Get live statistics and analysis of Vishal Verma's profile on X / Twitter

My Take on GEPA (Genetic-Pareto) : prompt optimization doesn’t need a judge , but if you want reflexive, human-aligned improvement, you’ll want one. GEPA evolves prompts through mutation and selection, but evolution only works if you can score the outputs. Without a good scoring signal, GEPA is just blind mutation. Some say: “We already have human-labeled data.” But those labels were created for one prompt. The moment you mutate the prompt, that mapping breaks, your human annotations don’t automatically apply to new outputs. That’s where a judge comes in. A well-aligned LLM-as-judge can generalize human intent across unseen outputs. When your human annotations include why something was rated a certain way (“too verbose,” “off-topic,” “incorrect reasoning”), that reasoning stays valid, even for new prompts. You can train that judge using human explanations and let GEPA optimize against it. This gives you reflexive feedback: a closed loop where evolution aligns to human preference automatically. Don’t want to build or align a judge? Then use a code-based check instead- especially if your quality metrics are measurable (factuality, format, correctness, etc.). But if you’re using an LLM to compare outputs against human expectations… surprise, you already built a judge. The real takeaway: Your human-labeled data didn’t go stale, your evaluation pipeline did. Teach your system how to judge, not just what to score. That’s how GEPA achieves real adaptive intelligence.

Once upon a time, emails had personality. You could hear the person , the abrupt ones, the overthinkers, the ones who forgot punctuation existed. Then came ChatGPT, and now every email sounds like it was proofread by a butler with a thesaurus.

Just spent 15 mins debugging why my code “was wrong.” Turns out Codex refuses to believe in __file__ It auto-corrected double underscores like they were a typo. Bro. That’s literally how Python works. 💀

We've been sending LLMs bloated data and wondering why our token bills are insane. Enter TOON: JSON, but designed for how LLMs actually work. { "tags": ["pen", "ink", "pencil"] } → tags[3]: pen,ink,pencil 40–60% token reduction. No information lost. We optimized the models. We optimized the prompts. Why didn't we optimize the data format?

Early customers are brutal teachers 😬. What’s the best lesson you learned from yours? Asking for a friend… okay, it’s me.

Before you build your next LLM: • Do you understand gradient clipping? • Can you explain layernorm math? • Have you tried curriculum learning? • Do you know your tokenizer quirks? • Can you debug loss spikes? • Have you checked for data leakage? • Can you detect attention collapse? • Do you understand learning rate schedules? • Have you profiled your data pipeline? • Can you interpret your validation metrics (beyond accuracy)? • Do you test for prompt bias + toxicity drift? • Can you spot mode collapse early? • Have you validated your eval set quality? • Can you read an optimizer config and know why it works? • Have you tested inference scaling (KV cache, quantization, batching)? • Can you explain how positional encodings affect long context? • Have you measured compute vs. data scaling trade-offs? • Do you know what your loss actually represents? If not, start there.

Seen this story before - my ex-colleague did a PhD in Physics, got married, left academia for an RL job because physics wasn’t paying, and now he’s obsessed with RL.

This is fascinating. Model auditing via output traces might end up as important as watermarking

Love seeing AI move beyond optimization into creativity. Curious how you defined and measured “novelty” vs “realism” in practice , that’s such a subtle frontier in creative domains like chess.

Aim to name your startup so well that when you’re acquired, they change their name.

Spent the weekend reading new survey on "Vibe Coding" where they formally modeled what most of us have been doing for months, generating code with an agent, running it, and shipping if it works. What we called “vibe coding” is now an officially described engineering paradigm. They analysed over 1,000 papers and built a Constrained Markov Decision Process to represent the triadic relationship between: > developer intent (what and why) > codebase context (where) > agent decisions (how) That framing changes everything. It formalizes what we intuitively knew: the locus of engineering has shifted from code writing to agent orchestration. One day we were writing functions; the next, we were validating model outputs and still calling it “coding.” We’re not, at least not in the old sense. Here’s the part that really unsettling : Automation bias is killing velocity. We trust the agent’s clean syntax, ship it, and when things break, we blame “hallucination.” But the truth is: we treated an autonomous system like a smarter Stack Overflow. Performance depends on: >the quality of feedback loops (compiler errors, runtime traces, human intuition) >the infrastructure surrounding the model (sandboxed execution, CI/CD integration, orchestration tools) >the structure of collaboration between human and agent The tools are fine. Our workflows aren’t.

This is spot on. I've noticed this too !! Many juniors treat AI like a vending machine for code

Once upon a time, emails had personality. You could hear the person , the abrupt ones, the overthinkers, the ones who forgot punctuation existed. Then came ChatGPT, and now every email sounds like it was proofread by a butler with a thesaurus.

This is spot on. I've noticed this too !! Many juniors treat AI like a vending machine for code

Just spent 15 mins debugging why my code “was wrong.” Turns out Codex refuses to believe in __file__ It auto-corrected double underscores like they were a typo. Bro. That’s literally how Python works. 💀

My Take on GEPA (Genetic-Pareto) : prompt optimization doesn’t need a judge , but if you want reflexive, human-aligned improvement, you’ll want one. GEPA evolves prompts through mutation and selection, but evolution only works if you can score the outputs. Without a good scoring signal, GEPA is just blind mutation. Some say: “We already have human-labeled data.” But those labels were created for one prompt. The moment you mutate the prompt, that mapping breaks, your human annotations don’t automatically apply to new outputs. That’s where a judge comes in. A well-aligned LLM-as-judge can generalize human intent across unseen outputs. When your human annotations include why something was rated a certain way (“too verbose,” “off-topic,” “incorrect reasoning”), that reasoning stays valid, even for new prompts. You can train that judge using human explanations and let GEPA optimize against it. This gives you reflexive feedback: a closed loop where evolution aligns to human preference automatically. Don’t want to build or align a judge? Then use a code-based check instead- especially if your quality metrics are measurable (factuality, format, correctness, etc.). But if you’re using an LLM to compare outputs against human expectations… surprise, you already built a judge. The real takeaway: Your human-labeled data didn’t go stale, your evaluation pipeline did. Teach your system how to judge, not just what to score. That’s how GEPA achieves real adaptive intelligence.

We've been sending LLMs bloated data and wondering why our token bills are insane. Enter TOON: JSON, but designed for how LLMs actually work. { "tags": ["pen", "ink", "pencil"] } → tags[3]: pen,ink,pencil 40–60% token reduction. No information lost. We optimized the models. We optimized the prompts. Why didn't we optimize the data format?

Early customers are brutal teachers 😬. What’s the best lesson you learned from yours? Asking for a friend… okay, it’s me.