Tag: ai

Weekly Notes 2026/37

  • 🏃️I’ve completed 700 km of running. More than 3 months are left. I should be able to finish my 1000 km run this year without trying extra hard in Nov and Dec.
  • I’ll be starting a new job from Oct this year. My role will be focused on (Rust) tooling around infrastructure management for self-hosted, air-gapped, or hybrid infra. Think of software supply chain management and lifecycle management of apps. Pretty much like docker/Kubernetes on one hand; and ansible/puppet on the other.
  • A friend got excited about newly opened something called Brick Oven in our neighborhood. We went there for dinner earlier this week. I found the place to be OK. Among such places, I still prefer Onesta. Usually, I prefer to go to traditional south Indian eateries with self-serving. Those are most efficient eateries and food quality is very good. But if you are catching up with someone, I guess places like Brick Oven, Onesta, CCD are fine.
  • Lately, I’ve not been feeling like visiting HN. I am spending more time on lobste.rs and RSS feed. I also discovered something called Webring (https://webring.xxiivv.com/) — interesting idea but I won’t see myself browsing them just yet.
  • I feel more energetic when I write code by myself. In retrospect I now understand why. Because there are occasional dopamine rush when I discover a better pattern and whole codebase is now slightly better and when I learn something on the job. This works because getting better and learning matters to me emotionally. I am almost certain that a long-term LLM usage will burn me out in a year. I’ve been seeing signs.
    • Another solution to this problem is that I use 30% of my time coding using LLM and 70% of time solving problems and designing systems by hand. This will also work for me. What is not going to work in long run is being a developer and only using LLM.

Weekly Notes 2026/36

  • 🌦️As of today, the southern peninsula is very rain deficit. Interesting summer ahead of us. Some twenty five years ago, when I went to Chennai for college, I was pleasantly surprised to learn that rain-water harvesting was mandatory for every building. Why it is so hard for such good-practices to become popular?
Bengaluru is extremely rain deficient this year. The whole southern peninsula is ~30% rain deficit.
  • The generative AI is on my mind for last couple of weeks but I don’t want to write to think about it too much. The genie is out of the bottle and we have to deal with it somehow.
    • Most people think of throughput when they think of AI. IMO, throughput is not a value-add of AI though the limelight is on the throughput. It doesn’t increase throughput without costing you in long-term. I think the real advantage of AI is getting to a higher altitude easily. That also means you should know the questions you need answers to. AI can’t ask good questions. Also see “Do you sill read the code“.
  • It is now pretty obvious that many (perhaps most) don’t see tech (at least the IT) as inherently good. We have squandered most of the goodwill that our predecessors accumulated.
    • Software companies needs to see FOSS and community engagement as long-term investment. Unfortunately, not many companies can think beyond a few quarters.
  • I spend some time on my old FOSS work, without using any AI. I felt a little better. The temptation to use it very strong to use it on boring part but when you touch it, you don’t usually stop there!
  • Creatine supplements have positive impact on my fatigue levels.
  • I love dua: I use it to analyze disk usage, and recently they’ve added a feature to clean up temporary files easily. Oh joy!

How good are LLM agents? Depends on whom you ask!

Depending on whom you ask, you get very different answers.

Depending on whom you ask, you get very different answers. I’ve been noticing a pattern.

If you ask a generalist who is B+/A- on any job you throw at them, they are likely to find LLMs very useful.

If you ask a specialist who has spent a lot of time doing a few things and has deep expertise in them, they are more likely to call LLMs garbage.

I think both of them are right.

My own experience with LLMs is that they are very good at doing “average”. Being a generalist, I only know a few things very well (learnt during my Ph.D. and startup), and I can see LLM sloppiness in those areas. But for most other things, I’ve no way to evaluate LLM output—a plausible answer looks as good as a correct one.

I totally understand when Zig maintainers or database developers ban the use of LLMs. I also understand when a good application developer, who understands some part of her system very well, relies heavily on LLMs for other important jobs like adding a demo site, integration, or a front-end in a language she doesn’t know.

What about enterprise and large organizations? If their products are technically “average” like a social media client, then I see some advantage of using coding agents (though I’m not sure if long-term costs are worth it). If their product has critical components that require very specialized knowledge — such as compilers, database engines, or encryption algorithms — I’d be running away from them if I get to know they are trading their hard-earned tribal knowledge for short-term, LLM-induced gains.

