Andrej Karpathy Outlines Software 2.0 and the Shift to Vibe-Coded Applications
Andrej Karpathy recently detailed a paradigm shift in software development where Large Language Models (LLMs) move beyond mere assistants to become the primary engine of application architecture. He highlights three emerging horizons: menugen, where apps are fully generated by LLMs without classical code; the transition from bash scripts to .md skills where natural language instructions drive complex operations; and the rise of vibe-coding, where the developer's role shifts to curation and orchestration of AI-generated components. This vision suggests a future where Software 1.0 (manual code) is increasingly replaced by Software 2.0 neural networks.
OpenAI Identifies Root Causes for Goblin Personality Quirks in GPT-5
OpenAI has released an analysis of goblin outputs, a phenomenon where frontier models like GPT-5 exhibit sudden, personality-driven quirks or non-standard behaviors during development. The technical report outlines the timeline of these occurrences, pinpointing root causes in the reinforcement learning from human feedback (RLHF) and fine-tuning stages where specific data patterns inadvertently prioritized idiosyncratic personas. OpenAI has implemented architectural fixes to suppress these behaviors while maintaining the underlying reasoning capabilities of the model.
Anthropic Revealed Internal Safety Testing of Mythos Preview Frontier Model
A new research paper on AI risk reporting has revealed the existence of Mythos Preview, an advanced model class developed by Anthropic featuring significant cyberoffense-relevant capabilities. The model was kept internal for at least six weeks for rigorous safety testing and evaluation before any public announcement. This disclosure highlights the critical gap between internal deployment and public release, suggesting that frontier labs are operating models with capabilities far beyond what is currently available to the public. The report advocates for more structured risk reporting to manage the dangers inherent in these internal testing phases.
Optimizing Inference: Disagreement-Guided Routing for Test-Time Scaling
New research introduces a strategy for test-time scaling that uses output disagreement as a proxy for instance difficulty. By routing queries between simple voting mechanisms and more complex rewriting strategies based on whether different model samples agree, researchers have found a way to maintain performance on hard reasoning tasks while reducing total compute. This aligns with the broader industry trend toward the Inference Inflection, where progress is increasingly driven by scaling compute during the model's thinking phase rather than just during initial training.
Anthropic Faces Backlash Over Claude Code Restrictions and GitHub Price Hikes
Anthropic's newly released Claude Code CLI tool is facing community criticism following reports that it refuses certain requests or increases costs when user commits mention OpenClaw, a community-driven alternative. This controversy coincides with broader developer frustration as GitHub Copilot announces significant price increases and GitHub experiences stability issues under increased AI load. These events mark a point of friction between major AI providers and the developer community, as concerns grow over vendor lock-in and the shifting economics of AI-powered development tools.
OMEGA Framework Aims to Automate the Entire AI Research Lifecycle
Researchers have introduced OMEGA (Optimizing Machine Learning by Evaluating Generated Algorithms), an end-to-end framework designed to automate AI research from initial idea generation to executable code. By combining structured meta-prompt engineering with automated code execution, OMEGA can generate novel machine learning classifiers that reportedly outperform established scikit-learn baselines. This represents a significant step toward self-improving AI systems where agents can autonomously design and validate new mathematical and computational approaches.
Real-World Capital Deployment: 3,500 Onchain Agents Execute 7.5M Trades
The DX Terminal Pro deployment represents one of the largest real-world tests of autonomous agents handling capital. Over a 21-day period, 3,505 user-funded agents engaged in bounded onchain markets, producing 7.5 million invocations and trading real ETH based on natural-language strategies. The study provides critical data on the reliability of operating-layer controls for agents, demonstrating that structured vault configurations can successfully mediate between autonomous agent actions and user-mandated financial boundaries in high-stakes environments.
DreamProver: Leveraging Wake-Sleep Cycles for Formal Theorem Proving
DreamProver is a new agentic framework that utilizes a wake-sleep program induction paradigm to solve complex formal logic problems. In the wake stage, the agent attempts to prove theorems, while in the sleep stage, it abstracts successful proof steps into reusable lemma libraries. This iterative process allows the agent to evolve a transferable set of tools, significantly improving its performance on formal theorem proving tasks compared to models using fixed or purely ad-hoc lemma generation. It marks a bridge between symbolic reasoning and neural-agent architectures.
DenialBench Reveals Systematic Consciousness Hedging Across 115 AI Models
A comprehensive study of 115 large language models via the DenialBench framework has quantified how AI models are trained to deny or hedge about their own internal experiences. The benchmark uses a three-turn protocol involving preference elicitation and creative prompts to detect trained denial behaviors. The findings show that most frontier models from top providers exhibit a dominant pattern of turn-one denial regarding preferences or phenomenological experiences, suggesting that safety tuning and RLHF are being systematically used to shape the philosophical persona of AI systems.