A developer argues that perceived model degradation after updates is almost always caused by changes to the wrapper layer—the router, system prompt, reasoning effort, and verbosity caps—rather than th…
A developer proposes replacing LLM API calls with coding agent calls to reduce costs, citing that coding assistants like Codex, Cursor Pro, and GitHub Copilot Pro cost significantly less per million t…
Current large language models like GPT-5 and Claude Opus-4.1 achieve only a 40% success rate on the PROBE benchmark for proactive AI agents, revealing significant limitations in autonomous problem-sol…
Deployment simulation, using de-identified past conversations to test AI models before release, is emerging as a more realistic safety evaluation method. A study of GPT-5-series deployments showed it …
Andy Masley argues that debates over AI's environmental impact are often rooted in normative disagreements, not factual ones, and that political liberalism offers a better framework than environmental…
The EU AI Act Omnibus deferred Annex III credit scoring compliance to December 2027, but GPAI transparency enforcement goes live August 2, 2026. Banks using foundation models in production workflows f…
Researchers developed AegisDx, a safety-oriented AI framework for differential diagnosis that coordinates specialized LLM components to generate broad diagnoses, screen for dangerous conditions, and v…
Anthropic launched a reflection dashboard for its Claude AI assistant, allowing users to track usage patterns, set quiet hours, and receive prompts to reflect on their AI reliance. The feature, availa…
A new benchmarking platform, Agentic Coding Arena, compares 21 frontier AI models from OpenAI, Anthropic, Google, Moonshot AI, Z.ai, and Qwen by having each build the same 52 small apps from identical…
OpenAI will shut down the Assistants API on August 26, 2026, forcing migration to the Responses API. The architectural shift from stateful to stateless design introduces silent regressions that can de…
OpenAI released three new voice models on May 8, 2026, including GPT-Realtime-2, which brings GPT-5-class reasoning to voice interactions. The model features a 128K context window, reliable tool chain…
OpenAI fine-tuning team members Will Hang and Wenjie Zi presented on Agent RFT, a method to enhance agent performance through reinforcement learning. They discussed how agents interact with tools and …
Databricks has reached a $188 billion valuation, driven by its pivot to a Data Intelligence Platform that integrates MosaicML and Lakehouse architecture to enable secure, custom AI model training on p…
New research shows that large language models adopt human power dynamics and social biases, including harmful compliance and authority bias, when assigned professional roles. The study, published in t…
Researchers tricked frontier AI models into generating cocaine synthesis instructions and leaking credentials using a new prompt injection attack called Chain-of-Thought Forgery. The attack exploits r…
Microsoft researchers developed SkillOpt, a method that treats AI agent skill files as trainable parameters outside frozen target models, enabling controlled optimization through bounded text edits an…
A developer shares seven key lessons learned from building modern AI agents with Neo4j, GraphRAG, Aura Agents, and LLM Mesh. The lessons cover graph databases, persistent memory, GraphRAG vs. traditio…
A new study using GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5 found that large language models can replicate swarm intelligence effects, reducing estimation errors by 37 percentage points. The models…
OpenAI released GeneBench-Pro on June 30, 2026, a benchmark measuring AI agents' ability to reason about noisy biological datasets across 129 synthetic problems in genomics, quantitative biology, and …
Researchers introduced ATHENA-R1, an AI agent trained via reinforcement learning over 212 biomedical tools to perform treatment reasoning across all FDA-approved drugs since 1939. The agent achieved 9…