#idea

System 1 and 2

I forget where this was mentioned (either in one of the AI podcast episodes, or this numenta video), but basically we can think of GPT-3 as being the first almost perfect copy of system-1 human thinking, which is how Kahneman chose to dichotomize how our brains work—essentially, system-1 is the fast, intuitive thinking, while system-2 is the deliberate, rational, logical thinking.

Pattern recognition is basically system-1, and it’s where all the problems of correlation ≠ causation occur, since it’s just focused on predicting things by association. And that’s basically what GPT-3 is capable of doing.

The question is then how do we get to system-2 thinking, which is pretty much our competitive edge—deliberate thought. Here’s a random #idea that I had on my run, and I suspect someone has already thought about: what if system-2 = system-1 + simulation? It seems to me that the crucial piece of the puzzle is basically being able to simulate the world, or at least some very crude model of it. Once you have the capacity to simulate the world, then you can run your system-1 inferences, and see how things compare to the truths of your simulation, while also making sure to update your model of the world against reality.

Project Misinformation

Goal: detect disinformation

Resource: Awesome list

context: twitter/facebook/social media.
data:
- the text/content: you have a tweet that is basically a headline of an article, for instance
- source: where is the headline from? (e.g. nytimes)
- user covariates: demographic information of the sharer
- information cascade: this is catch-all phrase for everything that happens with the sharing of the post
  - what is the sharing like? grass-roots or shared by influencers
  - what are the responses like (content in the retweets, say)
  - demographic of the sharers/clustering?
I make the distinction between mis- and dis-information because I think dis-information is the much more pernicious problem. dis-information is not about the politics, but just the self-aware nature of the information.

Literature Review

#paper paper on misinformation with neural network
- they use something known as “cascade model”, which takes advantage of the twitter architecture to capture responses/retweets, and use the content of the retweets to help classify the truthiness of the original tweet
- #idea there must be something here that allows you to merge ideas of nlp/tweet responses/the underlying social network
#paper Bias Misperceived: The Role of Partisanship and Misinformation in YouTube Comment Moderation
- this is a little different, but it also deals with partisanship: there’s a dataset that has partisanship scores for websites (which then gets linked to Youtube videos in some weird)
- thesis: is there political bias in terms of youtube comment censorship
#paper survey on misinformation 👍
- covariates:
  - source
  - content
    - lots of “descriptive” results on the traits of fake news headlines (longer titles, more capitalized words)
  - user response (on social media) (cascade)
    - propagation structure
- methods:
  - cue/feature, which is basically the pre-NLP era way of doing linguistic analysis
    - lie detection: linguistic cues of deception
  - deep learning based methods: this is what we want to target
    - #paper FakeNewsNet: A data repository with news content, social context and dynamic information for studying fake news on social media
      - supposedly, this paper shows that these kinds of methods have bad prediction scores (on the new dataset)
  - feedback-based (covariates/secondary information)
    - propogation
    - temporal
    - response text
    - response users
- something that we haven’t even talked about, is intervention: what kinds of methods are available to combat these bad actors.