Week #75
Oh the middle of end of 2025, and the beginning of 2026. This is a wired post. I let myself back, reflected more. I know it was a fast paced year for software development. People are here to prove it. I just laid it out straight that its never been valuable to be a human, a distinct, natural and earnest human.
I must say I completed a non-technical goal of 2025, I have my novel’s first draft in my hands(in my google drive). Its very rough, I want to revise it, I want to remove the fluff, but it is there in its entirety. 33 Chapters, 85K words. 45 days of writing span across 6 months, I did it. I am pumped to revise and write the next novel. The hunger to write has never been higher for me.
Quote of the week
“Grit is never bad, grit with wrong intent, for wrong purpose is definitely bad”
I had a grit to write a story, I will write it. It was a grit, not a goal. I had goals for it for the past 4 years, yet when I made it a grit, here it is, I am on top of it. If a grit was something to think bad of someone, or look down on someone with grudge, it is a bad thing. A wrong grit is something born out of desires, rage, frustration, anger or even jealously, None of it could be bad per see. But if the intention of those desires, rage, jealously is for ill of someone then it could be bad, not if it is born to uplift yourself from the ground, after having thumped by life.
Sometimes, grit is something life give us, throws at you, you need to learn to handle it. It can’t be thought, it needs to experienced. You have to fail, you have to struggle. But here for me, the struggle, the resistance to write finally bent its knee against my grit (not boasting, nor arrogant). Just fierce grit born out of the desire to be a better person, to give back to a human, to say a kind thank you.
Wrote
Feels good to write this posts. I never miss them! I write less post this year. Just 2 or 3, but I have written a lot of SQL like learning log posts (53 of those).
Not to forger I am also writing my thoughts on the things I learn and read here. So 52 articles for each week. Easily making up 100 writing pieces.
I completed Advent of SQL with 15 posts too, here are the remaining, which I completed on the weekend. I learnt a lot, it was a good one, the problems ramping up gradually, then the lore for each post was so good.
Read
technical practices that can dramatically shrink delivery lead times while improving reliability and reducing the cost of change, with or without “AI”
A banger of a post. This is expressing that AI is just a shift in a toolset, or maybe even a abstraction of the language. We will still have ambiguity that a human needs to understand in order to deliver a software.
We can see from his experience, developers were written obsolete from time to time, and each time it was different, more potent than the other, but here we are.
On top of all that, “AI” coding assistants are really nothing like the compilers and code generators of previous cycles. The exact same prompt is very unlikely to produce the exact same computer program. And the code that gets generated is pretty much guaranteed to have issues that a real programmer will need to be able to recognise and address
This is true too, we are just automating and generating crap faster, code is always crap until distilled and refine with each iteration to the needs. We just now have a better or worse iteration cycle, a machine that can spit out code like tirelessly, we need to vet and test it.
2025, The year in LLMs: Simon Willison Weblog
Boy! That is a lot! I have been saying “overwhelming” word was not sufficient to describe this tend of LLMs in 2025. This explains the reason
We had LLAMA falling, Gemini gripping, OpenAI still on the top yet cornered neck to neck suddenly with Chinese Labs and Anthropic in its own league. We saw the sudden rise and sudden dip in vibe coding, people thought “We can be programmers! We don’t need developers anymore, hehe” to “Damm! Do I need a developer to debug this?”. That was a funny thing to watch (as a developer)
The images and 6 second video clip generated by AI are mind boggling, we saw from Sora and Nano Banana what havoc they can wreck if put in untamed hands.
Local models are getting good, but the speed of the cloud and advancement over the other side is rocketing. There is also this trend of cli based agents. Claude code just set the trend and let 100s of cli agents rip off in the months to follow. Those are still released by new companies every now and then.
Slop, yeah! We had less human slop than we needed AI right?
Thanks to Simon Sir for this awesome blog. It finally gives me a relief to read so many thing have happened at a glance
Andrej Karapathy’s 2025 LLM Year in Review
This was more of a reflection post, of how his mental model has changed and how things are building up. I like it. It was a interesting and highly technical perspective.
His opinion of LLMs as Ghost is so liberating, as it actually threatens me from my identity if we compare it with humans. Ghost makes sense, even dismissive it as a slave sort of relation right? Not in a bad way but kind of inferior relation for LLMs with humans.
Agent that lives in the terminal is practical, for a developer or a human who understands what they are doing, they know what they want, its just too much menial for them to spend the energy on. I agree.
There is a lot of work to be done, developers, don’t strap your belts, hone your hammers, its going to be needed.
Also his post:
Could this have been more accurate! Right note to end the year.
Vibe coding last year, now this is the trend we are surfing on, this will last decades.
Local LLMs are how nerds justify a big computer they don’t need
Curiosity gets the better of them. I have a 8GB device, I can barely run a 1B parameter model. I get frustrated but have nothing to complain. I can use ChatGPT in temporory mode, or incognito mode if I don’t want it to attach it to the memory. I don’t see using local models on scale is justifiable just yet.
You are not dumb, you just lack the pre-requisites
Yeah! I have started to learn SQLite and since 2 years made a Brilliant org streak. I feel good taking on advanced concepts soon.
Basic and a good foundation helps you pivot and branch off to wide possibilities.
13 Tips for Writing a technical book
A handy little thing to remind myself, this is inevitable for me. I would write one. Not this year probably. But I would surely write one, my gut, my instinct is not false on this.
I would this then. Great advice for just being curious.
This is a good post to say that we have different ways of using LLMs at this point and nothing is permanent. Every month or weeks, this is changing. Adopting a new workflow is like juggling circus art.
Codex is something I haven’t even touched, Claude code too, never. I have used Amp, Gemini CLI, Warp and Cursor the most.
