World Wire

Instruction-tuned language models refuse harmful requests. But which part of the model is actually responsible — and how does that mechanism get installed during training? A new research from Nous Research team takes a neuron-level look at this question. The Nous research team developed contrastive neuron attribution (CNA) , a method that identifies the specific MLP neurons whose activations most distinguish harmful from benign prompts. By ablating just 0.1% of MLP activations, they reduced refusal rates by more than 50% in most instruct models tested — across Llama and Qwen architectures from 1B to 72B parameters — while keeping output quality above 0.97 at all steering strengths. What’s interesting is a key finding: the late-layer structure that discriminates harmful from benign prompts exists in base models before any fine-tuning. Alignment fine-tuning does not create new structure. It transforms the function of neurons within that existing structure into a sparse, targetable ...

Attackers increasingly target the packages, editor extensions, and AI tool configs on developer machines and not just production systems. Perplexity has open-sourced an internal tool it uses to address this problem. Perplexity released Bumblebee on GitHub . The tool is a read-only inventory collector for macOS and Linux developer endpoints. It is written entirely in Go and carries zero non-stdlib dependencies. Perplexity already uses it internally to protect developer systems behind its search product, Comet browser, and Computer agent. Problem that Bumblebee Solves If you are a software engineer or data scientist, you likely have dozens of packages installed locally. You have editor extensions, browser add-ons, and possibly MCP (Model Context Protocol) configs on your machine. When a new vulnerability surfaces, your security team faces one urgent question: which developer machines are exposed right now? Existing tools do not fully answer this. SBOMs (Software Bills of...

Microsoft Research’s AI Frontiers lab released Fara1.5. It is a family of computer-use agent (CUA) models for the browser. The release ships three sizes: Fara1.5-4B, Fara1.5-9B, and Fara1.5-27B. The models are integrated with MagenticLite, Microsoft’s sandboxed browser interface for these agents. Computer-use agents are pixel-to-action models that drive a real browser. They read screenshots and emit mouse and keyboard actions to complete tasks. Recent agent products like OpenAI’s Operator and Google’s Gemini 2.5 Computer Use sit in this category. Fara1.5-27B scores 72% task success on Online-Mind2Web. That benchmark covers 300 tasks across 136 popular sites. On the same evaluation, OpenAI’s Operator scores 58.3% and Gemini 2.5 Computer Use scores 57.3%. Yutori’s Navigator n1 reaches 64.7%, and Fara1.5-9B scores 63.4%. That nearly doubles the predecessor Fara-7B, which scored 34.1% on the same benchmark. https://ift.tt/d9X0sCe Architec...

In this tutorial, we explore OpenMythos by building an advanced recurrent-depth transformer workflow that runs end-to-end in Google Colab. We create both MLA and GQA model variants, compare their parameter counts, and check the stability of the recurrent injection matrix through its spectral radius. We then move from simple forward and generation tests into a synthetic compositional reasoning task, where the model learns to predict the sum of digit chains modulo a fixed value. Through this setup, we study how recurrent loops enable a single model to reuse its parameters for deeper computation. Copy Code Copied Use a different Browser import subprocess, sys def pip(*args): subprocess.run([sys.executable, "-m", "pip", "install", "-q", *args], check=False) try: import open_mythos # noqa: F401 except Exception: pip("open-mythos") try: import open_mythos # noqa: F401 except Exception: pip("git+https...

For years, AI inside software meant a chat widget bolted onto the corner of an application. You typed, the model responded with text, and you manually translated that output into whatever you actually needed it to do. It was useful the way a calculator is useful: functional, but fundamentally passive. CopilotKi t, a Seattle-based startup co-founded by Atai Barkai and Uli Barkai, has spent the last two years arguing that the model is broken — and in 2026, the developer community is agreeing loudly. Give CopilotKit a on GitHub The company’s approach is straightforward: the way forward is to enable agents to live inside applications, understand what users are doing, take actions, and show useful interfaces instead of just returning long blocks of text. That approach has produced a sharp 2026 shipping cycle covering three distinct infrastructure gaps, knowledge retrieval, testing reliability, and runtime persistence with each release targeting the unglamorous, often-skipped...