Sai Krishna

Air is a public good and in any society with common sense, it should be a top government priority to provide clean air.

Alas, as we wait for that common sense to dawn on us, let's at least mask up in the meantime.

#SkySky AirQuality

LogProbs LogProbs is one of the basic skills for a Prompt Engineer to have.

LogProbs is a cool Prompt Engineering skill.

🤖📈 #Statsky #Stats #AI #LLM

medium.com/@saikrishna_...

One key insight about the impact of LLMs on creative writing from Andrej Karpathy’s LLM Deepdive:

medium.com/@saikrishna_...

#AI 🤖📈 #LLM #WritingCommunity

When I write stories, I ensure that I have the climax first with all its payoffs 😁

#Poetry #TinyPoem

Survey of 416 residents. The polling group has a sampling error of 3% in their familiar territories. 95% Confidence Interval falls below 50%.

Yet the audacity/ignorance to claim that entire Greenland supports to be adopted by the US.

🏛️ #rstats #PASky 🤖📈 #Stats

When to brush your teeth? A good ANOVA study! I found this paper which did a simple ANOVA study to find out when should one brush their teeth!

𝐖𝐡𝐞𝐧 𝐬𝐡𝐨𝐮𝐥𝐝 𝐲𝐨𝐮 𝐛𝐫𝐮𝐬𝐡 𝐲𝐨𝐮𝐫 𝐭𝐞𝐞𝐭𝐡?

1. Brush twice a day — Morning and Night after dinner.
2. Morning, it’s better to 𝐛𝐫𝐮𝐬𝐡 𝐚𝐟𝐭𝐞𝐫 𝐛𝐫𝐞𝐚𝐤𝐟𝐚𝐬𝐭.

If you find it scandalous, read this blog!

medium.com/@saikrishna_...

#rstats #Stats #StatSky 🤖📈

Don’t wanna be the best

#Poetry #TinyPoem

Greed

#Poetry #TinyPoem

Trident of endeavour

Science works, finds truths
Business pays, fulfils desires
Power lies, frauds morals

#Poetry #TinyPoem #haiku

Happy New Year!

#Poetry #TinyPoem

Farts of a city

#Poetry #TinyPoem #AirQuality

Trying my hand at writing poems 🌝

Thanks to Douglas Kearney’s course! www.coursera.org/learn/poetry...

#Poetry #TinyPoem

TL;DR

Paper tried to compare air quality between land cover classes using Kruskal Wallis tests
link.springer.com/article/10.1...

But practical significance seems neglected for statistical significance.

What do you say?

#Stats #rstats 🤖📈

medium.com/@saikrishna_...

Grove air quality sensor + ESP8266 reporting indoor air quality to a google sheet every 10 minutes

#IoT

github.com/d-saikrishna...

“Air quality data saves lives. Where air quality is reported, action is taken, and air quality improves.”

Oh I so wish 🌝

Which place is this?

Real estate broker working with Linear Regression on imbalanced data I used Housing price data for this analysis. Previous blog based on the same dataset are:

Depends on what we are doing with the data. A regression on it will optimise for global fit - thereby compromising predictions/inferences over the local imbalanced parts of the data.

medium.com/@saikrishna_...

🏛️ Interesting argument on lack of democratic accountability 🏛️

osf.io/preprints/os...

1. It’s not the lack of awareness. But political partisanship is an obstacle to accountability.
2. Lack of social capital. Public support to policies that preserve public goods reduces if there is personal cost

We wrote about gaps in monitoring data and its interpretation of air quality.

urbanemissions.info/wp-content/u...

Stats gaps of representativeness, selection bias etc., creep into these datasets. Media, policy making should consider them before forming opinions.

WQU | Apply for the Applied Data Science Lab Join the Applied Data Science Lab: Access application, online program requirements, course and certificate information, FAQs, and more.

I’m assuming this is Python for Data Science - this is a good free course www.wqu.edu/adsl-apply

Suggestion: Add support for unpenalized logistic regression · Issue #6738 · scikit-learn/scikit-learn LinearRegression provides unpenalized OLS, and SGDClassifier, which supports loss="log", also supports penalty="none". But if you want plain old unpenalized logistic regression, you have to fake it...

This is my yearly reminder that published results of coefficients using logistic regression in Python are likely wrong because they are L2 penalized by default: github.com/scikit-learn...

#rstats #statistics #digitalhumanities

Bayesian probabilistic forecasts using categorical information | Part 1 In this blog, I will make Bayesian forecasts of Ozone concentrations.

If I ask you to forecast pollution in Delhi on some random day, you'd say a number X

If I further ask you to forecast on a Winter Day, you'd top up X with some y

This is Bayes theorem running in your head.

This blog attempts to quantify it

medium.com/@saikrishna_...

#rstats #stats 🤖📈

Bayesian Feminist My friend believes that women are bad employees.

I wrote a book in Telugu where I addressed a few harmful beliefs (astrology, homeo, superstitions, biases) using basic stats concepts.

This blog of mine is one such attempt.

medium.com/@saikrishna_...

In a Gaussian distribution. An uninformative prior need not have any closed bounds.

So MLE would be identical to Bayesian estimate.

But give a closed bound (say mean height is uniformly distributed but within 100-200 cm) — Bayesian posterior estimates will consider this close bound — not MLE

No answers :/

Anyways.

Even without an uninformative prior, we provide a bounded prior in a Bernoulli trial.

P(p) = 1 (when 0<=p<=1)

P(p)=0 (otherwise)

So the bayesian estimates are informed of this bounded nature of p. The MLE doesn’t consider that. Hence the difference.

#rstats #stats 🤖📈

Yes. I mentioned uniform prior. Should have mentioned mean estimate as well. But the question remains! Why mle estimate is different for a Bernoulli trial.

Mle estimate of mean and Bayesian estimate (mean of posterior with gaussian prior) is the same for a normal distribution

Quiz time!

Why is the Maximum Likelihood Estimate (MLE) of probability of success (p) in a Bernoulli trial not equal to the Bayesian Estimate?

p_mle = r/n

p_bayesian = (r+1)/(n+2)
(Assuming uninformative prior)

r = number of successes
n = number of trials

#rstats #stats 🤖📈

So if x^2 is the confounder, adding x to the model may not completely remove the Omitted Variable Bias 😬😬😬

Seems obvious now. But wow, that’s an interesting thing! Makes causal inference more complex

😂😂😂😂😂😂