Travis scodt

@11vats11 @raghumahajan But it kinda was- low energy supersymmetry was proposed to be discovered at LHC, and the whole field bet on it! A large fraction of initial LHC papers were SUSY!

@raghumahajan i suppose the supersymmetry research won't be capital heavy if it's mostly theoretical atp so all the chaos shouldn't affect it much? idk if it needs a lot of compute... the physics undergrads are less likely to be influenced by the media fervor anyway.

In popular media these days, some very mathematically deep ideas in theoretical physics are portrayed as unworthy to pursue since they have not yet made contact with experiments, despite having been studied for decades. Supersymmetry is one of them.

@raghumahajan We are not studying science for its economic value but it must eventually make contact with experiment to be science. SUSY was proposed to solve the hierarchy problem which it didn’t. Now you need to make a testable prediction, or prove that it is inevitable.

@HelloVyom So the ones doing research aren’t the ones taking risks, but founders are? The two in physics went on trajectories that are next to impossible from India. Even the one at openai. These are the minds building the real future!

24 years of JEE toppers, and... Only 1 became a founder. Let that number sit for a second This is the top 0.01% of teenagers in India, who have braved their way through one of the toughest exams in the world Where are they now? 7 became researchers/professors 6 became Big Tech employees 5 are in finance/quant 1 is an IAS officer 1 is at OpenAI 1 is a founder India finds the smartest kid every year, and celebrates them in newspapers and TV. What happens next? They go build someone else's dream in another country We have built a system where the most brilliant minds in the country optimize for safety, not leverage The real test isn't a competitive exam It's whether you have the guts to bet on yourself

@octonion Look for either multimodal peaks at centres or quantify anomalies >5\sigma from this distribution to detect individuals.

@octonion 1. Yes, they do, but I am saying that a gaussian fit to fat tailed outliers will be even wider yielding more outliers than baseline. 2. I agree, we would need data in all dimensions like subject, paper and center. Then model the CDF using some regularized distribution and …

Assume n≈60000 students, with a rough between-paper Gaussian delta D_i ~ N(μ,τ^2). If μ is small, "dozens" of observed 100+ gaps corresponds to τ≈26-30. Now M = max |D_i| ≈ τ sqrt(2 ln(2n)), yielding M ≈ 126-145. An observed max in this range is in line with this model. 1/

@octonion This provides no information whatsoever about anomalies or cheating!

@octonion You have not only assumed normality, you have also assumed dozens of 100+ gaps is normal- thus assuming a standard deviation \tau . By “fitting” a gaussian this way to a fat tailed distribution( someone plotted it) you will predict many more outliers than there are!

@agrawalmanindra @yomritoyj @saketkc Aah that is exactly what I am contesting, there is absolutely no set of scores with upper and lower bounds that could violate chebychev’s inequality. Only if you sample from something like a cauchy distribution with infinite variance could you violate it.

@yomritoyj Chebyshev gives upper bound, and if numbers are above it, it certainly raises alarm. In reality, numbers are much lower. To get more clarity, further analysis is to be done. That is done by @saketkc nicely.

Prof. Agarwal is a world-renowned theoretical computer scientist, so I'm sure he had some subtle argument in his head. But on paper this does not make sense. Chebychev's inequality gives an upper bound and not a lower bound.

@agrawalmanindra Dear Prof. Agarwal, is there any scenario you would detect cheating statistically? Do you try anomaly detection using modem ML and stats on paper scores by subject and center? And compare it to baseline/trend over years? That is what these kids are asking for.

Claims like below show a lack of understanding of basic statistics. When there are 60,000 students, some will have large gaps in marks for two papers. Chebychev's inequality quantifies it. It has been the case in JEE Advanced since two papers were introduced. 😊

@pravesh If you could cook up any distribution of bounded numbers like scores that violates chebychev’s inequality then it would be relevant. This is just a fun exercise checking 1+1=2. It has nothing to do with anomalies or cheating by extension. You know this.

This is amazing. I taught a course in statistics this year and this is such cool illustration of what is known as Chebyshev inequality. Basically, it says that for any probability distribution, ANY, the probability that your sample from the distribution is away k times the standard deviation (from mean) is upper bounded by 1/k^2. The random variable here is D=P1-P2. The difference in marks of two papers of a student. The population is 60000, so you have 60000 samples. The mean of D would be nearly zero (or some positive number of P1 is easier than P2). So even if the variance is say 20 marks, you would expect say around (for k=4), 1/16 * 60000 students (nearly 375) to have difference of 80 marks.. and even 200 odd students to have difference of 100 marks between their two exams. These are upper bound of course. So what has happened is neither statistically unusual, nor does it indicate any cheating. And there is a good chance that the distribution is Gaussian, which means the upper bound is likely tighter than we think..

@ravihanda That is exactly why I tried to make him clarify his point, since he is answering tweets! Especially for a school audience without statistics training.

People are arguing with him on basic statistics in comments. He is a winner of Godel Prize + Fulkerson Prize + Director at IIT Kanpur and he is still answering comments with calm and logic. Teachers are built different.

@agrawalmanindra How do you quantify outliers? I feel the distribution of 2 papers is impossible to model, except you could compare with previous year distributions say with kl divergence or look for multimodal peaks, but you could only quantify group cheating not individual!

@Travidscot Sure. My point is that existence of outliers does not provide proof of cheating, as the original poster was trying to suggest.

@tishasaroyan is of indic origin and not from indo aryan migrants. Also, since we haven’t deciphered the script or the language of ivc, a lot more of hinduism and language may have had continuity from pre indo Aryan migration. Whether pushpati is specifically shiva is irrelevant!

@tishasaroyan If hinduism is defined to begin with composition of the rigveda, then everything that inspired it and everything it evolved from obviously predates it. The fact that the pushpati seal is in a yogic pose and evidence for yoga goes back to ivc implies that atleast some of hinduism

Amish, you're an idiót! Yoga's origins are not in Hinduism, in fact Yoga predates Hinduism by at least 500 years. Yoga is evidenced in Indus Valley and then adopted by the Vedic people arriving in waves from 1500 BCE. Read the Atharva Veda if you have half a brain and this should become clear as a day!! Therefore, the figure is not Hindu simply because it sits in a yogic posture. There's no Hinduism/Vedic religion in the Indus Valley, or anywhere else in the Indian subcontinent before 1500 BCE. Furthermore, Shiva as a yogi is a Puranic construct, the Vedic Rudra, which comes before Shiva is not a Yogi. If seal 420 figure was some form of Shiva it would be closer to Rudra in its form and features and not to the Puranic attributes of Shiva, who comes chronologically after Rudra.