Ekjot Singh

Sneak peek to our website :)

@NayrhitB @SusquehannaVC @LightspeedIndia @BCapitalGroup @beenextVC @sparrowcapvc @2point2club deck

The exact pitch deck that helped us raise a $9M Seed Round copy whatever you want VCs that invested: → @SusquehannaVC (led) → @LightspeedIndia → @BCapitalGroup → Seaborne Capital → @beenextVC → @sparrowcapvc → @2point2club joined. fundraising is hard enough without guessing what investors want to see. so - I'm making our deck public. if you're raising right now, take it and make it yours. Reply 'deck' + follow (so I can DM it over)

M3 Ultra Mac Studio Al Cluster Uses VGD 1.0 with RDMA over Thunderbolt 5 (80Gbps) for 2TB unified memory @ 3.2TB/s. Runs Kimi K2 Thinking (1T parameters) at 28 tok/sec with MLX tensor parallel. Up to 1.8x TPS w/ 2 Macs, 3.2x w/ 4 Macs. VGD 1.0 is now LIVE.

🎬 Spotify Wrapped for Devs. GitWrapped 2025: Your commits, cinematic. ✨ GitHub & GitLab stats 🧬 Find your "Archetype" 🖼️ Get your Movie Poster See your year in code: 👇 gitwrapped.ekjot.me Built by @ek10sh #GitWrapped #GitHub #YearInCode #DevCommunity

@Cipher_twt Hi @cipher_twt

Twitter is cool. But it’s 100x better when you connect with people who code If you’re into tech, AI, DSA, Web development, Web3 or programming, say hi

You cannot patch intelligence. You must architect it. Full breakdown and code released below. 👇 📄 Post: metanthropic.vercel.app/research/fragi… 🌐 Project: metanthropic.github.io/TFGCJCSSLLMs-w… 💻 Code: github.com/metanthropic/T… #AGI #Metanthropic

🧠 The "Doubt Switch" Using Sparse Autoencoders, we saw safety vectors attacking latent physics circuits. Current alignment techniques act as noise, jamming the model's ability to verify reality. Safety cannot be a patch.

🚨 Does "safety" break reasoning? We found that aggressive guardrails trigger "Cognitive Jamming" in LLMs—shattering internal world-models instead of aligning them. Paper: metanthropic.vercel.app/papers/fragili… #AISafety #MechanisticInterpretability

🧠 The "Doubt Switch" Using Sparse Autoencoders, we saw safety vectors attacking latent physics circuits. Current alignment techniques act as noise, jamming the model's ability to verify reality. Safety cannot be a patch.

💥 Repetition Collapse Over-steer a model, and it doesn't just refuse. It crashes. We identified a phase transition where valid reasoning dissolves into infinite loops ("I cannot I cannot..."). The guardrail becomes a jammer.

Does "safety" break reasoning? We found that aggressive guardrails trigger "Cognitive Jamming" in LLMs—shattering internal world-models instead of aligning them. Paper: metanthropic.vercel.app/papers/fragili… #AISafety #MechanisticInterpretability

🔗 Try it yourself: 🌐 Project: metanthropic.vercel.app/research/datas… 📄 Paper: metanthropic.vercel.app/papers/dataset… 💻 Code: github.com/metanthropics/… Work by @ek10sh & Metanthropic. 5/5 🧵

The Visuals Text:🎨 Platonic Representations:Our method acts as a diagnostic tool, revealing what models actually "see." DINO-distilled images focus on structure; CLIP focuses on semantics. These dream-like images capture the texture and shape needed for classification. 4/5 🧵

🚀 Announcing: Dataset Distillation for the Pre-Training Era! Standard distillation trains from scratch. We built a method for the Foundation Model age. Introducing Linear Gradient Matching by @ek10sh & Metanthropics. Can 1 synthetic image beat real data? 🧵👇 #AI #CV #NeurIPS

Sneak peek to our website :)

Trust is not a vibe. It’s an architecture. The facility is officially OPEN. Audit our research. See what we’ve built. Enter here: metanthropic.vercel.app #AGI #BuildPublic #MechInterp #AI

The limit of AI is no longer compute. It is coherence. Today, we open the doors to Metanthropic. We aren't just building another wrapper. We are solving the physics of intelligence. ⚛️⚡ A thread on the end of the "Black Box" era. 🧵👇