As tech & VC twitter are exploding in excitement over the human-like capabilities of GPT-3, I'd like to remind everyone that at short time scales, "intelligence" is in the eye of the beholder. Here's of thread of examples where it is easy to fall victim to anthropomorphic bias:

Here's a video of Google's QT-Opt system that tries a couple times to pick up an object. You'd think that it's "retrying" with new grasp poses, but the model actually is memoryless.

Ke Jie initially declined to play AlphaGo / AlphaZero for fear that it would "copy his style". AlphaGo is memoryless, so it could not possibly "adapt" to Ke Jie's style.

Here's a fish swimming upstream. You might think that this task requires considerable (neural) intelligence, with sensing current and performing some optimal control.
https://vimeo.com/44887922 

But the fish is dead.

In Merger Arb: one might think that in order to be profitable, one must take the daunting task of predicting which mergers succeed and which ones don't. But it turns out that 90% of the returns can be replicated by assuming all mergers go through and simply indexing.

Animators exploit anthropomorphic bias all the time to craft relatable stories to inanimate objects like sacks of flour
https://www.deviantart.com/tamhorse/art/Finished-Flour-Sack-Walk-282804902

Predicting certain financial quantities like SPY volatility is usually easy. 90% of it can be forecasted with simple historical data and simple models. 99% of it can be forecasted with some additional datasets and simple models. Last 1% requires a human brain.

A recurring theme here is the "long-tail" of intelligence in human behavior. Hypothesis: large % of our convos, thoughts, and behavior can be replicated with simple circuits, and that's why their output can seem "humanlike" (because they are indeed typical within a short window)

https://twitter.com/cweichen/status/1284716329199706112?s=20