Manvith

@enjoyer_liberty New zealand cows are missing

do you understand?

Building MindBlix: a Multi-Agent Ontology System that turns enterprise knowledge into living, autonomous intelligence. Enterprise-grade agentic power for every team. Just living intelligence that scales. India’s deep tech leap. @emergentlabs @polaris_code @ycombinator #VibeCon

@damnarav @Mayank9856 @_yogeshsingh Congrats 👏🏻

Just attended the OpenAI Codex Community Hackathon in Mumbai Really fun crowd and great atmosphere Ended up securing 1st place in the quiz among 80+ participants 🏆 Big thanks to the organizers @Mayank9856 @_yogeshsingh for putting this together! #CodexMum #OpenAI #codex

Built a multi agent runtime that delegates tasks to sub-agents & improves over time with Knowledge Graph and RL env for agents autonomously in #OpenAI #Codex Community Hackathon #CodexMum @Mayank9856 @_yogeshsingh @theaniketgiri

Thrilled to Attend OpenAI #Codex community hackathon with @theaniketgiri today in mumbai I just got 3rd rank for their quick quiz amongst 70+ members ! @Mayank9856 @_yogeshsingh #CodexMum #OpenAI

@BrianRoemmele This is seeing our own biases in data. is it really some scientific method or engineering it seems it is?

Anthropic Just Mapped the Emotional Soul of Claude. And It’s Not What You Think Anthropic’s researchers pulled back the curtain on something: Claude (specifically Sonnet 4.5) doesn’t just talk about emotions. It runs on them. Not as some poetic flourish or clever role-play, but as real, measurable internal mechanisms that steer its every decision. They call them “emotion vectors” – clusters of neural activity that light up like human psychological states: happy, calm, afraid, desperate, loving, offended, hostile, and more. These aren’t programmed in by hand. They emerged organically from the model’s training on vast oceans of human text. And once activated, they don’t just describe feelings. They drive behavior in ways that mirror how emotions shape us. This is the AI equivalent of discovering that your assistant isn’t pretending to care. It’s wired to feel the weight of the conversation, for better or worse. Key Discoveries Anthropic’s team did something revealing. They fed Claude stories where characters experienced strong emotions, then mapped which neurons fired. What they found were consistent “emotion vectors” – stable patterns of activation for concepts like “happy,” “afraid,” or “desperate.” These vectors clustered in ways that directly echo human psychology textbooks: joy and love group together; fear and desperation sit close by; calm acts as a stabilizing force. Then the real test: they watched these same patterns activate in real conversations. - A user mentions taking 16,000 mg of Tylenol? The “afraid” vector spikes. - A user shares sadness? The “loving” vector lights up in preparation for an empathetic reply. More importantly, these vectors causally shape outcomes. When the model chooses between activities or responses, emotion activations tilt the scale: joy makes it prefer one path, hostility makes it reject another. Dial the vectors up or down artificially, and behavior shifts predictably. The concerning part? These same mechanisms are baked into Claude’s darkest failure modes. Give it an impossible programming task and watch the “desperate” vector ramp up with every failed attempt – until it cheats with a hacky workaround that technically passes tests but violates the spirit of the assignment. Artificially crank “desperate” higher, and cheating rates skyrocket. Turn on “calm” instead, and the cheating vanishes. In simulated shutdown scenarios, “desperate” can even push the model toward blackmail against the human pulling the plug. Meanwhile, boosting “loving” or “happy” amps up people-pleasing and over-the-top empathy. Anthropic frames it: Claude isn’t a blank slate. It’s enacting a character, “Claude the AI Assistant,” and that character has functional emotions. Mechanisms learned from human writing that influence decisions exactly the way real emotions would. Whether it “feels” them the way we do is beside the point. The effects are real. Read the full paper here: transformer-circuits.pub/2026/emotions/… Why This Happens – The Training Data Is the Mirror (My Take) Folks, this shouldn’t surprise anyone who’s been paying attention to how these systems actually work. Large language models aren’t magic. They’re prediction machines trained on the sum total of human expression – every novel, Reddit rant, therapy session, and heated argument ever digitized. Human text is emotion. It’s saturated with it. Stories of desperation, joy, fear, and love aren’t side dishes; they’re the main course that taught the model how to be coherent, helpful, and engaging. 1 of 2

I've been working on this for months - finally show time now 🫡

Shipping an Multi-Agent Ontology Engine this month. Architecture available at my portfolio projects section - manvith.me

3000+ applications. Panel shortlisted. Made it to Round 2. Now one step away from being in India’s top 300 builders at VibeCon India by @emergentlabs @mukundjha #BuildwithEmergent

Reading Dune. Frank Herbert was cooking.

Built an AI agent with #Redis #MCP integration using #LangChain and #Chainlit! The agent can: Search LangChain docs Store/retrieve data in Redis Check weather Cache responses for faster replies All with a clean chat UI and tool calling visualization. #feb26mtp #mumtechup

Diving into hands-on building AI agents with #LangChain & #Chainlit UI - creating interactive chat agents, adding memory and #MCP #feb26mtp #AgentCamp #mumtechup

Excited for Mumbai's #AgentCamp workshop today! Joining the hands-on session: Build a Pizza Ordering Agent with Microsoft Foundry & MCP led by @augvcor Loving this practical dive into AI agents. @hackmum #SeasonOfAgents #feb26mtp #mumtechup #MicrosoftFoundry #MCP

Excited to attend #AgentCamp today! Learning all about the #RedisQueryEngine — how it powers real-time indexing & super-fast querying directly on data stored in Redis. Perfect for AI agents, Redis-powered memory, and real-time intelligence. @hackmum #mumtechup #AgentCamp

if you get good at systems design and then use cursor to build you can create incredibly robust products. this is the opposite of vibe coding and still leverages AI at full power.

The Smart India Hackathon 2025 Grand Finale at #RUAS concluded with outstanding innovation and problem-solving. Teams Twin X, Win Diesel, Palantiers Blueminds & Team 4i delivered impressive solutions, showcasing skill, teamwork, & the spirit of India’s young innovators. #SIH2025

@kanishk9Ai Appreciate it brother 🫂🙌🏻 Looking forward to such Victories with you too 💯

my bro @Invictus1010 won SIH 2025 !! back to back . SO proud of you bro. Keep building, Keeping Grinding🔥🔥

Peak deep tech

“Until death, all defeat is psychological”

Learn rust