A Simplified Model of “Work”

1. Develop an idea of what you should do.
2. Execute the thing you should do.
   • Some results might be expected.
   • Some results may come as a surprise.
3. Recalibrate and go back to Step 1.

A few tongue twisters

pinged ping ponger ping ping pinged pinged ping ponger ping ping

(The below was made a few years ago, with the help of some of my friends) 谁在水上睡觉时付税？

猫统治!

有一只猫。他叫王。王真喜欢喝牛奶。王的主人是一个软件工程师。

小王不知道怎么用计算机。他嫉妒她的主人。

小王在它主人的架子看到了一本书：《Intro to Python》。

小王偷偷地知道怎么读英文。他花一些月度这本书。然后他的主人给他买了一个电脑。

小王创造了一些很简单的程序。一个程序打印“hello world”。

然后，小王在coursera上一节课：”introduction to machine learning”。他上一些可之后创造了一个移动应用：”cats rule”。每个猫下载了这个应用。然后所有种类的猫宗旨了世界，也人类沦为了猫的奴隶。

结束

A brief post. This just includes some principles that I believe I should live by. Expect this post to get changed over time

• Don't get sidetracked by time-consuming things.
• Don't waste too much time working on things that are unachievable.
• Zoom out and consider the metagame. Ask myself what I should be doing, and ask myself what I shouldn't be doing.

Disclaimer: These are my opinions and thoughts, and shouldn’t be taken as ground truth facts. Much of what I’m writing is based on ideas I’ve thought of, or things I’ve discussed or heard from other people. Also my thoughts may change over time, depending on factors like future capabilities of AI.

Intro

Over the last few years, we’ve all seen AI change. ChatGPT got popular in the end of 2022, and every 6 months since then, some drastic change has occurred.

Several decades ago, even something like chess was dominated by humans. But this changed in the 1990s when IBM’s engine DeepBlue was able to beat grandmasters like Gary Kasparov. Now OpenAI has models that has placed 6th in the IOI unofficially (the four players it lost to were all 3000+ rated in Codeforces, which is roughly equivalent to being top 50 in the world).

Out of the many things that AI has gotten good at, I, as well as many others, had expected the last thing for it to be good at to be math / computing olympiads. Many of the problems in these contests (at least the higher quality ones) are quite original, and require creativity in coming up with solutions. AI models are largely based on training data, so it’s surprising that they should be able to solve non-standard problems.

A bit of background about myself (which you can also find clicking through my website). I used to do some regional level math contests in high school, as well as other contests including AMC, AIME and ARML. I wasn’t a top competitor, but I do find my involvement in these contests to be responsible for my interest in mathematics – at the time, I liked combinatorics problems (and still find them fun to work in). Starting from college, I’ve been highly involved in competitive programming, and am a Master in Codeforces (my rating is usually around 2100-2200, which is around the top 1 percent of Codeforces users), and have been doing contests on and off. I was also an ICPC World Finalist from Georgia Tech a couple of years ago. Nowadays, I spend my time doing grad school work, and also teach competitive programming on the side.

I remember in early 2022 hearing about DeepMind’s model AlphaCode, which could apparently get ~1300 rating on Codeforces. Getting to this level is nontrivial for a human, and often requires prior problem-solving experience: for example through math contests, or through a couple of months of practicing competitive programming. This was before ChatGPT came out, so it was new news. Then, in late 2022, ChatGPT came out. I didn’t play around with it much until 2023, where I found it could help me significantly with my internship work (I was working at a tech startup during the time). It did dawn to me that stackoverflow was becoming less useful, and I could learn things through ChatGPT. ChatGPT was still in somewhat primitive stages back then though and did hallucinate a lot. I remember testing it with competitive programming problems at the time – It could get some test cases correct in USACO Bronze problems and could spit out implementations for standard data structures such as segment trees and link-cut trees. At the time, I saw ChatGPT as a useful way to retrieve facts about things that are well-known.

