The explosive claim that’s upended the AI world

illuminem summarises for you the essential news of the day, reviewed by our editorial team. Read the full piece on MIT Technology Review or enjoy below:

🗞️ Driving the news: Nick Cammarata, a researcher at AI startup Goodfire, argues that neural networks may be more understandable than previously assumed
• His 26,000-word blog post challenges the widespread belief that AI models are fundamentally tangled and difficult to interpret, raising hopes that researchers could eventually better understand, monitor and control advanced AI systems

🔭 The context: Cammarata revisited earlier research on superposition, in which individual neurons appeared to represent multiple unrelated concepts
• Studying an image-recognition model called Inception v1, he found that a neuron responding to both cats and cars might instead recognise a shared visual pattern: a V shape
• He believes this suggests that seemingly confusing neural activity may contain more interpretable logic
• However, his latest claims remain largely a hypothesis, and the model he studied is far simpler than today’s leading language models

🌍 Why it matters for the planet: More interpretable AI could potentially support the responsible use of increasingly powerful systems in areas such as energy efficiency and climate research

⏭️ What’s next: Researchers are awaiting technical results from Cammarata’s work on modern large language models
• Experts including Stanford’s Chris Potts and Google DeepMind’s Neel Nanda caution that understanding individual neurons is not the same as explaining how entire networks operate
• Demonstrating that the approach scales to advanced models will be essential before claims of a major breakthrough can be substantiated

💬 One quote: Nick Cammarata, researcher at Goodfire, said: “We were looking at it wrong.”

📈 One stat: Less than 0.01% — Jacob Hilton of the Alignment Research Center’s estimate of how much researchers understand about the inner workings of top-tier large language models

Subscribe to our free newsletters to never miss a beat in sustainability.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.