16 articles — 🔴 2 critical , 🟡 11 important , 🟢 3 interesting
GPT-5.5 Instant is the new default ChatGPT model introduced by OpenAI on May 5, 2026. The model delivers smarter, more precise responses, a reduced hallucination rate, and improved personalization — accompanied by a new system card.
EvoLM is a post-training method that eliminates external supervision — a Qwen3-8B rubric generator outperforms GPT-4.1 on RewardBench-2 by 25.7% and SkyWork-RM by 16%, while the trained policy reaches 69.3% on the OLMo3-Adapt benchmark.
Google expanded File Search in the Gemini API to multimodal search, enabling native embedding and retrieval of images alongside text documents through the gemini-embedding-2 model. Two new grounding fields and event-driven webhook support for the Batch API were also added.
Microsoft Research presented DroidSpeak at NSDI 2026 — a system that shares KV cache across architecturally identical fine-tuned LLM variants, achieving up to 4× higher throughput with minimal quality loss in enterprise scenarios running dozens of domain-specific models.
A new paper shows that safety in clinical LLMs does not follow the same scaling laws as accuracy — cleaner evidence in RAG raises accuracy from 73.5% to 94.1% and reduces high-risk errors from 12% to 2.6%, more than any model scaling effect.
The UK AI Security Institute announced a partnership with Microsoft on May 5 covering frontier AI safety. The collaboration spans three research areas: evaluation of high-risk capabilities, testing of safeguards, and research into societal resilience to conversational AI.
Anthropic releases 10 ready-made agent templates for financial services, including a pitchbook builder, KYC screener, and month-end closer. Templates ship as plugins for Claude Cowork and Claude Code, and Claude Opus 4.7 scores 64.37% on the Vals AI Finance benchmark.
MEMTIER is a five-tier memory architecture for long-running autonomous agents — on the LongMemEval-S benchmark with Qwen2.5-7B, accuracy jumps from 0.050 to 0.382, and the tool execution success rate stops declining after 72 hours of operation.
AWS announced OS Level Actions for Amazon Bedrock AgentCore Browser on May 5 — a capability enabling agents to interact with the native OS interface outside the DOM. It introduces 8 actions and an action-screenshot-reaction loop, available with no additional configuration.
DataClaw is a new benchmark that evaluates the entire workflow of AI agents in exploratory data analysis — not just the final answer — revealing weaknesses in agents that reach correct results through flawed reasoning.
IBM expanded the Enterprise Advantage platform by launching Context Studio, a tool for building AI agents grounded in an organization's own data while preserving digital sovereignty. Providence Health reduced manager hiring time by 90%, and IBM projects operational cost savings above 25% within 18 months.
Anthropic released Claude Code v2.1.131, a hotfix that resolves two bugs: a VS Code extension activation crash on Windows due to a hardcoded build path and a missing x-api-key header sent to Mantle inference endpoints. Binaries are published for all major platforms.