Weekly Notes 2026/29

This week, Digit Insurance pulled a “scam” in me. It almost succeeded! They sent the following WhatsApp message.

I bought this car from Spinny almost a year ago. Under the impression that Digit is my car insurer, I went ahead and renewed the policy pending KYC and pre-inspection video. KYC was easy to do using PAN but I could not login in their app (that’s a different story). Frustrated, I cancelled the policy, refund pending.

So where is the scam? First, after wasting an hour on their app, when I double checked, my insurance is not expiring on Aug 3, not in next 5 days as they mentioned. And the insurer is United Insurance company, not Digit! Should’ve known better since they used the fear “fine of 2000” etc in the message to nudge me click!

I now understand that they advertise heavily in IPL and a famous cricketer is their scammer-in-chief or brand ambassador or whatever. Enjoy your cut, motherfucker!

I was “saved” by AI slop (refund is still not in my bank though). Their app didn’t let me login for inspection because my phone number starts with 91 and it confused it with ISD code (+91 part) and complained that mobile number is not complete but also won’t let me type more than 10 digits 🥁️🤣️. I had similar experience with these websites/apps when I was purchasing life-insurance. I ended up buying from LIC because their PHP website still works. Perhaps I should start using the govt. entities, they don’t seem to change their color-scheme and look & feel every other week and they sites seems to work better than most fancy looking websites I’ve lately used.

  • I had some recognition at work by my peers 🥹️. Its nice to be appreciated by your peers you appreciate. There is nothing more to the “career” IMO.
  • At work, there is a hard limit on AI tokens. I burned my limited monthly quota in a week or two and then went through withdrawal symptoms 🥹️. I’ve started writing some code by hand now. Most of the time, I do not ping anyone when I have a doubt about code-base. LLM works great because a lot of people invested time and effort in writing docs and great MR reviews.
  • The expectations from AI is a bit more saner these days but I don’t think the AI psychosis is going away anytime soon. We had an internal hackathon where AI was the judge! So did Kaggle last week.
    • I am yet to see AI increasing revenue per head or keeping folks focused enough to finish it off. There are some productivity gains here and there. More people are now talking about it.
  • I learnt about this ecological event in Mizoram last week. Every few decades mass blooming in their forests causes a rodent boom, and devastation to crops.
  • I filed my ITR for this year. I think finally I know the process. I created a Notion template to use next year. I also filed my GST returns. The GST website is not bad, it is not as good as ITR website though.

Weekly Notes 2026/26

  • I got my 3D printer back from my cousin. He stopped using it for a few years. He sent it back to me when he had to move. I’ve been printing a lot of useless stuff, like the one below.
A 3d printed dinosaur stealing my Hajmola
Medieval India 1, edited by Irfan Habib
  • I have been reading “Medieval India 1” (edited by Irfan Habib). I found a few papers in it quite interesting — especially on social mobility and travelers accounts of Hindu religious practices at many famous places; many of them don’t exist anymore. Perhaps I will write a review later. Mr. Habib has an excellent, readable body of work on Indian history. Indian history was a very emotional subject when I was in college. Mere mention of a historian was likely to invite accusations of partisanship. I’ve mostly read what people would call “leftist” historians. Mr. Habib, D. D. Kosambhi, etc., were/are my favorites. Though I don’t touch Romilla Thapar anymore. I don’t like her as a person anymore after her stint at EPW.
  • The economy and the reality are catching up with the AI usage at work. Costs have been rising, and token-maxxing by a few is burning through the monthly budgets in a week. Given its effect on memory prices, I am a bit glad that I bought a Linux machine a few years back, though I didn’t need any. It is quite a weird situation; the promise of AI is replacing labor and jobs, then who’s going to buy it, and what is its point other than being a tool of economic destruction?
  • Congrats to Ireland and New Zealand for their wins over India and England. I didn’t watch any of those matches.
  • The government pulled a statistical sleight of hand to show the rural wage has increased by 13%! This data will go into official statistics soon.
  • My favorite podcast, In Our Time, continues after its host retired from the show :-). I’ve listened to a few new episodes. It will take some time to get used to the absence of Mr. Melvyn Bragg. Content produced by the BBC is still one of the best in the world, but this podcast had been out of this world.
  • Work was very slow this week. I also took a few technical interviews for a client who has been using AI to screen resumes. I am not a fan of using AI in the interview process, but it is what it is. Earlier, I didn’t like structured interviews, but now I prefer them since comparison across becomes a bit easier. I don’t have anything against unstructured interviews if you practice beforehand and make sure it doesn’t derail candidates. Now I wish the candidate also practiced interviews, especially how they answer the questions; please get to the point as soon as possible.
  • Something is perhaps churning in the PHP community: RFC for geneic was rejected! And I saw a talk about the RFC process. I think it is usually a VERY GOOD idea to make the lives of programmers easier.

