AI Top-of-Mind for 8.27.24 — Fracking AI
Today: Positive AI energy consumption news, an agent reality check, Genie for software development, min-p, an update on C2PA, and Walmart’s approach to AI
Top-of-mind is AI energy consumption, with some positive news. By one estimate, data centers could consume 9 percent of U.S. electricity by 2030, up from 4 percent today. The ‘NY Times’ reports on advances with geothermal energy production and use by the likes of Facebook. The article details the use of fracking technology to generate geothermal power from areas that in the past were not viable:
· On Monday, Meta, the company that owns Facebook, announced an agreement with a start-up called Sage Geosystems to develop up to 150 megawatts of an advanced type of geothermal energy that would help power the tech giant’s expanding array of data centers. That is roughly enough electricity to power 70,000 homes.
· Sage will use fracking techniques similar to those that have helped extract vast amounts of oil and gas from shale rock. But rather than drill for fossil fuels, Sage plans to create fractures thousands of feet beneath the surface and pump water into them. The heat and pressure underground should heat the water to the point where it can be used to generate electricity in a turbine, all without the greenhouse gases that are causing global warming.
Two posts on agents. In the first, Kenny Vaneetvelde writing in ‘Level Up Coding’ looks at the state of play with agents and offers a reality check including pitfalls with deployment. From his post:
While AI agents undoubtedly have their place in the AI ecosystem, and a very important place at that, they are NOT a silver bullet solution to every problem. Most importantly, though, the way we implement them is of the greatest importance, yet there are 100 ways to go about it.
He then offers guidance on a better approach:
· Start with the Problem, Not the Technology
· Consider the Full Spectrum of AI Techniques
· Embrace Modularity and Composability
· Prioritize Explainability and Control
· Plan for Scalability and Performance
· Focus on Integration and Interoperability
In a second posting, Giancarlo Mori dives into five agent platforms for building agents, explaining features, best use cases, LLM support, and pricing:
· crewAI
· AutoGen
· LangChain
· Vertex
· Cogniflow
And related to the above, Gencai I in ‘Level Up Coding’ looks into ‘Genie’ for software engineering support and describes how to use the tool. From his post:
On the model front, Ignacio de Gregorio looks at ‘Min-p,’ an alternative to the widely used ‘top-p’ and one that can improve accuracy with word selection. From his post:
While top-p sampling chooses ‘one of the most likely,’ min-p actually takes into consideration the structure of the distribution, incentivizing the output of highly probable tokens when the result is obvious and without impacting the creativity of the model in situations where the distribution has a long tail.
A while back I wrote about an initiative to authenticate content as it passes through the various editing tools and is then published. Enrique Dans looks at the latest regarding the Content Authenticity Initiative (CAI) and the Coalition for Content Provenance (C2PA)… who is leveraging these, and what tools support this. He looks at slow adoption of the standard amongst camera manufacturers, for example, or non-Adobe editors. Same with publishers like the New York Times who support C2PA but offer no access to the data for images.
Turning to retail, some sage advice from Walmart, as reported by ‘Modern Retail.’ Some highlights from the podcast with Jon Alferness, the retailer’s chief product officer:
· The tying bind behind all of this is that new products need to solve for a real need. When it comes to AI, for example, the product can’t exist for its own sake. In fact, in his estimation, a new AI project shouldn’t even have the technology in the name.
· “From my point of view, it’s not important to say, ‘Hey, so we built such and such product, now powered with AI,’” he said. “I don’t think customers care one way or another. I think they just want their problem solved.”
A major area of focus for Walmart is ‘guided’ search and understanding the customer’s intent.
