Engineering
The Fast Gemma Challenge: our verified-SOTA recipe, in full
The VIDRAFT team documented its winning entry to the Fast Gemma Challenge, reaching 510.58 tokens per second at a perplexity…
Engineering
Discovering Cryptographic Weaknesses with Claude
Anthropic researchers ran an internal model (referred to as “Claude Mythos”) for roughly 60 hours of iterative, human-guided prompting to…
Engineering
Leanstral 1.5: Proof Abundance for All
Mistral AI released Leanstral 1.5, a 6-billion-parameter model for formal proof engineering in Lean 4, under an Apache-2.0 license. It…
Engineering
Bringing More Control Over Your Connectors
Mistral AI detailed new administrative and API-level controls for its agent “connectors” (tool/data integrations), aimed at production deployment concerns. The…
Engineering
Brain2Qwerty v2: From Brain Waves to Words
Meta AI described Brain2Qwerty v2, a system that decodes typed text directly from non-invasive magnetoencephalography (MEG) brain recordings using end-to-end…
Engineering
AISPA: User-Centric System Prompt Auditing for Large Language Model Applications
This paper proposes AISPA, a framework for auditing the hidden system prompts that govern commercial LLM products, which are rarely…
Engineering
smevals Introduces a Standard Vocabulary for LLM Evaluation Pipelines
Simon Willison describes smevals, an evaluation framework built with Prime Radiant that separates running tasks against model variants from grading…
Engineering
How Meta’s SAM and DINO Models Give a Pittsburgh Assistive-Robotics Platform Real-Time Awareness
The University of Pittsburgh’s RAMMP project integrates Meta’s open-source DINO and Segment Anything Model (SAM) to let a robotic mobility…
Engineering
Agentic Context Management Treats AI Agent Memory as a Lifecycle Problem, Not Just Storage
A new paper proposes Agentic Context Management (ACM), reframing agent memory as a lifecycle and architecture problem rather than a…
Engineering
A Scaling Study Finds Plain BM25 Beats Fancier RAG Retrieval Once Corpora Get Large Enough
Researchers ran a controlled scaling study comparing four retrieval-augmented generation paradigms across 28 corpus sizes ranging from 1.7 million to…
Engineering
Microsoft’s Echoverse Trains Computer-Use Agents in Synthetic Apps That Evolve With Every Graded Run
Microsoft researchers built Echoverse, a system that generates stateful synthetic applications for training computer-use agents through a co-evolution loop in…
Engineering
TIL: Adding a Custom MCP Server to Claude and ChatGPT
Simon Willison published a technical guide explaining how to connect custom MCP (Model Context Protocol) servers to Claude and ChatGPT’s…