Have not read the paper in much depth yet but this looks like great work, super interesting. Thanks for sharing.
Question: in the example of prediction on untrained tasks, what exactly hasn't been trained? The paper talks about video being one of the trained tasks. Did you simply retrain model without video examples and then test performance?
The model was trained on video classification, image qa and image captioning. Video captioning and video qa is not trained, yet the model shows results on those tasks.
These are all very fair points and things I thought through as I put this together.
For example, your point around <$15K salaries. As I explain in the appendix at bottom of post, I exclude anyone with <$10K salary for the exact reason that this is not feasible for someone claiming to work as a software developer in the United States.
Your point around monthly vs annual salary - this was a question on the survey itself, and hence the salary numbers are already adjusted to be annual.
The source code includes a few data cleansing actions like the above. In the end this is inherently subjective, but I tried to be principled in my approach in order not to unduly skew the results.
Unfortunately this is the best publicly available data we have, input errors and all.
Thanks for reading through and for your comment, this is important stuff.
This is an important point that is commonly overlooked.
Traffic is a sign of something. It’s a sign that the benefits of transportation outweighs the costs of transportation. Given this, high traffic often implies that there is pent up demand out there - more cars that would be on the road right now if only the roads could support it.
When more roads are built, this unlocks some of the pent up demand, as the cost of transportation has effectively declined due to more available roads.
However, this reduction in cost / hassle etc and the previously pent up demand may be so large that in encourages much more driving, such that in equilibrium the traffic could be nearly as bad as before.
It’s a sign that the benefits of transportation outweighs the costs of transportation
Which, given the way roads are funded, is something of a broken equation. The costs of transportation are not fully borne by the user proportionally to consumption.
You pay fuel taxes & buy tires, yes, but they by no means cover the full per-mile cost of your trip.
Not sure why you are downvoted, drivers are subsidised to a huge amount, and that doesn't even include pollution, deaths, destruction of wildlife, etc...
Austerity can involve tax increases, spending cuts, or combinations of the two. It’s not limited simply to spending cuts.
That said, it very much matters what “kind” of austerity one engages in, as research indicates that tax increases and spending cuts have different respective effects on an economy.
Like most economic principles, there’s a big gap between theory and practice.
In reality Austerity is defined by tax breaks for the very rich, combined with spending cuts and privatisation - i.e. economic enclosure, debt peonage, and asset sweating - for everyone else. Sometimes that includes tax increases, but if som it usually means indirect taxation.
Austerity is purely ideological. Even the IMF acknowledges that recent Austerity regimes have underperformed expectations while examples of stimulus spending have had the opposite effect.
> In reality Austerity is defined by tax breaks for the very rich, combined with spending cuts and privatisation
This is objectively false, and a poor atempt at forcing a particular spin on a very objective definition.
Looking at Portugal, the bailout program enforced tax increases exclusively on the middle and upper class, along with publix sector workers who are by far more priviledged than private sector workers. For example, the bailout program saw the introduction of a new tax bracket for the richest taxpayers which forced a 53% income tax rate.
And also privatizations are implemented to avoid additional austerity pushes by providing the government with one-off sources of free cash thus avoiding further spending cuts or tax increases. Therefore, they are in fact an alternative to austerity, not a consequence.
> Austerity is purely ideological.
This statement is so mind-numbingly wrong that it boggles the mind how anyone in this point in time could be so disingenuous or clueless to keep parroting this silliness.
Let's look at Portugal, who doubled its sovereign debt fron 60% of the nation's GDP to over 120% in about 5 years prior to any mention of austerity, and did so by piling a string of yearly deficits of over 10% including a structural deficit of around 3%.
If your state is so dependent on overspending that it needs loans after loans to keep working and pay up salaries, who in their right mind will argue that they don't desperately need to cut spending and/or raise taxes to avoid going bankrupt or even to keep functioning?
If austerity was actually "purely ideological", how exactly would it be possible for Portugal to keep their 10% spending deficits?
What do you mean by "the very rich"? Rich by income, assets or consumption? Currently, lots of governments overtax income and sometimes assets, while undertaxing consumption. They tax the likes of Bill Gates and Warren Buffett a lot (right up until the point when they-- or perhaps their descendents-- actively donate their assets to some charitable trust or another) while leaving the crazy luxury spending of Larry Ellison and others like him largely untaxed! This is crazy, and is what a move to consumption taxes (what you call "indirect taxes" in your parent comment) helps fix.
Sure - perhaps alchemy in the sense that many practitioners are simply throwing things against the wall and seeing what sticks, but not in the sense that there isn't anything real behind all the math or engineering.
Many advancements in machine learning have significant backing in theoretical proofs that a given algorithm will result in unbiased estimates, or will converge to such and such value etc.
On some level, the high amount of experimentation necessary in machine learning is not so much a sign that the practice is faulty in any particular way, but rather, that the world is a complex place. This is especially true when attempting to predict anything that involves human behavior.
