Every time I tweet something skeptical of the current AI hysteria, I get a guy or two who swear they use it all the time for programming.

But 18 months into this supposed revolution I'm yet to hear about: 👇 - A new data structure in any way superior in space or time complexity to exiting ones made by AI
- A new general use algorithm superior in any way made by AI
- A security vulnerability discovered by AI
- Any SW project at all developed by AI

I little thread on how Gaussian (aka Normal) distribution arises, why it seems to be everywhere and why under closer inspection it is almost nowhere. Jump in👇 Gaussian distribution is so prevalent, because it arises as a limit of averaging out many independent random variables with finite variance. This fundamental law of statistics is called a Central Limit Theorem or CLT. en.wikipedia.org/wiki/Central_l…

Small thread: 👇1⃣ The surprising success of machine learning in the last decade stems from the fact that in many tasks roughly 80-90% of variance for classification/regression is contained in relatively low level statistical features. 2⃣ On the flip side, the source of apparent failure of ML techniques in robotics (Moravecs paradox) and mission critical autonomy is because the remaining 10% of variance is hidden in hellish complexity. Hence the march of 9's is becoming much harder with every 9.

The shutdown of Argo AI is certainly a silent signpost on the road to autonomy. But just as I been saying before, the road to autonomy we follow now has more than one step, in which the best idea how to solve it can be summarized by:

¯\_(ツ)_/¯ It all got ignited by DARPA challenge in 2005 in which a few teams managed to get a car through tens of miles of a desert road. Got noticed by Google, money started flowing. Next what followed was a deep learning revolution which at least on the surface provided the missing magic

For years I've been claiming that our #AI being based in statistics is not equipped for dealing real world data. That data is most likely fat-tailed, and hence not possible to characterize by sampling. Let me explain this in thread below: 👇 First of all, the basics: random distributions can either be short-tailed (finite variance such as Gaussian) or fat tailed (e.g. Pareto). The nice property of finite variance distributions is that we can learn almost everything about the distribution by sampling it.

A long time ago (1980's) in a country far far away (Poland) in reality that no longer exists, there was a science show on TV. A show unlike any other, unique to this day: Sonda. Thread👇 1/9 Each episode was focused on some piece of modern technology or recent scientific discovery. However, in reality, the technology wasn't ever the star of the show. It was the dialog between the hosts that really made this show unique.

For safety-critical applications, how things fail is potentially more important than how often they fail. Thread 👇 1/9 For example, a diving regulator valve can potentially freeze over. But it's made in a way that if it does, it does it in an open position, feeding air constantly rather than just when the diver inhales. That way, the diver immediately knows something is wrong and he needs to

I've had a lengthy discussion today about the SpaceX mission. I'm observing the following - people act like this is some kind of breakthrough, at the scale almost of the Apollo program. Now I don't want to demean the effort, space is hard, but the reality as I see it is: We are sending a capsule on top of a rocket to LEO. This has been done thousands of times by various space programs, and was done regularly in the US in the 60's. Aside from some external sugar such as LCD screens, this technology hasn't fundamentally changed since then.