Vikas Aditya

Flying on @SouthwestAir today. Plane is 30% full, overhead bins empty… but my backpack must go under the seat. We’re officially at the stage where humans are executing policies like badly fine-tuned AI models—zero context, zero judgment. Thanks Southwest Airlines 😄

Sometimes the most important product moments are the ones users barely notice — because everything just works. Really proud of the HackerEarth team for recently pulling off one of the largest campus hiring events for one of the most prestigious tech company: 125K student registrations, nearly 70K candidates in a single test, 0.5M+ code submissions, 1M+ test cases, 16M+ proctoring snapshots, and 2.5M+ reporting events. At that scale, reliability is the product. What impressed me most was not just the numbers, but the quality of execution behind them. The team strengthened the assessment experience, scaled evaluation and proctoring pipelines, reworked reporting for near real-time visibility, and built the platform to handle extreme bursts far more predictably. These are the kinds of engineering investments that don’t always show up in flashy demos, but they make all the difference when thousands of candidates are depending on your platform for a career-defining moment. Moments like this remind us that evaluation infrastructure is not only about capabilities — it is about trust at scale. When candidates show up for an important opportunity and companies are trusting your platform with their brand, performance is non-negotiable. Big shoutout to everyone across engineering, product, support, sales, and ops who made this happen. Grateful to be building with a team that cares so deeply about scale, resilience, and candidate experience.

At #aiimpactsummit, almost all major companies are there. They are showcasing the capabilities of their platforms. While LinkedIn is showing how jobs will be transformed, top of mind for most attendees is if Al will replace their job. We at @HackerEarth are doing a live poll of which jobs will be replaced by Al. And we have a mirror that says "You are looking at someone who will not be replaced by Al". Based on the interest attendees have in taking their selfie in front of this mirror and in participating in the poll, it's clear that we need to talk more about Al's impact on jobs more than anything else.

There is so much noise about models reaching human capabilities on several benchmarks. But not for real world tasks. Not even close. The highest rated model on vibecodearena.ai is Grok and sits well below human scores.

Good energy and vibes at #aiimpactsumit in India. @HackerEarth is presenting. We also see many students and non technical folks. Which jobs AI will replace (including their own) is on top of people's minds

AI Is Rewriting Software Engineering Jobs, and 2026 Hiring Will Reward "Aptitude Over Syntax," New Data Suggests prn.to/3YETJO9

So benchmarks like SWE-bench really matter? Models have learned how to crack benchmarks. We need to test them on real world apps. I used Gemini and GPT to generate code that draws city skyline. Here are the results: Gemini: produced London skyline with landmarks. Analytical, Structured and more efficient GPT5.2: produced New York skyline. It's more creative, deputy, artistic. Here is full analysis with working code: vibecodearena.ai/duel/4eccbccc-…

It looks like we'll be getting an upgrade to Gemini 3.0 Pro soon! Nice!

Another fun app I built at vibecodearena. This time using Claude and Kimi-K2 While Claude did score better, Kimi-K2 did pretty well and considering that it's open source, it's a tough contender Check it out here👇 vibecodearena.ai/duel/ac041a46-…

@victormustar Used VibeCodeArena to develop a 5x5 Tic Tac Toe game and compared Claude and Kimi-K2. Kimi-K2 did pretty good imho. Check it out 👇 vibecodearena.ai/duel/ac041a46-…

Kimi-K2 ✖️ Z-Image-Turbo (god I love open source)

What creative personalities have you noticed in the models you use? Claude vs GPT? KimiK2 vs Llama? Gemini vs anyone? Drop your observations below. Genuinely curious what patterns people are seeing. 👇

Entire analysis with generated code for both models is at vibecodearena.ai/duel/4eccbccc-…

I used vibecodearena.ai to develop p5.js code using GPT-5.2 and Gemini 3 Pro - the two leading models. Exactly same instructions: "make a p5.js illustration of different city skylines during sunset" Both model generated the code but they have completely different creative DNA. Thread on what this reveals about AI personalities 🧵

Gave GPT-5.2 and Gemini the same prompt: "create sunset skylines in p5.js". Both leading models. Gemini: 53 sec, 526 lines, accurate London skyline but feels icon based GPT-5.2: 297 sec, 1086 lines, Feels more artistic, atmospheric mood piece Gemini seems more analytical (left brain) whereas GPT5.2 appears more creative (right brain). But we need both, right? Full comparison at vibecodearena.ai/duel/4eccbccc-…

If you have used both Nano Banana Pro and GPT Image 1.5, do you really think this ranking is correct? What is your verdict?