Alchemists could do a lot, making gunpowder for example is non trivial. They simply worked from a bad model with little understanding of what was going on.
Consider, lead and gold are very similar substances and chemistry lets you transform many things into other things so it must have seemed very possible. Unfortunately, I suspect the current AI movement is in a very similar state even if they can do a lot of things that seems magical it’s built on a poor foundation. Resulting in people mostly just trying stuff and see what happens to work without the ability to rigorously predict what will work well on a novel problem.
> Now alchemy is OK, alchemy is not bad, there is a place for alchemy, alchemy worked, alchemists invented metallurgy, [inaudible] dyed textiles, our modern glass making processes and medications.
> Then again, alchemists also believed they could cure diseases with leeches and transmute base metals into gold.
> For the physics and chemistry of the 1700s to usher in the sea change in our understanding of the universe that we now experience, scientists had to dismantle 2,000 years worth of alchemical theories.
> If you're building photo-sharing systems, alchemy is okay. But we're beyond that.
> Now we're building systems that govern health care, and mediate our civic dialogue. We influence elections.
> I would like to live in a society whose systems are built on top of verifiable rigorous thorough knowledge and not on alchemy.
Thanks. Slowing realizing some of the benefits of dicts and trying to use them more in my code where I would have otherwise defaulted to another data structure.
They really are quite useful and intuitive to work with once you understand basic structure
I do sometimes worry about speed relative to other options though, especially in large chunks of data
I don’t think he gives himself nearly enough credit here
If I could pay net $137 and end up with an email subscriber list of 2.5K along with a burgeoning personal brand centered around an exciting topic I’m interested in, I’d take that any day!
Even just reading this makes me mind spin thinking of how I could replicate his success. Congrats are definitely in order
Think it should be kept in mind that what Taleb is arguing here is the limited usefulness of IQ in predicting success above a certain IQ level, let’s say the mean. He 100% acknowledges its usefulness for predicting success of lower IQ persons.
He mainly just wants to point out that any other measure that was as bad at predicting higher levels of performance as IQ is would not be held in nearly as high regard. IQ therefore is in a vulnerable position as a hallowed metric, which he detects and proceeds to attack, hence the Twitter thread.
Think he could have done it with less bluster and fluster, but that’s all he’s saying at the end of the day. Not nearly as controversial as some are making it out to be
Have you ever read any of his books? Bluster and Fluster is his thing. I wanted to put down "The Anti-fragile" until I figured out that his persona is kind of like his advice on prediction. A kind of caustic sorting hat. Now it's sort of his brand and it works fantastically for clickbait.
In my view the interesting part of IQ is not merely in its ability to predict human performance in areas where we expect intelligence to matter, but also in its underlying claim that there is even such a thing as a general intelligence (g), as opposed to bundles of domain-specific abilities.
Proponents of g don't claim that it's an alternative to domain-specific abilities. They claim that it exists in addition to domain-specific abilities, and is what explains the correlation between performance across domains. That is, people who do well on mathematical tests also tend do well on tests of verbal ability. The factor analytic model that yields g attempts to explain this by viewing mathematical ability as the combination of two factors: a domain-specific mathematical intelligence and a domain-agnostic general intelligence. A subject's performance on verbal tests would similarly be modeled as the product of their g and of a cognitive ability specific to verbal tasks. That a subject's performance on one type of test is partially predictive of their performance on the other type is then (according to this model) explained by the two cases sharing a single partial cause, g.
As another poster mentioned, don’t think that g is necessarily in opposition to bundles of domain-specific abilities.
I see g as an attempt at dimensionalality reduction - can the essence of what is admittedly a complex phenomenon be boiled down to a single metric which, while clearly not complete, is at least directionally correct / helpful.
g in some sense could simply be a low dimensional projection of these more complex bundles.
It is exactly that: think of it as the first principal component, of some group of tests.
Obviously this always exists. The argument is really about how big the 2nd component is -- if it were comparable, then talking about the 1st alone would be very misleading. And for a typical bundle of school-like topics (like math/english/biology exams) it's quite a long way down: IIRC a factor of 3 or 4?
If the bundle of subjects is different, then the size and meaning of the components will vary. In the tests the army does to slot recruits into all sorts of roles, I think manual dexterity is one of the things they care about which is almost uncorrelated with g. Eyesight is I think one of the only things negatively correlated with g (and it's thought to be nurture: bookish kids spend too little time outdoors).
Was IQ ever meant to be a measure of financial success? Or rather, did anybody, at any time, ever think that the “smarter” people were the more financially successful? I know I never thought that or observed that anybody else appeared to think that. Being “smart” meant that you were (under the just-world hypothesis) guaranteed a certain level of comfort: you’d end up with an office job where you say all day in an air conditioned office rather than breaking your back under the hot sun, but the super-successful types were either (depending on your biases) the ruthlessly selfish or the risk-takers.
Question: in the example of prediction on untrained tasks, what exactly hasn't been trained? The paper talks about video being one of the trained tasks. Did you simply retrain model without video examples and then test performance?