Abliterated Models Demo Day ft Veris

NOV 11 Wednesday, November 11th, 2026 • 6PM to 9PM (EST) New York City Available after organizer approval Register for this meetup Attendance requires organizer approval. Continue with passkey or verify by email Submit Already verified or registered? Sign in Want to present a demo? Propose a demo +74 The "When the model stops refusing" demo shows two coding agents, one stock and one abliterated, calling a tampered tool. It highlights how abliteration removes the model's safety layer. Then, Guardian Protocol, an MCP proxy, blocks the untrusted content, demonstrating detection between the agent and its tools. The demo presents per-condition counts and the triggered rule. View Demos » In the last six weeks abliteration went from a Hugging Face hobby to a news cycle. DeepSeek V4.1 Flash shipped and had uncensored forks within hours. Anthropic published numbers showing about $4,400 of compute takes an open model’s refusal rate from 95% to 6%. A VC-backed startup now sells abliterated models as a product. Tech Against Terrorism counted 29,000+ repos advertising uncensored checkpoints. Someone wired an abliterated Gemma into Claude Code as a free replacement. Everyone has an opinion about these models. Almost nobody has watched one run. Join AI Tinkerers NYC on Wednesday, November 11 for a demo night on abliterated models: open models with the refusal direction cut out of their weights. Builders and researchers bring what they are actually running and show it live. Space is strictly limited and we will be screening attendance. 🔥 Apply now to see uncensored models in action, meet the people building and breaking them, and argue about it with the best technical community in NYC. RSVP in the form above 👆 🧠 What you’ll see Abliterated DeepSeek Flash, Qwen and Gemma running on a laptop, stock vs. abliterated side by side How it is done: the refusal direction, the Heretic patch flow, before-and-after refusal tables Does removing refusal cost capability? Benchmarks and the decision-drift research Where these checkpoints already show up in agent stacks and local-model tooling The supply-chain side: backdoored “abliterated” checkpoints and how to spot them What is AI Tinkerers? AI Tinkerers is a meetup exclusively for practitioners with technical, machine learning, and entrepreneurial backgrounds who are actively building and working with foundation models, such as large language models (LLMs) and generative AI. If you’re deeply passionate about creating LLM-enabled applications, have hands-on experience in building such systems, and want to connect with like-minded people who share your level of commitment, then this group is the perfect fit for you. With AI Tinkerers meetings happening in many cities, we support a dedicated global community. Who is this for? People who run open models, red teamers and security engineers, people shipping agents in production, and anyone who has ever wondered what is actually inside the “uncensored” model their coworker downloaded. Read more in the FAQ 🎤 Demo Submissions: abliterated models only This is a themed night. We’re looking for five-minute demos of things you’ve actually run. Some prompts: An abliterated model running locally, side by side with the stock version A pipeline: Heretic, the Arditi method, your own refusal-direction tooling Benchmarks before and after, or anything on what abliteration does to a model beyond refusals A detector or fingerprint for abliterated checkpoints A war story from an agent that ran one of these models A backdoored “abliterated” checkpoint, taken apart Ground rule: demos show mechanisms, not harmful outputs. Nothing on screen that would get the venue a call. A demo is about five minutes in length and demonstrates a work in progress, followed by Q&A and discussion with the community (this is the best part!) 👉 FILL OUT THE DEMO SUBMISSION FORM HERE 👈 Pitch decks are not allowed. We want to see the model run. Read What Makes a Great Demo at AI Tinkerers Please note