Yash Rawat

SWE @Samsung Research

SDE day notes + AI that survives production

[Yash Rawat @thinkofyash04]

I write two things here:
1. What an SDE day inside a research org actually looks like
2. What I’m learning so I can build AI products that survive production

SWE @ Samsung Research. Building on the side.

No course. Just notes from the desk.

[Fei-Fei Li @drfeifei]

Agents can be AI, but agency is human. It is every individual’s responsibility to empower our own agency, with the help of tools. But the North Star should always remain human centered.

[Yash Rawat @thinkofyash04]

Exactly. My colleagues and I discuss this quite often now. Debugging has this whole extra phase where you first have to rebuild your understanding of the system before you can even investigate the bug. The fix might still be a few lines, but getting confident about those few lines takes most of the time.

[Yash Rawat @thinkofyash04]

We got GTA 3 running in a browser before GTA 6 😭

How? WebAssembly runs the game engine, your GPU draws the world, and assets load as you drive.

The game we used to install now opens in a tab.

https://t.co/s9QgQNymhI

[Yash Rawat @thinkofyash04]

@Polymarket The AI didn’t get distracted. It was performing… unsupervised reinforcement learning.

[Paras Chopra @paraschopra]

It's kind of crazy that SO MANY smart people have a goal of bringing full autonomy to the world, when it is clear that it'll leave everyone jobless.

You ask these smart engineers or researchers if they've thought about consequences of their actions, and they'll simply shrug.

[Yash Rawat @thinkofyash04]

@9to5mac @apollozac Codex + ChatGPT upgrade on the same model is what I’m most hyped for.
If coding quality jumps the way the leaks suggest, a lot of workflows get rewritten this week.

[Yash Rawat @thinkofyash04]

@kimmonismus Brutal timeline.
If Astra holds up outside the benches, this is the first model in a while that feels like it changes how we build, not just how we chat.

[Yash Rawat @thinkofyash04]

@diogo_valada @thenewstack The agentic jump is the part I’m watching.
Benchmarks are nice, but if it can actually run long workflows without babysitting, that’s the real launch.

[Yash Rawat @thinkofyash04]

@jxnlco Please. I’ve got features sitting in a branch waiting for this model.
Daybreak first is understandable, but the wait is brutal.

[Yash Rawat @thinkofyash04]

@M1Astra These numbers don’t even look real.
If coding + agents feel like this in production, a lot of current stacks just got outdated overnight.

[Yash Rawat @thinkofyash04]

@johnboy232321 Phased rollout + those benches is the combo.
98.6% ARC-AGI-3 and 100% ExploitBench is wild. Can’t wait to throw real product work at it.

[Yash Rawat @thinkofyash04]

@xhSamui Same. Refreshing ChatGPT like it’s a game drop.
Limited rollout first is fine - just let builders in already.

[Yash Rawat @thinkofyash04]

@testingcatalog “AGI today, but only in Daybreak” is the most 2026 sentence possible.
Still hyped. Staged rollout makes sense given the cyber capabilities - just refresh ChatGPT every 20 minutes like the rest of us.

[Yash Rawat @thinkofyash04]

@synthwavedd Limited Daybreak access first, then everyone else in a few days. Painful wait, but the numbers look like a real jump.
Most excited to see what it does on long coding + agent tasks.

[Yash Rawat @thinkofyash04]

@scaling01 98.6% on ARC-AGI-3 is insane.
If this holds in real use, a lot of this still needs a human in the loop workflows just died. Can’t wait to test it on actual product work.

[Yash Rawat @thinkofyash04]

@KobeissiLetter Generational leap + cautious rollout is the right combo.
Watching this from the builder side - if Astra is even half as good as the early signals, a lot of current workflows are about to get rewritten.

[Yash Rawat @thinkofyash04]

@StockSavvyShay people may look back at this model as AGI is a wild sentence to drop on a Thursday night.

[Yash Rawat @thinkofyash04]

@wallstengine generational leap for the model. generational chaos for anyone who ships without reading the diff.

[Yash Rawat @thinkofyash04]

@unusual_whales AGI or not, they just handed vibe coders a model that scores 100 on ExploitBench.

[Yash Rawat @thinkofyash04]

@OpenAI Vibe coders just unlocked creative mode.