAI could set legal precedents on AI’s use of scraped content.

New AI data centers are pushing electricity demands sky-high, leading to increased utility bills for American consumers. As tech companies expand their energy-hungry data infrastructure, utility planning documents reveal that power grid strains are forcing utilities to make costly investments. Neil Chatterjee, former chair of the Federal Energy Regulatory Commission, pointed out that while data centers promote economic growth, they come at a cost to consumers. Read more

Five new AI hacks for developers

  1. Model chaining: Developers are reducing costs by chaining cheaper models to summarize data before passing it to a more advanced model. Dara Ladjevardian from Delphi shared that he uses open-source models like Llama or Mistral to summarize information, which he then feeds to high-end models like GPT-4o, saving costs on context-heavy queries.

  2. Automatic prompt optimization: Prompt engineering is being automated with tools like DSPy and SAMMO. DSPy tests common prompt tweaks, while SAMMO offers broader capabilities, such as modifying prompt components for better responses, eliminating the need for manual prompt adjustments.

  3. Enhanced output verification: Ensuring accuracy in LLM responses can be challenging. Jason Liu suggests requiring LLMs to use exact phrases from reference documents, allowing developers to verify these responses by matching wording with source documents automatically.

  4. Synthetic data for product evaluation: AI-powered product testing often requires costly, manual testing. Liu advises using LLMs to generate synthetic data, like sample legal questions, to simulate real scenarios and improve evaluation processes. This method saves time and resources.

  5. Computer agents for app testing: With new AI capabilities, developers can use computer-using agents, like Anthropic’s, to autonomously test applications, detect bugs, and potentially even fix them. Guillermo Rauch from Vercel sees this as a future of streamlined app testing.

Meta’s AI leverages vast amounts of user data from Facebook, Instagram, and interactions with its chatbot to create highly personalized content, but offers no opt-out for most users. The data fuels AI-generated images and tailored feed content to drive engagement, aligning with Meta's ad-driven model. The company may soon use generative AI for personalized ads as well, making it a leader in AI-powered social media. Read more

AI search company Perplexity has launched an Election Information Hub to provide voters with quick answers to key election-related questions. Using data from Democracy Works and The Associated Press, the hub covers polling locations, candidate summaries, and will offer live vote counts on Election Day. Perplexity says it prioritizes non-partisan, fact-checked sources, though initial glitches highlighted the challenges of relying on AI for accurate election information. Read more

Generative AI is increasingly used to create realistic scientific images, raising concerns about a rise in faked data in scientific literature. Experts, like image integrity analyst Jana Christopher, are worried that AI could flood journals with untrustworthy visuals, making fraud harder to detect. To counter this, tools like Proofig and Imagetwin are being developed to spot AI-generated images, though human oversight remains crucial. Many hope these advances will eventually outsmart today’s AI-enabled fraud. Read more