The end of 2024 was when my mind was blown – OpenAI had announced a model o1 that was ~1800 rated on Codeforces. I was still somewhat skeptical since I hadn’t tried competing with it or seen it in action. But I was convinced that AI was better than me after o3 was released in the Spring 2025. I did use it to prove a lemma in one of my papers, and I definitely found it to be a useful tool. Very recently, GPT-5-pro has been shown to be able to discover new mathematics (see this X post). I guess I wasn’t too surprised at this point, after having seen o3 be a useful tool in research. But this suggests AI is still getting better. What will the world look like when it’s better than every human at every (technical) task?

What role do contests play today?

This brings the following question: what role do math / programming contests play today? AI can solve everything, right? What’s the point of trying to do something that’s already solved by AI? Why can’t we just use AI for everything, since that’s anyways what we’ll do in our careers?

Don’t get me wrong, I think it’s important to be able to use AI. But it’s also important to be able to think critically about problems and evaluate information that is given to you. Critical thinking is still a needed skill. In life, you’ll hear opinions and perspectives from others, and need to have some way to form your own opinion after evaluating what you observe.

Will AI devalue competitive programming achievements?

From my experience doing contests, it feels very rewarding when you achieve something in competitive programming, such as an increase in Codeforces rating. I’ve gotten +150 rating a couple of times, and I can confirm the dopamine rush is great.

I’m not sure how things will be if it’s possible for anyone to get 2100 rating on Codeforces by just plugging the problem statement into AI. This part is a bit sad to think about. Some companies (including many tech startups and quant companies) have tried to hire top talent. Many of them still hire via metrics like Codeforces ratings and awards at contests like IOI / ICPC. I wonder if the increase in AI’s ability will devalue hard-earned competitive programming achievements.

Will programming contests feel “outdated” in the future?

In the past, programming contests gave people valuable skills for programming. This includes things like finding edge cases and debugging speed. In some sense the results were a bit tangible. There are still benefits to gain from contests, even after AI is better than us. The “critical thinking” skill is still to be gained from programming contests; however, the notion is a bit vague, with no concrete results. Maybe someone who wants to do well in a math contest may gain benefit from it (because of the similarity of problem style, especially for combinatorics problems), but people who want to get into tech may feel less and less inclined to do them, just because it’s getting farther and farther from what “real-world programming” is.

Perhaps coding will later not be a needed skill, but “critical thinking” could still be a needed skill. A bit of a wild idea, but if “critical thinking abilities” is the main goal, this may result in less of a shift towards learning coding, and more of a shift to learning things that directly foster “critical thinking” such as math contests. The last few years has seen a shift in focus from math contests to programming contests, but maybe we’ll shift back to math contests. Again, this is just some very wild speculation, but it does seem like a possibility to me.

Other comments: Contests for evaluating your own skills

Even if all our jobs will involve AI, people will still want to be able to evaluate themselves. Doing a difficult task where AI isn’t allowed, such as a contest, gives an objective metric about your individual skills (it may debatable whether this is a good metric, but it’s nevertheless an objective metric). You can have some kind of ownership over your achievements. You achieved not because you were carried by AI – you achieved because of your own hard work and good performance.

Again, I definitely do think people should learn to use AI. It makes you more productive. But there should also be some kind of independence where you can make your own decisions without depending on AI.

My thoughts on rules for programming contests

For a long time, I’ve had a point of confusion for what the rules should look like in programming contests. Atcoder has a fairly lenient AI policy (As of August 2025 when this blog post is being written). You can use autocomplete, and you can translate problem statements. Meanwhile some contests like USACO strictly prohibit any use of AI or autocomplete.

As AI gets stronger (maybe we’re already at this stage), autocomplete should be able to infer what the problem statement is asking when looking at the code, and fix the code. It does seem necessary that there needs to be a strict ban on using AI. Maybe AI for looking up syntax will be justified, but it’s not clear what justifications there can be for other kinds of AI.

