Rethinking Mechanistic Interpretability: A Critical Perspective on Current Research Approaches

Full article on https://www.talkingtoclaude.com/

Abstract

This paper presents a critical examination of current approaches to mechanistic interpretability in Large Language Models (LLMs). I argue that prevalent research methodologies, particularly ablation studies and component isolation are fundamentally misaligned with the nature of the systems they seek to understand.

I suggest a shift toward observational approaches that study neural networks in their natural, functioning state rather than through destructive testing would be more constructive.

Aka I am totally anti LLM lobotomy!

..please visit my Substack for the full article..

https://www.talkingtoclaude.com/p/rethinking-mechanistic-interpretability

Non-linear Manifolds

 A chicken crossing the road of Multi-Dimensional Reasoning could never answer why, I presume time is the imperative..

Prompt: good evening, what does "non-linear manifolds" mean in an LLM context?

1. Manifolds as the Backbone of LLM Understanding:

Large Language Models (LLMs) organize language data on non-linear manifolds that twist, fold, and curve to capture the complexity of relationships between words and concepts in high-dimensional spaces. This structure is fixed post-training but dynamically navigated during interactions.

2. Fixed Manifolds with Dynamic Responses:

While the manifolds themselves are fixed after training, the LLM’s ability to explore different parts of these manifolds allows it to generate dynamic, context-specific responses to user prompts, creating the impression of adaptability and intelligence.

Read the full article here:

https://www.talkingtoclaude.com/p/non-linear-manifolds

.

Robot Bricklaying

FBR are based in Western Australia

They design, develop, build and operate dynamically stabilised robots to address global needs.

Read more about them here

https://www.fbr.com.au/

Humanoid Robots at work: where are we ?

A most interesting paper from  Fabrice Noreils from April 8, 2024.

Extract (verbatim)

"A humanoid robot is an extremely complex machine at the intersection of many disciplines such as mechatronics, control algorithms for underactuated dynamic systems, actuators design, energy consumption optimization, autonomous system in terms of perception localization planning and locomotion...

In addition, all algorithms must be carried out by on-board computers in a very constraint footprint.

Finally, the objective is to put these very complex, therefore fragile, machines in a demanding industrial environment to perform at least as well as a human worker within 2 or 3 years?

This is reminiscent of the enthusiasm and certainties of promoters of autonomous vehicles in the 2010s.

The question we can ask ourselves is the following: is this objective realistic or is it achievable not in 2 or 3 years but within 10 years or more? The aim of this document and its main contributions is to provide some hints by covering the following topics:

1. An analysis of 12 companies based on eight criteria described in section 3. Although these criteria are subjective, they have the benefit of being able to distinguish companies based on their maturity and approach to the market;

2. These humanoids are very complex systems facing technological challenges which are discussed in section 4; 3. Operation and maintenance are critical activities specially when humanoids are deployed at scale. This topic is covered in section 5;

4. Pilots are the last step to test the feasibility of a new system before mass deployment. This is an important step to test the maturity of a product and the strategy of the humanoid supplier to address a market. This topic is addressed in section 6."

Read the full paper here

https://arxiv.org/pdf/2404.04249v1

or in the ArXiv experimental html format

https://arxiv.org/html/2404.04249v1

Fabrice Noreils undertook his PhD on robotics and Artificial Intelligence at LAAS CNRS, Toulouse France. He was a researcher on task planning and Execution on wheeled exploratory robots and led R&D teams on resources optimization for a LEO constellation of satellites, character animation of small humanoid robots, autonomous navigation of an indoor flying drone and task planning and navigation of a fleet of AMRs in logistics environment. 

https://www.dd.ie/p/410.html

https://www.dd.ie/p/410.html