Weekly Notes 2026/25

  • 📚️I ordered a few of Prof. Andre Beteille books. He dies a few months back. He is perhaps the only sociologist that I could read. MN Srinivas’s “A remembered village” was also a great read.
    • I also bought a few books on software engineering in general, especially on refactoring, design patterns, and architecture. Books on software engineering becomes outdated pretty soon. I prefer blog posts or papers.
  • 🥭️ The peak mango season in Bengaluru is over!
  • 🏏️I think I’ve lost all interest in cricket. I can no longer watch a full game, even watching highlights takes some effort. While the heart misses the test matches but I don’t have that much of time or attention. Then the T20, the “pornography of cricket”, fun as it is, has already killed neural circuitry that appreciated delayed-gratification required for test cricket. Good riddance, perhaps!
  • 🤖️I learnt that Hyundai bought Boston Dynamics a few years back (80% stake) and now finished the deal by buying the remaining stakes. Automotive industry is betting on “physical AI”. Toyota, where I currently work, also have significant focus on physical AI.
  • 🛵️ It was a WFO week at work. I think I should once a week to office just to ensure that bike is in good shape. My office bag content remain relevant to seasons. I got caught in rain and I didn’t have my rain jacket in my bag.
  • This year is going to be most tumultuous as far as “coding” is concerned. Either we’ll figure it out how to use AI properly. Or I’d have gone through the stage of griefs and reached acceptance by then. I’ve been thinking about AI or pushing back on it till December this year.

Weekly Notes 2026/16

  • Very dry week. No sign of rain. The temperature is high and there is no forecast of rain either. I am hoping for next week to be bit cooler.
  • I gave a talk at a student club at APU.
  • I am joining Toyota Connected India next week. Pretty excited about writing Rust in safety critical systems.
  • I am getting a better grip on AI tools. I am still conflicted if I should continue to use them or go back to good old ways. Lets wait for a few more weeks.

Reading List

Weekly Notes 2026/15

  • First two weeks of April are usually the hottest weeks in Bangalore. Its less dry than the 2025 but only one light shower. The mangoes outside my house are growing nicely though.
  • Its been three weeks since I’ve written code by hand (except for a few tweaks). I am using Claude and sometimes qwen cli. Claude is definitely much better but qwen is quite capable at well scoped tasks.
  • I don’t think AI is replacing programmers anytime soon but sure making me dumber. It is becoming harder to focus, read docs and engage with a linear conversation. Also when I ran out of the credits, I had to push myself to open my editor and look at the code and I found it alarming.
    • My TLDR summary of AI is the following. It generates pig-iron and call it steel. You can build a fun Choo-Choo train with it over the weekend but please don’t build a proper train with it that carries hundreds of people. But can you resist?

Reading list

Weekly Notes 2026/14

  • For last three weeks, I am doing vibe coding for almost everything. It is helping me understand its power and limitation. I feel like everyone is right about AI. This reminds me a saying about India, “Whatever you can rightly say about India, its opposite is equally true.”.
  • Ookie is at her mama place for last two weeks. She will be back this Sunday. She is having very good time in Toranagallu.
  • I vibe coded an app https://github.com/dilawar/trim-dead-area to trim dead areas from a video (auto-crop to most interesting parts). It started off very fast but tweaking is extremely hard with mixed results. Learning about ffmpeg was a joy using Claude. More here https://dilawar.in/2026/vibe-coding-a-gui-to-auto-crop-dead-zones-from-a-video/.

Using Claude in a small C++ library

Long back, I ported LSODA solver to C++ (https://github.com/dilawar/libsoda-cxx) for a neural simulator I was working on during my PhD. This port is almost the same as the reference C code except for STL containers and some other goodies (and some OB1 errors).

Today, I used Claude to improve unit tests and docs and enhance CMake integration.

First, I asked Claude to look in scipy codebase, search for LSODA related tests, and add them to my project. It did a great job and wrote a better commit message than I would have done but that’s not saying much! The tests Claude wrote used assert which is a noop in release build i.e. these assert won’t be there in release build so passing tests in release are meaningless. When I pointed this out, it replaced them with runtime exception that works with both debug and release build. Not cool, Claude but not bad either. Its C++!

