Founder of Zenact AI | Founded Opti Owl Cloud | ex SRE Lead @Zomato | IIT Dhanbad | all opinions are my own
India
Joined March 2014
- Tweets 1,628
- Following 5,521
- Followers 740
- Likes 5,188
alright, with all the power vested in me
i hereby declare from now on, all uni ball one p 0.38mm gel pens are speculative tokens whose value directly indicates the value of Zenact AI
Srinivas Devaki retweeted
"building someone else's dream" is a pretty dogmatic view IMO
every great thing the humanity achieved, a lot more people believed in the mission than focussing on whose mission it was or who started that mission
duty and mission are real things
stop reporting issues
security teams need to be more solution oriented, flagging issues is just not enough
same goes for bug detection, its not enough at all, solutioning would soon become much more important for high performing teams
no wonder jevons paradox keeps coming up
for qa & testing though, its literally true, if costs come down, one could easily spend a billion tokens and cover 99% of the issues that only an expert human tester could spot
and the model is such a brittle model, certain keywords will make the model go crazy
"as little as possible" is one such keyword for today
just stop talking
gemini for agents is such an overworking model, just stop, you completed everything, just stop
Two reactions when people hear about the 996 culture:
😂 "That’s nothing, we’ve been doing 24/7 for years."
😂 "Bro, that’s fake news propaganda."
Either way, everyone’s tired.
there is no perfect way to interview
literally every metric is a proxy, and literally any lame proxy metric will work if the range is big enough and used consistently across candidates
doesn’t matter whether it’s leetcode or something lamer like typing speed or math blitz
its a superpower and a curse - we normalise magic and innovation soo fast
i wouldn't be shocked if aliens come to earth and everyone normalises it within a couple of weeks
most externalise all kinds of extraordinary events rather than internalising what it means to them
slower models like gpt-5 are very bad for coding workflows that require precise changes
its nothing to do with the model accuracy itself, the problem is with how operator is affected by the slowness of the model, it almost always means precision & care is lost
wish there is a way to have codex on high, get all the work done
and use the reasoning text generated before to now create docs with a smaller model
ability to solve complicated problems isn't same as solving complex problems
former is layered simple problems, later is tightly linked problems that relate to each other
somehow iits mostly produce talent that can solve the former than the later
in old saas world, many companies may not have any true IP
but with AI its definitely different, i guess it gets true whenever a new thing comes into the ecosystem where everything needs to be built on a new paradigm
and yes, this applies to even wrappers
