Alex Wissner-Gross

This notion of an economy that has an inner spiral that's going to ultimately consume and disrupt the rest of the economy, I think is front and center to the notion of the singularity.

To the first half of your question about whether LLMs not just have value, but where all of the competition goes, all of my friends at the Frontier Labs call it a rat race.

There is very much a race to the bottom in some sense of driving the cost of intelligence

So low that it's effectively too cheap to meter and used to be every year or so.

It's a way back five years ago when it used to be maybe an annual event that we'd get a new frontier model that would push the state of the art.

And then it was every quarter when we saw the move from LLMs to reasoning models.

We can talk more about that.

And then more recently, with models that are recursively self-improving and designing the weights or other properties of their successors, we're seeing new frontier models come out on arguably an almost weekly basis.

Soon, I think it's going to be daily, hourly, minutely.

We're going to reach, to the extent we haven't already, some sort of takeoff.

Sure, so one can point to a number of inflection points in the history of AI.

One can point to the 1980s when my friend Yann LeCun first developed convolutional neural networks, and then they were used to spot and to identify zip codes by the US Postal Service, but not very much else for their first few decades.

Fast forward to 2012 when the ImageNet large-scale computer vision competition

created the world's first data set, the first corpus of lots of annotated images, more than a million images that were curated and annotated and labeled with, this is a dog, this is a cat, this is a car, here's the bounding box within the image.

And thanks to that competition, we learned that convolutional neural networks, thank you, Jan, were actually quite good at image classification.

And we saw the first ML boom going from an AI research community where algorithms were chosen artisanally and religiously and wars, religious wars were fought over which approaches to AI were best and which ones were worst.

to a world where the benchmarks dominated and whatever did well at the benchmarks, that's what the community ran with.

That's 2012.

Fast forward from 2012 to arguably sometime around the summer of 2020 when we discovered, in addition to convolutional networks,

being really amazing, arguably a successor to convolutional neural networks and also long short-term memory networks, LSTMs.