Then I ask for a CMake harness so that folks can integrate my library into their CMake based workflow easily. I did OK at first attempt and made a boo-boo which was caught by Coderabbit review bot (free for open source). Claude fixed it later!

All of it took roughly 30 minutes!

If you are very very familiar with existing codebase then I don’t think Claude/AI tools add too much value. With hot cache, I could have done it in one hour. The advantages of 30 minutes of time saving may not be worth losing your touch with something you enjoy i.e. programming and story building in your head. I definitely feel that I am on “opium” when I use it for a long time.

Claude is very very good at things at which I am below average. Which is true for almost all the things I am doing with Claude thee days. So it is not unnatural that these tools enjoy great reputations among users. More objectively, good AI tools are perfect for things that feels like chores e.g. “translate tests from this Python implementation to C++” or “reproduce this bug and when you find it commit it on a branch” or “create a new project with this and that”.

At my last stint at a medium size org that lasted a month, we are asked to use Claude code as much as possible. Sure, I finished a task in half an hour that would have taken a person familiar with the codebase 3 hours but the PR is still pending review after 12 days! So where are the gains? Beware of Amdahl law as well!

As many have pointed out before that AI tools are very eager to achieve success. They will try to achieve it by any means. Once in a while they will stop and ask you but most of time, they will find a plausible solution. Ensure that your prompt has the definition of “success/done” and “properly” articulated. Discuss it with all stakeholders when a ticket is created otherwise you’ll get something that kind of work but it does not.

Weekly Notes 2026/13

Growing mangoes in my neighborhood
  • The mangoes outside my home are growing bigger nicely. It’s finally nice to see fruits grow in the neighborhood.
  • It was a tough month at work. After thinking hard about my team work culture, my interaction with my manager, and his un-willingness to find me a new team (unless I am his relative), and consulting two of my manager friends, I decided to submit resignation.
    • I had a narrow escape at Bellary road while coming back from office. That also significantly contributed to the decision.
    • They are also moving office to a new location and that uncertainty also contributed to the decision.
  • I had another interview just after I submitted my resignation. You should not do both things at the same day, please! It went horribly wrong.
  • One Microsoft interview was in pipeline. I failed the fourth round (I guess). I am not keep on giving more interviews this year.
  • I am planning to do independent consultant work. I have already have two clients and it should be sufficient to get started. Lets see how it evolves.

Reading list

Weekly Notes 2026/12

  • All rounds of Toyota interviews are finally over. The HR discussion lasted 90 minutes. I enjoyed most of it.🤞
    • And they rejected! Apparently I am over qualified. Didn’t they know from the resume? The reason is perhaps something else — a wrong pause between sentences! I was under the impression that things are going very well.
  • Early this week, a light shower. The leaves of the mango tree outside my house are cleaned! It wasn’t strong enough to clean the dust from the road and wash my parked car. Two days later, we had a good enough rain that washed my car. The mangoes on the tree outside my home are big enough to pickle :-). This year March is not as hot as the last one. Already a few showers.
  • Due to LPG shortage, the restaurant near my workplace is serving a few items, mostly dosa, tea, coffee! At least they are open. But only a fraction of people are working. Everyone pays for the war expect for people who started it?
  • At work, use of AI tools are now compulsory. I am learning to use them, begrudgingly!
    • I already have a big win. I was given .NET6 codebase with empty readme file — a language I never wrote. I asked Claude to act as a staff engineer/product manager and walk me though the codebase and quiz me later. The purpose of walk-through: I can contribute to the codebase. It did a much much better job than a co-worker would have done. It saved me so many pings and emails.
  • I learnt more about AI tooling yesterday from a talk by a coworker. This blog post was also pretty helpful explaining how context window size impacts the quality of response.
    • When the context is filled, the quality of response suffers. The context size is typically 20,000 to 1,000,000. An archive paper is roughly 10,000 tokens. So its not very large!
    • Either you drop some of the context (loose information) or do compaction (summarize the context) and loose nuance. While compacting, pinning important information helps. Perhaps claude is already doing all that. For a general purpose chatbot, I see why this could be a hard problem to solve. A single word may have the most nuance.
  • In other news, uv, ruff and ty joins OpenAI. These tools are created by Astral. Many AI companies have been buying tooling (especially cli tooling).