Techstructive Weekly #43
5 new LLM models released, devstral, gemini 2.5, Claude 4, Jules and what not, things are crazy in the AI World. Finding a place to un-overwhelm wit this edition of the week from 18th to 24th May 2025
Week #43
What a whirlwind of a week. This was expected, nothing new. New week, new models, every day feels closer to losing the developer job. But everytime, AI sucks and human grug brain works.
This week, I took a step back and tried to ease myself out from this AI slop. Want to really plan the things out and start the grind again. Need to re-align my core values and get back to the consistent and curious flaming torch in guiding myself in the AI abyss.
LLMs and AI Labs are out of control, they are releasing new models left and right.
Number of LLM models released this week: 5
Gemini 2.5 Flash Pro (improved)
Gemma 3n
Devstral
Vercel’s v0-1.0-md model
Claude 4 family (Sonnet and Opus)
This is a new section. At this rate, I am overwhelmed trying to keep track of which model is released at what time, since new models keep popping up constantly.
I spent a lot of time reading and understanding about Agents, and will be creating some form of content around it. For my understanding, as well as creating a resource for this AI hype, to get my hands dirty.
Some personal news, I migrated my blog from markata to my own SSG from scratch, named Burrow. Bold move! I don’t care, no one’s reading my post, so it doesn’t matter, I kept the links intact. A win for me.
Built from bare-bones golang
Only 3rd party library (goldmark) used for markdown parsing (thought of wiring own markdown parser some day to remove this :)
Github SDK for some automation and LibSQL driver for database syncing
Markata by Waylon Walker
Nothing wrong with this SSG, it was fantastic to write our own plugins, and most of the plugins were built-in. But I wanted to explore more about Golang. And this was the move I took.
The current burrow-ssg in golang is inspired by markata
Quote of the week
"Stay hungry, stay foolish."
— Steve Jobs
I listened to the speech of Steve Jobs, and there were three lessons that he tried to convey:
Follow your heart, let the dot connects
Keep looking, don’t settle
Do it as if you are living the last day of your life
This just are valuable life lessons that he might have spend 50 years on. These are given by wisdom and not knowledge.
Stay hungry, bring the curiosity, and keep fighting and keep throwing your hands and legs at it, you will be at the top one day.
Wrote
Thoughts
Nothing significant, because I have been in the cave, cooking stuff. This weekend it will be something. Pushing hard, coding, not vibe, just bug smashing.
Read
This is silly; the more detail you put, the more detailed your tests have to be
The more tests there are, the probability of all of them passing would start decreasing, and hence, it could cause subtle bugs
So I think it is a matter of convincing yourself to write more tests and documentation
Nothing makes me want to hire someone less than them showing an open-to-work badge
This is a good take, if someone has put an open-to-work badge, it might mean that the person is giving indication to his/her openness to work.
We can’t judge by that, and should not, as the person might be in a weird situation and can’t openly express their situation so the badge gives them a slight help in communicating without actually speaking about it.
Revenge of the junior developer
Ok, I spent half an hour reading this article and felt a mix of opinions
On one side, I am happy that I am a junior, I am dabbling with AI agents and tab tab things, but it’s kind of weird to use them.
I don’t like the review part, it feels like a daunting task
I really don’t like ChatGPT’s new memory feature change
I haven’t used this, but I can understand the feeling of this
This looks like a feature without a button, a default feature. Really, are these big companies talking about AGI and care about the values, and slopping users with a bare minimum experience?
In short, you should not use AI slop, pretty obvious
Glauber Costa on Technical Blogging
These are golden points here
Just write it down
Showcasing benchmarks is a bit hard
If you are not sure about it, reluctant about publishing it, just do it, as you don’t know what will work and what won’t until you haven’t done it.
Watched
Visual Studio Code + Copilot is not open source
VS Code is trying to save tools like cline and augment code
Basically, making developers stick to the VS Code ecosystem as closely as possible
This will make that happen, with Copilot abilities (not the server), access to make the abilities run within VS Code, so people don’t have to fork it.
Gemini 2.5 Pro and Flash are killing it with the new improvements
Gemini is getting integrated into almost everything at Google
This is cool, I talked about it in my thoughts as Appwrite is kind of entering the full-stack as a service ecosystem with this launch
They came #1 on Product Hunt this week
Datbricks bought Neon
This is a interesting buy
Data analysis with the powerhouse database ready for AI
Making a viable integration with the Neon database.
Great move by Databricks
Neon doesn’t have enterprise customers, they are low on funds. But they have a wide adoption and great AI integrations
Databricks has enterprise customers, but not much to offer on the AI side
Combine the,m and suddenly they both become a deadly combo
Learnt
Using Jules to get things done on GitHub
Write the thing you want to do
It will create a branch > you only have to create and merge the PR
Use Open Router to use Anthropic models since it’s not quite reliable with their own API
Tech News
Do we have a website to track which thing, product, model, or tool was released when? If not, I am going to create one specifically for LLM and AI stuff, because this thing is getting out of hand. We are seeing models drop almost every 2 days. This is crazy stuff. OpenAI, Google, Anthropic, Meta, Mistral, xAI, Alibaba, Cohere, in loops over months and weeks with new models. The never-ending cycle. Let’s see if we can predict what model will be next.
Ok, this week we say 5 LLM models are getting released, and those are not any normal models, those are quite a significant push in the state of the art for what LLMs can do.
Anthropic with Claude 4 family (Sonnet and Opus)
Quite solid in coding task, I just tried giving it a golang file and it corrected all the errors and weird quirks it had.
This is a 14 GB, 23 Billion Parameter model for coding task
Not sure about performance but it requires 32 GB Ram to run, so I am out of uck on that.
Waiting for a slight lightweight model. Qwen is a great model in that sense.
Google with the improved Gemini 2.5 Flash and Pro
This was short lived as Claude 4 dropped with all guns blazing
But Gemini is now getting into everything at Google
This is leaning on the magnetic route in every product
This is a bold move from Vercel, could be used by UI developers extensively
Tools
Ok, too much Google
Here are the 100 things that Google released at the IO 2025
Phew! That is a lot of news in a week. Google and Anthropic have nailed it. Google is going all over the place with a commanding position, Anthropic taking the focus to the developer ecosystem by making the model better and better at tool calling and coding tasks.
For more news, follow the Hackernewsletter (it’s a Boeing 747 edition), and for software development/coding articles, join daily.dev.
Which model do you think is the ground-breaking model?
What agentic tools/IDE do you find more useful and help in improving productivity?
That’s it from this 43rd edition of my weekly learning. I hope you enjoyed it, and leave comments on what you think about some of my takes or any feedback.