byte | sumit sute

██████╗ ██╗ ██╗████████╗███████╗

██╔══██╗╚██╗██╔╝╚══██╔══╝██╔════╝

██████╔╝ ╚████╔╝ ██║ █████╗

██╔══██╗ ╚██╔╝ ██║ ██╔══╝

╚█████╔╝ ██║ ██║ ███████╗

╚════╝ ╚═╝ ╚═╝ ╚══════╝

Some thoughts don't need essays. They just need somewhere to land. Bytes are the small ones: thoughts that appear mid-commute, mid-debug, mid-coffee. I post them from Telegram or the terminal, whichever is open, and whichever I'm probably procrastinating in when the brain buffer overflows. Stay connected on the telegram channel for latest bytes.

There is kind of debugging loop with AI where you keep asking follow-ups until the problem disappears. It works, but it also shifts the focus away from understanding what was actually happening.mar 27

xxx

The difference between asking follow-up questions to improve a solution and asking them to improve your understanding is subtle, but it compounds quickly.mar 27

xxx

Yes, comprehension has historically been built through friction, while agents optimize for removing friction and providing more delight in delegation. We should employ the agents to also simulate this friction to continue lab training our comprehension muscles.mar 24

xxx

But there are no perfect specs and there no complete tests. Systems are often understood in motion.mar 24

xxx

Cheaper code doesn't mean cheaper comprehension though. Comprehension is still the bottleneck that doesn't quickly scale with automation and remains stubbornly human.mar 24

xxx

Good architecture compresses context. It lets an agent load the right truth without having to reconstruct it from scattered clues.mar 22

xxx

The cost of bad structure is often seen in terms of scalability but it can also show up as hesitation every time you need to edit something.mar 22

xxx

Turns out agents don't do therapy well and no point in asking 'how do I get better?'. Switching to diagnose → hypothesize → debug.mar 20

xxx

Your inner engineer wants to babysit the agent. Resist. Turns out the job is less 'micromanage the robot' and more 'be a decent manager'. Leadership > Overengineering.mar 17

xxx

Reminder to self. LLMs are stateless. You pay for the whole story every time. As conversations with agents get longer, each prompt gets compoundingly expensive even with engineering solutions to optimize this with cache, compression, decomposition, parallel agents etc.mar 17

xxx