If you are in Mumbai from tomorrow for Kubecon and other events, apart from the main event - you should surely attend these SIDE EVENTS happening - which are more community led and can help you network with folks closely. All of them require approval, so make sure you regsiter if you are planning to attend.

I see @Zoho everywhere at Bangalore Airport 👀 @svembu

If you are in Mumbai from tomorrow for Kubecon and other events, apart from the main event - you should surely attend these SIDE EVENTS happening - which are more community led and can help you network with folks closely. All of them require approval, so make sure you regsiter if you are planning to attend.

I read this new paper by Hexo Labs called SIA and the framing is sharper than most self-improving agent papers I've seen. the field has split into two camps that don't talk to each other: rewrite the scaffold around a frozen model (prompts, tools, retry logic) do test-time RL on the weights while keeping the harness fixed SIA's move: let a Feedback-Agent choose, generation by generation, whether to edit the harness OR fire a LoRA weight update. same loop WITH two levers. the result that stuck with me: scRNA denoising task. harness-only iteration plateaued at 0.241 mse_norm for generation after generation. first weight-update checkpoint added two lines: clip and round outputs to non-negative integers. basic biological constraint. pushed it straight to 0.289. no amount of scaffold rewriting ever found that fix. the lesson: the harness shapes how the agent searches the weights shape what it actually knows you can't rewrite your way out of a knowledge gap.

Finally it’s the DGX Spark day, tell me what should I build or use it?

was this pre-written :D