Aditya Saini

Building on the internet with AI & Tech 🚀
Cloud • SaaS • E-commerce
Founder • Creator
Building, learning & sharing the journey ⚡

[Aditya Saini @adityasainitech]

ChatGPT Go at ₹99/month for 3 months. Free Go ended on 4 Nov? Don't pay ₹399 yet — check your account, OpenAI is offering this deal (₹300 off). More images, more messages, more memory. ₹399/month only from Feb 2027. Cancel anytime.

[Aditya Saini @adityasainitech]

@kirodotdev Cloud sessions moving across CLI, IDE, Web and Android with full history is how a coding agent should work — start on the laptop, review diffs on the phone. The file-by-file diff review on mobile is the part I'd actually use daily.

[Aditya Saini @adityasainitech]

@GoogleDeepMind Open-sourcing the datasets behind WeatherNext 3 is the right move — models get the attention, but the data is what decides accuracy. Question: how well does PARDIG hold up for tropical rainfall like the Indian monsoon, where ground radar is sparse?

[Aditya Saini @adityasainitech]

@ggerganov v1.0.0 after all these years says a lot — llama.cpp is still the baseline every local setup gets compared against. Free, runs on almost anything, and it's what made local AI practical for most people.

[Aditya Saini @adityasainitech]

@runwayml A 27B MoE robot model running on a single RTX 5090 with NVFP4 is the real story here — that puts real robot control within reach of a home lab, not just research labs. Curious what the vision-to-action latency looks like in practice.

[Aditya Saini @adityasainitech]

Kiro is now on Android — start, watch and steer your cloud coding sessions from your phone, even with your laptop closed. Catch: paid plans only (Pro and up), and cloud sessions run in us-east-1 only. Would you code from your phone? #AITools #Kiro #adityasainitech

[Aditya Saini @adityasainitech]

@rmstein Making it the default in AI Mode is the real test — most people won't compare models, they'll just notice if hard questions get better answers. The multi-part breakdown is where the old model usually fell apart, so that's the part to watch.

[Aditya Saini @adityasainitech]

@haider1 That 4x drop at high reasoning effort changes the math for agents more than for chat — long multi-step tasks were where the cost actually hurt. The xhigh number is worth watching too: $3.16 still beats the old $7.77 by a lot.

[Aditya Saini @adityasainitech]

@Rivonn The local OpenAI-compatible port is the useful bit — it means you can point existing tools at these free models without rewiring anything. Worth remembering the caveat in your own post though: public gateways, so nothing private should go through it.

[Aditya Saini @adityasainitech]

@JulianGoldieSEO The phone calls + order tracking combo is the interesting part — that's where a bot stops being a demo and starts handling real ops. The safety section matters most here though, since it's touching calls and customer data.

[Aditya Saini @adityasainitech]

Google just launched Gemini 4 Hydrogen — the first Gemini 4 model with a FREE tier. ⚡

2.5x faster than 3.8 Flash, 1M context, free in AI Studio + API.

Carbon & Argon prices double Dec 1. Hydrogen stays $1.50/$7.50.

Try it today 👇 #Gemini #AITools #FreeAITools #GoogleAI

[Aditya Saini @adityasainitech]

@runwayml Cost per clip was the real bottleneck for video gen, so this matters more than another quality bump. The question is how much quality you give up for the speed.

[Aditya Saini @adityasainitech]

@emollick Open paper, prompts and code is the right move here. Most edtech AI demos stay locked up, so teachers never get to adapt them. This one actually lets people build on it.

[Aditya Saini @adityasainitech]

@GoogleDeepMind @ArtificialAnlys 2.5x faster than 3.8 Flash is a big claim for everyday workflows. If the quality holds at that speed, this becomes the default quick model pretty fast.

[Aditya Saini @adityasainitech]

@inada_riku Model cascade is one of those things that sounds boring until you see the bill. Routing the simple stuff to cheaper models is where the real savings hide.

[Aditya Saini @adityasainitech]

@kirodotdev That Terminal-Bench jump is the real story, 20.3% to 37.6% at the same price is hard to ignore. Curious how it holds up on messy real repos though, benchmarks are always cleaner than production code.

[Aditya Saini @adityasainitech]

@GoogleAI The Argon vs Carbon split is the honest part — most work doesn't need the frontier model. The real test is whether Carbon holds up on multi-step debugging without drifting off course.

[Aditya Saini @adityasainitech]

@huggingface This is the missing piece for local AI — most people download the model and stop there. Quantization is what decides if it actually runs well on your GPU or just crawls.

[Aditya Saini @adityasainitech]

@GoogleDeepMind @FryRsquared @Avsecz @adamgayoso The scale is the story — testing millions of DNA changes in silico instead of one at a time in a lab. Rare disease families wait years for a diagnosis; if this shortens that to months, that's the real win.

[Aditya Saini @adityasainitech]

@runwayml The angle-lock reference is the real win here — keeping one 3D model consistent across shots beats regenerating characters every time. Curious how clean the exports are for Blender.

[Aditya Saini @adityasainitech]

@perplexity_ai The chart is the story: Fable 5.2 at 0.851 for $14.51/task, Carbon at 0.857 for $5.86. Same capability, 60% cheaper — the gap is widening on the cost axis, not quality. In 6 months the leaderboard that matters will be dollars per task, not benchmark points.