Netcompany Snippets #12

?? Curious about our work? Visit us right here: https://netcmpy.com/Snippets

Looking for low-hanging fruits in the garden of science

We discovered general relativity and quantum mechanics over 100 years ago – what have we done since? A few things, but likely none as big. The already classic paper Are ideas getting harder to find? proposes a hypothesis: there are fewer low-hanging fruits than there used to be, leading to declining returns to R&D. Maxwell Tabarrok of Maximum Progress, however, is critical : Could the explanation not just as well be misallocation of resources?

Unsure what to major in? Try the classics

The Vesuvius challenge awards $ 1.000.000 to whomever uses machine learning and computer vision to read the Herculaneum Papyri, a potential treasure trove of interesting information. Now, the initiator of the challenge, Nat Friedman, announces the first breakthrough: the word "πορφυρα?", meaning "purple dye" or "cloths of purple" has been uncovered by a 21-year-old computer scientist. Byrne Hobart immediately jumps to the witty conclusion : “There’s no alpha in majoring in math or CS because you’re optimistic about AI. You have to consider downstream effects and major in classics instead”.

Look for the heavy tail – and sample heavily

“Most important things in life are heavy-tailed”, Ben Kuhn writes . Height is light-tailed: the tallest human is only a few inches taller than the shortest human. The impact of ideas, however, is heavy-tailed. The top 100 most-cited papers have over 12,000 citations, while the median paper has just one. The takeaway? If you are sampling from a heavy-tailed distribution, look for outliers and search the entire space properly.

On hype, bullying, and the scientific environment

The publication of biologist Lee Cronin’s “Assembly theory explains and quantifies selection and evolution” in the journal Nature has caused quite a stir . Evolutionary biologist Yogi Jaeger writes on X that “assembly theory clearly does not explain evolution or selection” and clarifies why in a lengthy blog post. Cronin’s idea is that early evolution favored or “selected” some chemical reactions over others. Later, Darwinian selection favored objects that were good replicators. Over time, objects with longer and more complex histories were built. Above a certain complexity threshold, only a lifelike process could have created the given object. Does this explain and quantify selection and evolution?

The EU is more productive than the USA

“The EU economy is now 65% the size of America’s in dollar terms, down from 90% just ten years ago”, The Economist writes . Adjusting for Purchasing Power Parity, or PPP, the EU’s GPD is roughly 95% of America’s. And adjusting PPP for hours worked, countries like Denmark and Norway exceed America in productivity.

Are generative models world models?

A common critique of generative models such as ChatGPT is that they cannot plan. However, a new paper argues that “modern LLMs acquire structured knowledge about fundamental dimensions such as space and time and learn... literal world models”.? Some researchers are not convinced .

Anthropic achieve AI safety breakthrough

If we can understand individual components of a neural network, that’s a good step towards understanding the network as a whole. At least that’s the working hypothesis of the area of mechanistic interpretability. However, one problem is that individual neurons respond to several things simultaneously, such as faces and cars. Formally, they are polysemantic. Until now, this has posed a huge problem for understanding neural networks mechanistically. It has also meant that individual neurons might not be the right level to look at. But now Anthropic has released a research paper that uses a novel method to discover monosemantic features of a neural network. Chris Olah of Anthropic writes on X that “I’m now very optimistic. I’d go as far as saying [mechanistic interpretability] is now primarily an engineering problem”.