Aditya

20, engineer who can yap about tech & code

old account : @astrobot_me

[Aditya @maxthirty3]

Kinda milestone for me, guess which one I'm? 🤗

[YAS Ministry @YASMinistry]

At Viksit Bharat Young Leaders Dialogue 2026, young innovators selected through Design for Bharat Challenge and Hack for Social Cause presented their impactful ideas and prototypes to Hon’ble Prime Minister Shri Narendra Modi. Reinforcing the role of youth in nation-building through an innovation-led journey, moving closer to the vision of a Viksit Bharat by 2047. 🇮🇳✨

#YoungLeadersDialogue2026 #MYBharat

[Aditya @maxthirty3]

Due to AI mode in google search now its almost me every time doing google search and looking documentation of something then to starting a new Claude or Gemini or gpt chat

#AImode #Gemini #googlefordevs

[Aditya @maxthirty3]

Since couple of days im working with @openclaw , even for doing cli setups you have an agent called cres. which you can chat with and setup

I'm seeing where it's going, asking llm for something that you can itself look into documentation and do it yourself by spending some time

[Pawan @PawanKChandana]

It’s been 24 hours now…

Vikram-1’s launch still feels surreal.

Knowing just how hard this is, I couldn’t contain the excitement at Mission Control as the rocket we built with so many dreams successfully inserted satellites into orbit — on its very first attempt.

Just phenomenal! 🔥🚀🚀

[Kimi.ai @Kimi_Moonshot]

Meet Kimi K3

[Gareeb Scientist @gareebscientist]

What a dayy what a day ❤️🚀

[Aditya @maxthirty3]

wtf

[Aditya @maxthirty3]

I have adhd with my system monitor, and I keep on killing unnecessary tasks or processes that I see on the system monitor

peakk resource management

[Aditya @maxthirty3]

@Ayussheth I would like to join

[Aditya @maxthirty3]

We're hosting a bootcamp

Join us if you're nearby

Best regards
Aditya
(Lead, GDG RIT)

[Hack2skill @hack2skill]

Roorkee, get ready 🚀

Build with AI Bootcamp is coming to Roorkee Institute of Technology on 5 June 2026.

Hands-on AI learning with:
✔ Certificates
✔ Guidance from Google experts
✔ Real building experience

Register now: 🔗 https://t.co/mN17cclkaK

#BuildWithAI #GoogleAI

[Aditya @maxthirty3]

I'm stuck in a loop🔁

A: Loads 5b ~ 10 gb model
B: Loads 800, 000 chars
C: Start the Training
D: My notebook Crashes after a few iterations due to GPU vRAM

And here I'm at step A again! WTH

[Aditya @maxthirty3]

@insanekrishnaa Interested

[Aditya @maxthirty3]

@pallavishekhar_ Good content 👍

[Aditya @maxthirty3]

So the improvement is real in some ways, but not always as generalizable as the benchmark numbers suggest.

[Aditya @maxthirty3]

So the key takeaway is:

SWE-bench measures performance on a specific , 12 Python repos and their issue distributions — not necessarily “general software engineering intelligence.”

[Aditya @maxthirty3]

The caveat is that if a benchmark tests on these same repos (like SWE-bench does), there's a risk of data contamination — the model may have seen the issues during training, inflating apparent capability.

[Aditya @maxthirty3]

Why the massive leap?

Once the benchmark paper revealed the exact 12 repos, AI labs had a clear optimization target. Models were heavily tuned/trained/evaluated against those repositories and issue patterns.

[Aditya @maxthirty3]

Fast forward to now: leaderboard scores suddenly jumped to ~76%+ on SWE-bench.

But this jump needs context.
Once SWE-bench became the industry-standard benchmark, the exact repos, issue styles, and evaluation setup became widely known.

[Aditya @maxthirty3]

One of the most cited coding benchmarks for LLMs is SWE-bench.

It evaluates models on ~2,000 real GitHub issues collected from 12 Python open-source repos.

Back in 2024, even top models like GPT-4 and Claude solved only ~1–5% of tasks.

[Aditya @maxthirty3]

Agent2Agent Protocol is not supported for typescript based adk agents? @googledevs

Or docs are not updated