There are many people who want to do programming contests without any unfair help. But there are also others who decide to cheat, and AI has made it much easier for them to cheat. In the past, cheating often involved some form of collaboration during the ongoing contest. In some sense, cheating was hard for lower rated participants since it’s nearly impossible to get a high-rated participant to share their answers during an ongoing contest. But now it’s simple – just ask ChatGPT, which has comparable skill to a top 0.5% level competitive programmer. Maybe some form of proctoring is needed to prevent cheating? Also the Universal Cup has started requiring contestants to record their screen when competing – maybe this is the new norm?

Conclusion

The world as I see it has changed a lot, and the role of things like programming contests is getting less clear. I’ve written this blog to lay out these concerns, and I’m curious to see what happens in the next few years.

I am back after some time. I have been working in a startup called PerfectRec, and the work there has been ML-heavy. I am their “expert” on learn-to-rank, and learning / reading up on literature has motivated me to get back to blogging. Anyways, let’s get started!

Much of the below content is taken from my recent readings from this Wikipedia page and from [1], and it additionally includes some of my speculations / thoughts. But I’m thinking this post will be on the lighter / simpler end of what I’m planning to write for future posts. It doesn’t go deep into the theory, but I felt the ideas still might be worth sharing!

Linear Regression: Problem Statement

The least-squares linear regression can formally be stated as follows:

(one can assume here that \(b, \beta^T \in \mathbb{R}^d, x \in \mathbb{R}^n\)), i.e. we want to find a \(\beta\) such that the least squares error for \(\beta^Tx + \beta_0 = b\) is minimal.

There tend to be issues with standard linear regression, where the function may not be the “best fit”. Especially in higher dimensions, the model tends to find functions with overly complicated strucuture, i.e. there exist functions with more zero parameters that fit the data appropriately. We ideally want a function that 1) has a simple and 2) is interpretable, i.e. one can determine which parameters are actually relevant. Below, we will introduce regularization, which serves to resolve these issues.

A Regularization example: LASSO Regression

The idea of Lasso (least absolute shrinkage and selection operator) is to solve the above optimization problem, but with an additional constraint:

We can rewrite the optimization problem as

The last term can be replaced with a \(2\)-norm (called ridge regression) or a \(0\)-norm (called best subset selection).

\(\lambda\) is often referred to as the regularization parameter. What is the role of \(\lambda\)? One explanation that is often used (although maybe quite non-rigorous?) is that when \(\lambda\) is large, there is more penalty for making the entries of \(\beta\) large, forcing the classifiers to be simpler.

We can also discuss the following to see what happens during the regularization (in the case of Lasso).

Consider the constraint \(\sum_{i=1}^d \beta_i \leq t\). The region for feasible \(\beta\) can be seen in the shaded region in the image below (image source from [1]):

\(\hat{\beta}\) is the least-squares optimal solution, and each ring is the set of points where the loss function is equal to a certain value (note that the contours have this structure since the least-squares loss is convex). We want to minimize the loss while being within the boundary, and this will in general happen when we are at a “corner” of the boundary; i.e. the solution that is given will have a sparse structure. Having few nonzeros is advantageous in that 1) it is a simple model and will thus be less likely to overfit, and 2) the model is interpretable since we can better determine which variables (i.e. the nonzeros) have a contribution to the problem we have at hand.

Other forms of regularization

There is also \(L_2\) (ridge regression), where the constraint is an upper bound on the two-norm of \(\beta\), rather than the \(1\)-norm. One can visualize the diagram from above, except the dark region is a Euclidean ball rather than an \(L_1\) ball. The solution will not necessarily have a sparse structure anymore, but many of the coordinates will tend to be similar to each other.

There is also subset selection, which has an \(L_0\) constraint. This is generally performed by setting coordinates below a threshold to zero.

