Technology

Art

Jonathan Frankle: AI has fusion-grade models but lacks the 'power lines' to connect them to human intent

Ideas powered by Invisible Machines

25d ago· 5 min readen

technology business ai infrastructure model evaluation

Summary

Jonathan Frankle, Chief AI Scientist at Databricks, argues that the AI industry has built "fusion-grade" language models but lacks the infrastructure—specification, testing, and applications—to connect these models to real human intent. He draws a parallel to electrical grids: powerful models exist without the "power lines" needed to make them useful. Frankle critiques academic benchmarks like HellaSwag for being disconnected from what users actually want, and traces his thinking from his earlier work on "data mixology" at MosaicML through to the challenges of scaling AI at Databricks.

Source

UX MagazineJonathan Frankle: AI has fusion-grade models but lacks the 'power lines' to connect them to human intentuxmag.com

Key quotes

· 3 pulled

Three years after data mixology, Frankle argues we built fusion-grade intelligence but not the power lines—specification, testing, and applications that connect models to human intent.

Academic benchmarks like HellaSwag tell you almost nothing about what ChatGPT users actually want.

We built fusion-grade intelligence but not the power lines—specification, testing, and applications that connect models to human intent.

Snippet from the RSS feed

Jonathan Frankle on fusion-grade models without grid infrastructure, specification beyond benchmarks, and the through-line from Mosaic mixology to Databricks scale.

You might also wanna read

Examining the Limitations of Transformer Models and the Gap to Human-Level AI

The article presents a skeptical perspective on claims about imminent Artificial General Intelligence (AGI), arguing that current transforme

dlants.me·4mo ago

A Chinese AI gains ground on the most powerful models of Claude and ChatGPT

Ara·3d ago

Exploring FORTH and Associative Languages as Better Fits for Transformer Architectures

The article explores the idea that FORTH and associative/applicative programming languages might be better suited for transformer architectu

rescrv.net·5mo ago

ADViCE

turing.ac.uk

Databricks Co-Founders on Building an Open Frontier Ecosystem for the Agent Era

In a rare double-interview, Databricks co-founders Matei Zaharia and Reynold Xin discuss the company's evolution beyond the lakehouse into a