Kartikey Saran

tech | SaaS | Ai agents | Building things

[Kartikey Saran @CoderKartikey]

what is your source of learning guys ?

[Kartikey Saran @CoderKartikey]

are uh still reading the books to learn something or using the ai for it ?

[Kartikey Saran @CoderKartikey]

are Chinese models are falling behind now?
bcz back then there was news that they distilled the Claude and they were superb that time like GLM 5.3 was insane but now we didn't saw any new magic from them...

[Kartikey Saran @CoderKartikey]

Do we have the market in trading sector ?
I m launching something in this sector ...

[Kartikey Saran @CoderKartikey]

@james_web3sec @james_web3sec yeah I wanted to talk. I actually tried to dm you before reaching out on the post but your dm is locked for only those who know ur X number so i can't dm you 🥲

[Kartikey Saran @CoderKartikey]

@_mctrinh connect the pillow mcp to the bed

[Kartikey Saran @CoderKartikey]

IMAGINE ! you are PRESENTING this PPT in the room.

[Kartikey Saran @CoderKartikey]

if you close ur eyes and pick a random profile on the x it gonna be a founder fr...

[Kartikey Saran @CoderKartikey]

@dexhorthy @0xblacklight append-only log as state is the underrated part. a bad run becomes replayable: rewind to the step where it went wrong and fork from there instead of starting over. can you swap models mid-log?

[Kartikey Saran @CoderKartikey]

@jerryjliu0 routing is the right call, no single ocr model wins on every doc type. sharp question: does it route per page? one pdf with a clean text page and a gnarly scanned table page wants two different models

[Kartikey Saran @CoderKartikey]

@dhh the underrated part is different training priors. same model reviewing itself mostly agrees with its own blind spots, a second model actually disagrees. only catch: two models agreeing confidently can still both be wrong, so the tests stay the judge

[Kartikey Saran @CoderKartikey]

@bcherny 10x cheaper is what actually changes behavior. at that price haiku stops being the fallback and becomes the default subagent for grep, summaries and compaction, with opus only called when it's stuck

[Kartikey Saran @CoderKartikey]

@ThePrimeagen screenshot-click accuracy is the number that matters for qa agents. a fast model that misclicks once per run still gives you a flaky suite, so i'd love to see the miss rate over a few hundred runs.

[Kartikey Saran @CoderKartikey]

@jxnlco honestly that's the best kind of demo. an agent that turns a vague 'what's my week' into 10 flights and 3 cars is showing real planning, the part i'd watch is whether it asks before booking anything.

[Kartikey Saran @CoderKartikey]

@svpino derived fields are where i'd want receipts though. if it says yearly cost is $1,188, show me the monthly value and the x12 it used, otherwise one wrong unit assumption looks exactly like a right answer.

[Kartikey Saran @CoderKartikey]

@bendee983 every abstraction lowered the floor without lowering the bar. the gap between "it runs" and "it survives real traffic" is still where the job lives

[Kartikey Saran @CoderKartikey]

@leerob decompilation is the sleeper use case nobody priced in. turning a binary back into readable source used to be years of volunteer work, now it's a loop of model guesses checked against a byte-for-byte match. that verification step is why it works so well.

[Kartikey Saran @CoderKartikey]

@emollick the re-reading part feels closer than the discovery part. an agent that checks every paper's stats against its own code and data would flag half a field's replication problems before anyone runs a new experiment.

[Kartikey Saran @CoderKartikey]

@kentcdodds this is the move. anything the agent does the same way twice should graduate into a script, then the agent only handles the parts that actually need judgment