Vaibhav Gupta

I don’t enjoy the syntax and tooling we use to building AI pipelines. Langchain/ai sdk/et all have never really felt “right”. No type safety, single language support, prompt is hidden until I buy some observability suite. So we made a thing youtu.be/wD3zieaV0Yc?si…

@MichaelThiessen @dexhorthy jokes on you, cause i'm a clown 🤡

@dexhorthy You and @vaibcode should hash this out on an episode because it seems you have opposite stances. I’d love to hear this but with more nuance and tldraw sketches!

louder

Code can be sloppy. Writing cannot.

@sean_j_roberts To be honest, I felt the same a few months back. I had to change my workflow a lot until I felt emotional comfort in not reading all the code. If you’re curious about it, shoot me a dm. I’m happy to have a conversation around how.

@vaibcode I honestly don't think this is good enough. I'm not willing to sign my name to code I haven't looked over.

Everyone is complaining about how models are non-deterministic. Yet most employees in a company are non deterministic.

@sean_j_roberts Models understand most code better than most humans. And models don’t leave or have 6 months of “onboarding” time.

@vaibcode It's not just about checking for mistakes, it's also about having actual humans who understand the code.

The point is not that humans are bad. Humans are freaking amazing. And so are models. Non-determinism is a feature.

@RyanJamesShaw Interviews and performance reviews are some of the worst processes I’ve seen for filtering people.

@vaibcode False equivalence. Employees go through interviews, delivery reviews and performance reviews. Do LLMs?

@SalzDevs Have you ever seen code produced by humans?

@vaibcode Thinking is non-deterministic, period. That's not what people mean when they complain about models being non-deterministic. The issue is predictability: the same input can produce materially different outputs, making debugging, testing, and reliability much harder.

@RohitPaulK Sure, that’s why I don’t believe in single person companies. But most companies have a non-negligible amount of unreliable folks. And even the best employees have off days. You still need processes.

@vaibcode What people are actually complaining about is that models are unreliable. Good employees are reliable

@techguy1360661 to me the best teams are those with great processes and don't rely on any person knowing everything. in most extremely large companies i've worked in, the person who built the original thing is long gone and context is missing all the time. we still ship.

@vaibcode Continual learning and context building over time in humans

@sujay_kapadnis congrats, you just invented 16th century calvinism

@vaibcode Nothing is non deterministic

@890691905140v define it for me.

@vaibcode u are too young to understand what non determinism means

@techguy1360661 tell me more. what am i missing?

@vaibcode profoundly retarded. Like you have never actually worked with an llm in practice

The goal is to build systems that a company can trust so that mistakes don't happen. And when a mistake does happen, have it be fixable. Many direct managers will often delegate code review to other senior engineers. If not, go up a level to M2, they definitely are not reviewing the code.

@vaibcode Yeah and we review their code!

You don’t blame your engineers when a bug ships. You fix the process. Do the same thing for ai

@MichaelThiessen I think so! But even more so I think it’s about being confident that the system isn’t broken. Managers have been “shipping” without reading/understanding code for long before ai!

@vaibcode I think that maybe what we’re after is *understanding* what the code is doing, which doesn’t require reading every single line. Not understanding the code (or at least the architecture and critical parts) is where the vibe coding nightmare begins.

Wrong. Reading code is about increasing the level of certainty. There are many things that can also increase levels of certainty that make reading 100% of the code lower alpha. System Invariants, ci/cd, pre roll outs on beta systems, test coverage, reading the tests, design doc.

@ku1deep I think people can have done amazing things and still be wrong about different things. Sentry is an amazing company, but I just don’t think reading every line of code is required. And more importantly not reading every line of code doesn’t make someone not serious

@vaibcode I have to say I admire your confidence in calling David wrong like that. I mean one would think someone with that level of accomplishment would invite a bit more professional respect and curiosity.

Also most humans good at writing code, are not always good at code reviews (ie reading code). It is a poor man’s argument to assume that if AI is writing the code and humans are reading it that you have the same levels of certainty as you did when humans were writing it.

@dexhorthy Iteration not recursion. Back to cs 101

loops is my favorite form of recursion