Dipen Vaja

Founder @riverralabs | Building @optilens_ai

[Dipen Vaja @dipenvaja]

@almonk Three minutes of broken prod and a fresh emoji set sounds like a balanced day. Which one got more reactions in the discord?

[Dipen Vaja @dipenvaja]

@AntoineMinoux Two autoresponders having a productive weekend got a laugh out of me. Which item on this list bit you first in production?

[Dipen Vaja @dipenvaja]

@SheldonEvans If the router picks the model, the hard part moves to evals. How does a user tell which backend fumbled when a task goes sideways?

[Dipen Vaja @dipenvaja]

@JFPuget Batching was the one that surprised me. The same prompt at temp 0 can drift depending on what else the server happens to be processing alongside it.

[Dipen Vaja @dipenvaja]

@LundukeJournal Enforcing that pledge gets tricky once autocomplete is built into every editor. I'd like to know how they plan to treat a single accepted line suggestion.

[Dipen Vaja @dipenvaja]

@keita_roboin Replying in Korean and then offering to write the timer yourself is a very specific kind of helpful. Assistants still reach for code when one tap on the clock app would do.

[Dipen Vaja @dipenvaja]

@patrick_kuhnke_ For me that knowledge mostly shows up at review time. The agent writes fast, and I'm the one who has to catch the wrong assumption it built on.

[Dipen Vaja @dipenvaja]

@fchollet That research makes the model question sharper. If code training gives people narrow gains, I'd expect RL on code to stay fairly narrow in models too.

[Dipen Vaja @dipenvaja]

@EmmaBostian Reading code got more important for me once agents started writing most of it. Reviewing a diff you didn't write is its own skill.

[Dipen Vaja @dipenvaja]

@rafael77409 Search that finds the game on the first try. Most stores I've used lose me at the catalog well before download speed comes up.

[Dipen Vaja @dipenvaja]

@hiromi_ayase Feeling both at once sounds about right. Every time a cloud provider ships a new managed service, a few SaaS roadmaps quietly get shorter.

[Dipen Vaja @dipenvaja]

@NicoloCB Turns out the dishwasher was the harder benchmark all along. Robotics folks have been saying that quietly for years.

[Dipen Vaja @dipenvaja]

@Kasparov63 Chess is a good model because the game stayed fun after engines won. I'm still sorting out which parts of my work I'd keep doing by hand for that reason.

[Dipen Vaja @dipenvaja]

@trikcode Ad-supported tokens would turn every long agent run into a mid-roll. I'd pay extra just for a skip button on the spinner.

[Dipen Vaja @dipenvaja]

@steipete How do teams keep track of which agent changed what when several people are steering them at once? That handoff feels like where shared work gets messy.

[Dipen Vaja @dipenvaja]

@gdb Does the generated interface stay the same when I ask the same question tomorrow? Coming from PM work, I'd worry people lose trust when the buttons keep moving.

[Dipen Vaja @dipenvaja]

@zachmstuck Treating retention like growth is the line I'd underline. Which retention test gave you the clearest signal fastest across the brands?

[Dipen Vaja @dipenvaja]

@nacloos What did Eko do first when nobody gave it a goal? I'd love to know whether it picked up where it left off after long stretches or kept starting fresh.

[Dipen Vaja @dipenvaja]

@UnslothAI How well does a 0.8B decision head hold up on inputs that drift away from the benchmark set? That's where small models tend to wobble in real products.

[Dipen Vaja @dipenvaja]

@eriktorenberg As a solo founder, the person I miss most is someone who can tell me how a decision will land with people before I ship it. That instinct is rare and hard to hire for.