I love those, those are cheap or even free, they help me understand what I was about to do wrong. They have never produced anything right 100%. I always needed to understand what was I supposed to do.
Is this true “>The important decisions these days are language/ecosystem and dependencies” Maybe but I don’t see that. Its kind of true, but not in a big way. The major things are the flow, the edge cases and the intuition for the problem for it to be ale to understand.
This actually surprised me”> Go wasn’t something I gave even the slightest thought even a few months ago, but eventually I played around and found that agents are really great at writing it, and its simple type system makes linting fast.” I want to try it now. I have ton of go projects.
The internet is a net negative
We’ve maximized information and accidentally drowned wisdom
Hits home. This is good observation and a perfect critique, not over cynical, nor too loathed. Its just helplessness to avoid the battle of the mind and the heart.
The business of the world has forced humans itself into a trap. What an irony we live in, creating a cage for ourselves. Besides slaughtering nature into it too.
All that time, that irreplaceable human attention, fed into machines that convert consciousness into quarterly earnings.
That hurts badly. We are loosing are attention to these machines. We need to get it back. The time, the wisdom and the boring tone to our lives.
The optimist in me is still here. Still hoping.
This is great piece of writing. I love it. Want to write essays like this.
Thanks for writing this Kenneth, you have inspired some spark for me.
What I learned writing Gleam, after coming from Python
Top to down approach. This just shifts from taking the problem and boiling it down to the input and output. Wow! This just made so much sense now.
We can define the main API as the function that takes something and returns something. In between the intermediate steps, we can then decide what each component of the result will come from.
I need to try hard on learning functional programming this year.
Engineering is becoming bee-keeping
I like this comparison quite a lot. Swarming agents is what its happening. And the realisation that code was the thing that doesn’t matter, the thing that matters is did we solve the problem
Honey shows up at the end. That’s what matters.
And bees can sting. Without the right gear and practices, you get hurt. The protective suit, the smoker, the careful movements. In code, that’s patterns, documentation, tests. The guardrails that keep the stings to a minimum.
Working like this is exciting. There’s a playfulness to it. You can try things without committing. You can explore without sunk costs weighing you down. You can work on three features at once because you’re not holding all the context in your head anymore.
This is so subtle, yet perfect. You lead by good or a bad way.
You cannot say I was not the leader when you are the only person building and maintaining it. You lead by examples, good or bad.
Average sucks you know? You are either good or extremely bad. You cannot not lead.
Wow! This post is so perfect, not only fits the manager but also every human, a elder human trying to teach or lead a younger one. The younger one learns from the examples and behavior of the leader.
Git Integrations is ten years away
This is hillariously funny. I can’t imagine VS Code team coping up with git integrations in 2025
I don’t blame them entirely, at least they realize it is missing. With these LLM assisted coding, they decided to ship it finally. We have one instance of AI assisted coding helping VS Code ship faster (after 10 years).
Learn Git, true. I alway 100% of the times use the cli. NO aliases, no agents, just CLI commands. git add, commit -m, push, pull, merge, rebase whatever. If I don’t know, I google it, read the ai overview and straight to the keyboards.
Watched
CMU Databse System #3 Database Storage: Files, Storage, Tuples
This was a good lecture on the different storage hierarchy of the storage. The top there is the pages, the blocks of memory that database fetches for individual records or tuples. Then there is the blocks of memory on the databse file itself, and the actual disk of storage.
CMU Database System #4 Memory Management and Buffer pools
Ok OS is not our friend, we need to manage our memory ourselves. This went wild, I thought managing memory was like shooting yourselves on the foot, but not for DBs.
So we load the database file, from the disk into memory not as full, but chunks called frames, where each page is contained in the buffer pool. Interesting, this is done in the actual ram or the memory not full at once.
So this makes it the different algorithms to decide which frames/pages to keep and evict (remove)
There is a difference in lock and latches, a lock is something that protect the database logical content from other transactions i.e. the data to write or avoid corrupted reading
However a latch is something that helps in preventing the database internals from other operations, its only for an operation not a query. Its like a mutex.
We can’t rely on OS, as OS doesn’t know what are we querying.
There are like half a dozen implementation of replacement caches like LRU, Clock, LFU, LRU-K, ARC, etc.
How I parsed billions of rows for every user in 2 seconds
Wow! I like these videos. I learnt a lot too. It was passive knowledge true. But I came to know that these things are at least possible.
Clickhouse as a database, the queries, we can use Material views which can be used as a CTE almost but on the fly, Endpoints to query them as a URL. WOW!
The optimisation was based on the clickhouse features only, not sure if it would have been possible without it.
Learnt
SQL Recursive CTEs
We can define a recursive CTE by referencing the CTE within it
We have a single row (could be multiple as well) as the base case
Then we define the recursive part, by referencing the cte as the table we are fetching the records from with the data queried to it as the parameter.
Used it to solve Day 14
FTS in SQLite (Full text search)
I learnt how to write a query for FTS and construct like a index for searching across tables.
This is efficient from the string comparison as we don’t have to define how to look it up, we just define what we want. The algorithm and the query planner does it efficiently for us without storing it separately on disk.
Golang might just be better than python for writing LLM generated code
Its simple has a type system. So it makes it easier for LLMs to generate valid code with correct checks in place.
I need to experiment it with to understand the nuance this has.
Tech News
Thanks for reading Techstructive Weekly! Subscribe for free to receive new posts and support my work.
Well! That was a thumping start to the year 2026. I have bright ideas and a new canvas to paint. Looking forward to have things running and working in my favor over this year. After a slog and slump for 2 years, its time for me for redemption. I can see a hope, hopefully you can too. If not, you will soon.
Happy New Year!
Happy Coding :)