AI Top-of-Mind for 5.29.24 — Black Boxes
Top-of-mind is new insight into the inner workings of LLMs from Anthropic. What their researchers uncovered is a ‘brain map’ of sorts showing where different ‘features’ are in play. I’m thinking that, in the same way we have maps of word associations for LLM vectors, maybe the same idea for LLMs could show associations between different concepts. There are also parallels to DNA mapping.
From a related ‘NY Times’ article:
· One of the weirder, more unnerving things about today’s leading artificial intelligence systems is that nobody — not even the people who build them — really knows how the systems work.
· The researchers looked inside one of Anthropic’s A.I. models — Claude 3 Sonnet, a version of the company’s Claude 3 language model — and used a technique known as “dictionary learning” to uncover patterns in how combinations of neurons, the mathematical units inside the A.I. model, were activated when Claude was prompted to talk about certain topics. They identified roughly 10 million of these patterns, which they call “features.”
· They found that one feature, for example, was active whenever Claude was asked to talk about San Francisco. Other features were active whenever topics like immunology or specific scientific terms, such as the chemical element lithium, were mentioned. And some features were linked to more abstract concepts, like deception or gender bias.
Closer look into a specific ‘feature’ and how it can be modified:
Related to the safety issues above, an update on Apple’s plans to ensure privacy. ‘The Information’ and ‘The WSJ’ both report on the concept of a ‘virtual black box’ where employees won’t have access to any user data stored in Apple’s cloud. The approach, termed ‘Apple Chips in Data Centers,’ will rely on the secure enclave feature in Apple’s hardware.
Are you caught up in the Gen AI craze, and not sure when and when not to use it? Cezary Gesikowski in ‘Generative AI’ looks at the current state of play, including some analysis by Gartner. Some good guidance as you look to introduce AI into your enterprise.