Vardhan Shah

I solve problems & build systems.

[Vardhan Shah @durden8055]

@ApoStructura we don’t need to imagine, we have seen this in Star Wars - 2005 movie - parliament of “World Government”

[Vardhan Shah @durden8055]

you’re precisely wrong, when you say bad integration & developer centric abstraction.

To control any system, the entropy of the system needs to be kept bare minimum in regards to growing complexity by design.

Otherwise it can go out of hand created unintended consequences (worst case feedback loop based loops)

That’s why system design even implementation design becomes extremely important - to keep the entropy of the system bare minimum.

[Vardhan Shah @durden8055]

@CausalEngineer @fchollet you have no idea what is software design like LLM doesn’t have. see design not just in software but in reality.

[Vardhan Shah @durden8055]

this is a good idea but don’t let them know the methodology. as they can easily optimise for it. just someone needs to define start state & end state that can be achieved by claude opus high mode autonomously every day. end states shall be comparable with some variations. honestly i dont think they’re nerfing but if they’re doing it, no point asking them. they may nerf to a 30% or so cohort . changing that every day. it’s super easy to do stealthily

[Vardhan Shah @durden8055]

@acdlite are you serious? whole world is running on OOPs. please!

[Vardhan Shah @durden8055]

@ibuildthecloud absolutely. it's for stamping

[Vardhan Shah @durden8055]

@anulagarwal indexing would have been building in the background.

[Vardhan Shah @durden8055]

even if the code is unmaintainable, someone had to take a stab & at least clean area where the next set of development optimisation needed to come.

only way you can run system with unmaintainable code is that there is no evolution left, every demand from the system (limited as it maybe) is fulfilled.

To modify/extend/optimise any aspect of this fragile system in future, resilience of the coder will be needed even more!

[Vardhan Shah @durden8055]

@zeeg So aligned. Believe me people who build real stuff encounters this. IMHO, people building claude should definitely be facing this. If they’re not agreeing, I m not sure if everything going in claude as marketed.

[Vardhan Shah @durden8055]

This is absolutely true. It can’t manage the complexity despite being at minimum context, high thinking mode. It makes many mistakes as thinking doesn’t require verifiability - you can make stuff up & once first step is in wrong design everything overwhelms & snowballing starts. Honestly at this stage, model doesn’t have capacity to hadle any complexity

[Vardhan Shah @durden8055]

@trq212 @zeeg if any senior meaningfully want to contribute they would have to lock in & deliver complete unit. otherwise seemingly harmless code is overengineered slop buggy code.

either one has to constantly nudge or someone else will have to pickup the work.

[Vardhan Shah @durden8055]

@trq212 this was so needed! thanks

[Vardhan Shah @durden8055]

@championswimmer so true!
context fill management is superimportant but super annoying

[Vardhan Shah @durden8055]

@paraschopra *more you understand..

knowledge is meant to be understood.

[Vardhan Shah @durden8055]

task switching truly fries the brain. my capacity is 1:30-2 hrs. then depletion non linearly accelerates

[Vardhan Shah @durden8055]

once opus 4.6 overengineer.

it can’t cleanup.

it would overengineer more if you ask to simplify- because in their context it try to find meaning in the slop & create more of mess.

[Vardhan Shah @durden8055]

can safely say providing all the context, skills & detailed implementation approach also don’t help. as it can’t do sophisticated causal thinking. in itself in bypass permissions it can’t do. if you keep an eye & steer & nudge, it may do well, but then task switching comes into play as it is very hard sit idle, stare screen & ai do slowly the work.

[Vardhan Shah @durden8055]

@thdxr same here!

[Vardhan Shah @durden8055]

Are LLMs falsifiable?