This article explores whether an open-weight LLM (GPT-OSS-120B) can adapt a trained PPO control policy to changed dynamics by directly editing its weights. The experiment tests exact repair and adaptation scenarios in …
Researchers at MIT CSAIL demonstrate that frontier LLMs like GPT-6 Astra can natively read and write neural network weights without gradient descent, identifying trained behaviors from raw weights and writing …
An exploration of using GPT-OSS-120B to directly edit the weights of a trained LunarLander policy in response to physics changes, without gradient updates. The experiment successfully repaired reversed thrusters and …
A deep dive into the architecture and engineering principles behind always-on, persistent AI agents that can run autonomously for extended periods, covering context management, durable workflows, sub-agent orchestration, …
TypeSafe AI introduces System One Models and Jev, a new class of frontier AI models optimized for fast, structured, type-safe decisions rather than text generation. Jev achieves comparable intelligence to existing LLMs …
This paper presents a validation-centric workflow for AI-assisted GPU porting of CReSS, a 250,000+ line legacy Fortran weather simulation code, achieving a 5.1× speedup while preserving scientific validity across 162 …
ActiveGraph introduces an event-sourced runtime where an append-only log serves as the primary source of truth for AI agents, with graph state as a deterministic projection of that log. This architecture enables …
EngramaBench is a new benchmark for evaluating long-term conversational memory across five query types including cross-space reasoning, temporal inference, and adversarial abstention. A controlled comparison of GPT-4o …
FuzzingBrain V2 is a multi-agent LLM system that combines semantic code analysis with fuzzing-based verification to automatically discover and reproduce software vulnerabilities. The system achieved 90% detection rate on …
Andrej Karpathy discusses the transition from traditional programming to AI-assisted development, exploring concepts like ‘vibe coding’ and ‘agentic engineering’ while examining how LLMs represent …
