AbsurdumRedux

1==0

[AbsurdumRedux @ReduxAbsurdum]

If the rumours are to be believed, Jane Street’s ARR > 100B and HRT > 40B. I am assuming CitSec > 20B. So, three companies, less 10000 people and certainly less 2000 researchers making 160B+. Only a matter of time when we will see new trading companies being funded everywhere.

[AbsurdumRedux @ReduxAbsurdum]

@ChasingRainbw @MohapatraHemant In the field in which I work (Quant Finance), many top firms are outside of the US. India also has a couple of exceptional firms. One can in fact say that AI is closest to my field and London is as big as any other place (London founders in London not offshoots of American cos)

[AbsurdumRedux @ReduxAbsurdum]

@ChasingRainbw @MohapatraHemant Additionally, if the general direction is that AI will penetrate all areas, specialisation based on context will matter. Indian context, local innovation and feedback over time might make our models more valuable even if for deep science, SOTA models might still win.

[AbsurdumRedux @ReduxAbsurdum]

@MohapatraHemant I don’t think we should worry about SOTA (even though it would be nice to be SOTA). So long as the model is reasonable to get commercial traction, this will be rewarding in the long run. Commercial traction will come by way of most use cases not needing to be SOTA. +1

[AbsurdumRedux @ReduxAbsurdum]

@MohapatraHemant Neither our nuclear nor space program is SOTA. Bhabha and Sarabhai could despite the tech being closed and sanctions risk, we should be able to do so here.

[AbsurdumRedux @ReduxAbsurdum]

@MohapatraHemant I don’t think we should worry about SOTA (even though it would be nice to be SOTA). So long as the model is reasonable to get commercial traction, this will be rewarding in the long run. Commercial traction will come by way of most use cases not needing to be SOTA. +1

[AbsurdumRedux @ReduxAbsurdum]

@agrawalmanindra @WarriorGam65730 If one assumes a model of errors, one can use Chebyshev or other concentration inequalities to invalidate the model itself. That idea forms the basis of statistical tests. However, you may not use the data itself to estimate the variance and then expect Chebyshev to be violated.

[AbsurdumRedux @ReduxAbsurdum]

@ritwikshukla @Shubham45856917 @agrawalmanindra @pravesh Sorry, can you tell me what changed in 2018?

[AbsurdumRedux @ReduxAbsurdum]

@saketkc @pravesh @agrawalmanindra Or is there a different application of Bonferoni that you are alluding to?

[AbsurdumRedux @ReduxAbsurdum]

@saketkc @pravesh @agrawalmanindra Under the independence hypothesis (in the absence of corruption), it is unclear how 4x at 100 isn’t alarming. I do note that the kurtosis you report is less than that of the fitted distribution (and which explains the tails) which explains this but it remains anomolous to the fit

[AbsurdumRedux @ReduxAbsurdum]

@Shubham45856917 If one models a student as an oracle with some noise, the score obtained by the student is random variable which becomes increasingly concentrated towards the true score (skill). Mean increases lienerly and std by sqrt(N). Under this prior, 100 marks diff is highly unlikely.

[AbsurdumRedux @ReduxAbsurdum]

@Shubham45856917 @agrawalmanindra @pravesh No identity revealing questions please!

[AbsurdumRedux @ReduxAbsurdum]

@Shubham45856917 @agrawalmanindra @pravesh Correlating the corruption with real life variables may be the only way to rule out cheating. That is labour intensive and does require a fair bit of work.

[AbsurdumRedux @ReduxAbsurdum]

@Shubham45856917 @agrawalmanindra @pravesh However, three is no reason to believe that such a tail phenomena will be stable. Small variations can change the process and hence the observed corruption rate. For instance, if the papers are different in one year, this rate may be higher.+1

[AbsurdumRedux @ReduxAbsurdum]

@SinaYanka @Shubham45856917 @agrawalmanindra If we do pursue this systematically, it might be worth studying all of this. Ideally, all occurrences should be either truly random or accounted for by extraneous factors. I am not sure if it is worth your time though.

[AbsurdumRedux @ReduxAbsurdum]

@SinaYanka @Shubham45856917 This is correct. To rule out cheating, one needs to study the tail and show the absence of confounding factors. Note that this isn’t as simple as geographical proximity is both a corroborating and a confounding factor (imagine excessive heat or mass cheating). Tail stats is fun!