Tamojit Roy

SWE | TECHNOLOGY GEEK | COMPUTER SCIENCE GRAD | THINKER | YOUTUBER | FOUNDER & CEO AT @AstroDestroyer146

[Tamojit Roy @AstroDestroyer]

I've been using the Gemini 4 Argon model for quite sometime now and have been testing it thoroughly for various purposes. I've developed 22+ static websites (clones + original web-apps) using it. These are some of my observations:

● It takes a much longer time to think. Behind the rolling of the "Thinking it through..." one-liners, there's a codon stream of complex computation, experimentation and rapid decision-making involvement. Even for a very trivial task, it refuses to go the easy way and takes it's time.

● It *almost* never asks for additional instructions: I felt this quite annoying. The model almost never showed it's "work" before jumping to the conclusion. I got the bundle/build-file/zip straightaway 99.99% of the times.

● It does make "mistakes" occasionally: Although this should depend on the "prompt" quality, I noticed it when I instructed it to build a "JioHotstar" clone. The UI was broken for 1-3 Catalogue Cards. How crazy is that?

● It rides roughshod over when it gets overwhelmed: Imagine AI getting *overwhelmed* due to a prompt. As uncanny as it sounds, it leads to unexpected consequences. My Gemini became unresponsive and it even refused to entertain any further requests in the same chat context! All for just giving the requirements? I never talked technical, it did. Could it be that it seemed to get lost in its own hyper-intelligence. Who knows?

● It can choose *shortcuts* when it's confused: There was a requirement for a feature which would fetch and populate latest AI news on UI, but the feed should be dynamic and constantly update 1-2 days. It simply made the same 3-4 cards change their order periodically every 48 hours, no dynamics whatsoever!

● It still doesn't have "PROMPT → VIDEO" generation capability. The best it does is create a slideshow of *stylised* and *wobbly* HD images. Not acceptable.

● It nails even the hardest problems on 'Codeforces': I gave Argon to solve few "HARD" coding problems of both *Leetcode* and *Codeforces* and I was impressed. Not only did it gave the accurate solution and passed all test-cases, but also it suggested >1 alternatives! (with proper explanation; algo used; time, space complexity comparison; approach) This was epic! Is this the end of coding assessments? I mean if someone can somehow slip into a different system w.r.t. the system he's taking the online assessment, and somehow manage to route the stream of questions (which includes even the most difficult of coding problems) it's *GAME OVER* for the examiner! It would become very hard to answer any question 'wrongly'. Thoughts?

● Runs out of quota quite easily: I wouldn't say I went easy on it with tasks, but I personally felt it rans out of gas quite easily as compared to its predecessor models. The computing power and usage rate-limits are very strict! Waiting for it to renew felt frustrating.

● Less interactive: Prompting it felt like a Half-Duplex communication. It never asked for feedback nor anything which it couldn't understand. It just came up with something after thinking for long.

● Has a long way to go: Although, it came up with stunning and mesmerizing UI designs for my web-apps, it still was not at par on the system design, backend and functional front, once again reiterating why frontend developer jobs will fade away faster than backend engineering jobs (or, it has already happened?).

In conclusion, Gemini 4 Argon is by far the best flagship AI model which Google has released, but it still doesn't perform promisingly when it comes to understanding nuanced conversations, getting the hang of things (especially when the main instruction seem to get overshadowed by other irrelevant details present in bulk). The subscription price is reportedly going to double, after November. It is not good. Never been a big fan of benchmarks, hence will not comment upon it. It has limited public access, high costs, and mixed real-world performance compared to competing frontier models.

[Tamojit Roy @AstroDestroyer]

Actually expected more from Gemini 4 Argon. Wanted cinematic trailer of one of the famous Durga Puja clubs but seems a bit off with terrible music. Maybe, the prompt was not cut, but who knows? Says, it doesn't have direct access to a live-action neural text-to-video generator.

[Tamojit Roy @AstroDestroyer]

Gemini 4 Argon impresses with competitive introductory pricing, enormous token capacity, and versatile capabilities. Despite mixed opinions and limited personal testing, its potential makes me curious about AI’s future.

[Tamojit Roy @AstroDestroyer]

Created a video on my recent shopping experience in Lenskart.
https://t.co/GrKGiuF8jR

[Tamojit Roy @AstroDestroyer]

@LenskartSupport @Tamojit I've dm-ed. But, not able to send attachments. Kindly share email. Thanks.

[Tamojit Roy @AstroDestroyer]

Paid ₹3399 at a local @Lenskart store on Saturday. Money debited, card machine printed a slip, but store says order is "cancelled" in system. No glasses, no refund yet! App support promised resolution by today EOD but still waiting. Please resolve this ASAP! @lenskartsupport

[Tamojit Roy @AstroDestroyer]

Go, check it out if you haven't already!
https://t.co/5DhW3nzM9H

[Tamojit Roy @AstroDestroyer]

Ever heard of Kratos Gear Up? Well guess what. I've been using its Ear Buds for quite a few days now and the battery life is amazing. Although, it has been forecasted to last only around 6-7 months, let's be optimistic. Also the bass is strangely captivating. Enjoying it so far.

[Tamojit Roy @AstroDestroyer]

New Unboxing Video Alert!! ⚠️ Check it out. 🤩
https://t.co/FHzS71NW2t

[Tamojit Roy @AstroDestroyer]

@justbyte_ Because that's what legends do

[Tamojit Roy @AstroDestroyer]

@prasenx Make smoothies

[Tamojit Roy @AstroDestroyer]

@takeUforward_ https://t.co/ScJF7ovOxq

[Tamojit Roy @AstroDestroyer]

@AdityaShips https://t.co/ScJF7ovOxq

[Tamojit Roy @AstroDestroyer]

But, meanwhile the original goal remains unaccomplished (which was to create a stable, progressive and growth oriented hive mind). We are masters of introducing ideas, losers in understanding problems and creating meaningful solutions.

[Tamojit Roy @AstroDestroyer]

Agentic AI and vibe coding isn't even funny anymore! Earlier what was introduced to bridge the gap between data and humans is increasing it even more. It seems like it's bridging the gap between data and AI (leaving humans out of the room).

[Tamojit Roy @AstroDestroyer]

So, you need more people, the size of the team increases, which often leads to more thoughts, PoCs, eventually introducing something else entirely after AI.

[Tamojit Roy @AstroDestroyer]

@shiri_shh https://t.co/ScJF7ovOxq

[Tamojit Roy @AstroDestroyer]

@Manixh02 Ctrl + C → Ctrl + V 😋

[Tamojit Roy @AstroDestroyer]

In a few days, we're going to step into the unknown challenges as well as opportunities which adorn the brand new year of 2026! Let's hope for a better and inclusive future for all and be eager to embrace AI as much as possible, knowing that it's a truth which can't be ignored.