Swayatta Daw

AI Researcher - Cisco
🎓 MS by Research - IIIT Hyderabad

[Swayatta Daw @hesozakmi]

@ProfTomYeh Awesome, how about the modern variants - like sparsemax/ entmax?

[Swayatta Daw @hesozakmi]

@vaibcode Faced something similar at our team - it gets worse for monorepos when multiple devs are vibecoding prod stuff in the same repo!

[Swayatta Daw @hesozakmi]

@DevanshuXi Hey, was exploring something similar - there is a startup @deeptuneai who is building this

[Swayatta Daw @hesozakmi]

@willchen500 Can it talk to any legal knowledge base and do retrieval/search?

[Swayatta Daw @hesozakmi]

@A_K_Nain What are you writing kernels for though?

[Swayatta Daw @hesozakmi]

@yogesharma92 Well am from Kolkata, so you can only imagine how cheap it is

[Swayatta Daw @hesozakmi]

@willchen500 Assume you are using some external data to feed context, right?

[Swayatta Daw @hesozakmi]

@willchen500 Hey, just wondering - can we build an equaivalent of this for any India? Our (Indian) legal system is slow and hidden behind tons of paperwork - where did you get the data from

[Swayatta Daw @hesozakmi]

@A_K_Nain Yep, and this is same old self / weak supervision concept - the quality would highly depend on the grounding data , as well as keeping number of iterations low, with HIL

[Swayatta Daw @hesozakmi]

@rsnkyx Great article, came across this one for evals by @AnthropicAI : https://t.co/YdTosO0Xgj

[Swayatta Daw @hesozakmi]

@_avichawla What I have noticed is that spawning subagents often incur an overhead of latency - for systems where we need to provide result in say under ~30s, subagents can cause a significant hit in latency

[Swayatta Daw @hesozakmi]

@Pseudo_Sid26 Just went through the paper : https://t.co/myr3kFhgpO - the authors claim something very interesting at the end, XSA can basically be thought of as an implicit attention sink, where the undesired attention comes into self projection and gets removed!

[Swayatta Daw @hesozakmi]

@ndrewpignanelli Agree with this take. Only place where KGs make sense is when grep + embeddings is not enough

[Swayatta Daw @hesozakmi]

@NielsRogge Curious to know how does these top ranked harnesses have so less traction? Have not heard of any of these apart from droid

[Swayatta Daw @hesozakmi]

@AlexFinn Thanks for sharing, am going to try it tomorrow. Pretty hard to dethrone Claude Code imo

[Ujjwal Chadha @ujjwalscript]

"English is the new programming language" is largely BS.

Everyone on X timeline is saying: "Just vibe code it! You don't need to learn syntax anymore, just prompt the AI in plain English!"

HOT TAKE from the trenches of reviewing pull requests all day: English is the worst programming language ever invented.
It is ambiguous. It is emotional. It lacks strict constraints.

Code isn't hard because of the brackets or the syntax. Code is hard because of the precision. When you write in Python or React, you are forced to define the exact boundaries of reality. You have to explicitly handle the edge cases, the null states, and the architecture.

When you write in "English", you are just hoping a probabilistic engine guesses your intent correctly.

I am watching people use AI to "vibe code" entire backends in a weekend. The result? It functionally works for a demo, but it's an absolute disaster under the hood.

No scalability, zero security considerations, and multiple responsibilities crammed into single, unmaintainable components.

We aren't building the future 10x faster. We are just generating legacy spaghetti code 10x faster.

The engineers getting promoted on my team right now aren't the best "prompt whisperers." They are the ones who know exactly why the AI's "English-to-Code" translation just introduced a silent memory leak into the system design.

Stop learning how to "chat" with a bot. Start learning how to architect systems.
Ambiguity is the enemy of scale.

[Arpit Bhayani @arpit_bhayani]

With agentic slop, we are trading software reliability for shipping velocity and calling it progress. It isn't. Systems are more fragile than ever, and engineers building them no longer trust their code to hold up in real-world edge cases.

I am pro-AI, but this will backfire - big time.

[Swayatta Daw @hesozakmi]

@pbteja1998 @AndrewWarner What sort of subscriptions are you taking, is claude pro enough?

[Swayatta Daw @hesozakmi]

@HustlerzHub @grok who is he

[Scott Adams @ScottAdamsSays]

A Final Message From Scott Adams