Prashanth Pappu

Founder. Coder. Producteer.

[Prashanth Pappu @prashanth_pappu]

this day that age

[Tamal Saha @Tamal0401]

A military conference to deliberate as how to hold “free and fair” elections in #Bengal

Unbelievable - Chiefs of BSF, CRPF, SSB, ITBP & CISF - all came together in #Kolkata to discuss one state ‘election’?

Is this #BengalElection2026 or Operation Sindoor? Or operation Bengal?

[Prashanth Pappu @prashanth_pappu]

😂

[JNS @_devJNS]

me and Claude building an app at 3AM

[Prashanth Pappu @prashanth_pappu]

We used programs to synthesize the best models. Now we will use models to synthesize the best programs.

(I.e., not just generate program code but mutate, regenerate, score and optimize the best variant.)

This has huge implications for regulated industries with systematic billing errors, underpaid royalties, and wrong classifications. The ground truth is sitting in public regulatory records. These problems require joins, hierarchies, and counterfactuals. That's not a model synthesis problem. It's a program synthesis problem.

These industries will now see the best audit programs that can detect every incorrect application of their regulations.

[Prashanth Pappu @prashanth_pappu]

Everyone's afraid of AI that lies. The real threat? AI that agrees.

AI sycophancy will cost knowledge workers more than AI hallucinations ever will.

(Devs shipping fatal bugs they "confirmed" with AI. Sales sending pricing proposals "locked in" by AI.)

[Prashanth Pappu @prashanth_pappu]

*any leadership*

[Steve Huynh @ALEngineered]

A surprising amount of “tech leadership” is just being the adult in the room when everyone wants to work on shit that doesn’t matter.

[Prashanth Pappu @prashanth_pappu]

💩 take

[X Freeze @XFreeze]

Elon Musk predicts that AI will bypass coding entirely by the end of 2026 - just creates the binary directly

AI can create a much more efficient binary than can be done by any compiler

So just say, "Create optimized binary for this particular outcome," and you actually bypass even traditional coding

Current: Code → Compiler → Binary → Execute
Future: Prompt → AI-generated Binary → Execute

Grok Code is going to be state-of-the-art in 2–3 months

Software development is about to fundamentally change

[Prashanth Pappu @prashanth_pappu]

Lol

[Neil Zeghidour @neilzegh]

Me defending my O(n^3) solution to the coding interviewer.

[Prashanth Pappu @prashanth_pappu]

Contextual Expertise maybe closer to what’s being described.

1. It’s raw business hints: “if you are checking vendor certificates during onboarding note that civil contractors must submit both health and environmental certificates”.

2. Not sure if it is necessary to structure this data (into graphs!)

But,

3. There is value in capturing this expertise from real enterprise users.

4. There are elegant ways to condense and retrieve this expertise in the right context for AI actions.

[Prashanth Pappu @prashanth_pappu]

AI's big category error:

AI is not ready for autonomous processes.
AI is ready for collaborative processes.

[Prashanth Pappu @prashanth_pappu]

@gdb TypeScript is better.

Agents need to be often composed with complex transformations and TypeScript captures dependent types far better.

[Prashanth Pappu @prashanth_pappu]

This.

[Krishnan @cvkrishnan]

They all have one thing in common.

[Prashanth Pappu @prashanth_pappu]

@gdb Two big themes for enterprise agents in 2026 will be agent+human collaborative UIs and continual learning.

[Prashanth Pappu @prashanth_pappu]

@DimitrisPapail Very good Continual Learning can be achieved with input aware context engineering that incorporates user feedback. This will reach production scale in 2026.

[Prashanth Pappu @prashanth_pappu]

AI to code to X >> AI to X

[Prashanth Pappu @prashanth_pappu]

Functional frameworks are more amenable to intent encapsulation and rich compile time verification. For AI, that's transformative because code becomes a verifiable intermediate step for generating anything.

For example, to generate a contract or music track from multi-page instructions, you have a few options:

Direct generation — tuned foundational model outputs the artifact (uncontrollable & unreliable)

Agentic steering — orchestration layer guides the model (too manual & restrictive)

Intermediate code — model generates high level code that renders the output (more controllable, still unreliable)

Intermediate functional code — model generates code in a functional framework with compile-time verification (controllable & provably reliable)

[Prashanth Pappu @prashanth_pappu]

AI also enables a new type of specification that's very non-programmer friendly. Call this new specification language "formalish" i.e., english with optional formalisms that itself is composable and reusable. PMs and users can write and review formalish specs with AI support with a very shallow learning curve.

[Prashanth Pappu @prashanth_pappu]

@trssnd For real enterprise use-cases. Anywhere you need a "customization layer" on top of a generic software platform.

[Prashanth Pappu @prashanth_pappu]

Formalish: a new programming language for non-programmers.

- Formal + English
- Composable, Sharable, Reusable
- 100% Verifiable AI Code Gen

[Prashanth Pappu @prashanth_pappu]

Investors: AGI!

CEO: User talks, AI executes.

CTO: User talks, AI auto builds an agentic workflow and then executes it.

VP: User talks, AI auto builds an agentic workflow with human-in-the-loop and then executes it.

Product: User talks, AI presents a visual editor to co build an agentic workflow with a human developer.

<Developer sees one visual node connector with nested JSON schema fields to be carefully updated manually>

Engineering: Here are a few carefully engineered, tuned and secured agentic workflows with human-in-the-loop with select customer configurable parameters.

Investors: Vertical AI!

[Prashanth Pappu @prashanth_pappu]

MCP+UI is the new HTTP+HTML.