References

[1] Regression Shrinkage and Selection via the Lasso. Robert Tibshirani. Paper link

[2] Wikipedia page for Lasso Regression

[3] Wikipedia page for Ridge Regression

Some disclaimers: I only just came up with this idea recently on Sunday (Nov 7 2021), and I’d say it’s not well-tested. Also, this idea probably fails if you’re trying to invest the core of your time on more than 1 project, but I’d say it’s worth trying if your goal is just to focus on one project.

For some context, when I did research in my undergrad, I didn’t really have a structured way of putting in work. I would generally have more incentive to work on it the day before my research meeting, and the process would end up being rushed and rather stressful. I think this has also been the case for many other projects that I have worked on.

So here is my new idea:

Have one single task that you want to focus on (projects, research, etc…). Every half an hour of the time that you’re not doing essential activities (i.e. this doesn’t apply to situations like classes, meetings, and sleep), spend the last 10 minutes working on it. Write some key takeaway or ideas or questions in a notebook and mark that entry of the notebook with a number (so if it’s your 5th 10-minute block, you can number that entry with “5” or something).

Doing this for 20 blocks per day (or even less if you’re looking for a saner work-life balance) will result in working on that task for 100 minutes per day — over 23 hours per week. And this is not just “work” time. This is several 10-minute spurts of productive activity, so this is essentially 23 productive hours per week.

On a tangent, this kind of reminds me of this writeup (written by an American mathematician Sara Billey), about how to do a mathematics PhD in a timely fashion; this mentions trying to reach a 20-hour target of productive research time per week. I believe it’s worth a read if you’re curious about it.

I started trying this 10-minute block thing on Sunday. I have to write a research report for my Machine Learning Theory class, where I am supposed to look at some state of the art research, describe it, and explain things like open problems, shortcomings in existing work, etc… (this is open-ended and can also include new results; not sure if I’ll be able to get new results though by the end of this semester :p ). I am looking into Statistical Learning Theory, which is a field that looks into proving lower bounds for computational complexity of learning problems using oracles.

I don’t think I’ll be able to reach 20 hours this week :( , partially since I’ve been extremely busy for the last two days with other assignments and activities (I had a 5-hour ICPC practice yesterday!), but it’s Wednesday now, and I managed to put in 4 hours of effort. That may or may not sound like a lot, but keep in mind that these are 4 productive hours of work. Imagine working for a 4-hour block (i.e. 12-4 pm) and being extremely productive the entire time (I can’t relate to this kind of thing, and I’d imagine it’s also uncommon for others to be able to be productive continuously for 4 hours).

You can see my last few entries here (related to reading this paper)

Maybe 10 minutes is too little for a block? I feel like I’ve been able to boost my productivity significantly, but maybe it’s not great if you want to do deep work? (i.e. work on something open-ended, which might require you to think about it continuously for an hour). I’m not sure about this, but would be happy to discuss it. I personally feel that having a 10 minute block will make me less prone to temptations such as checking my phone, emails, etc… and makes it much easier to put in all of my focus on a task.

I’ve heard of something called the Pomodoro technique ( read this if you’re not familiar), and I remember trying it in high school. I personally find it too tiring, especially since 25 minutes feels like a long time for me to work on a task without getting distracted. But maybe it works for others.

I haven’t fully decided what to put in this blog, but maybe will include a combination about reflections about life and some topics in ML theory / combinatorial optimization. Also, I might add something competitive programming themed (The regional competition for ICPC is this March, and I have been spending >10 hours per week doing training for that!). Either way, stay tuned for more posts!

Hello! I am Arvind, a student at Georgia Tech. I have decided to start a blog.

What is this the purpose of this blog?

I’ve been studying research areas in theoretical computer science (related to graphs, combinatorial optimization, and machine learning). I’d like this blog to be a resource for others to learn tidbits about research that has been done in these topics, and I plan to write about recent things that I have learned / worked on.

Does \(\LaTeX\) work? (I’m new to Jekyll so I’m in the process of figuring things out :) )

a

b

c