AI Daily — May 8, 2026
2026-05-08
deepmind.google
AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields
Google DeepMind details expanded real-world deployment of AlphaEvolve, its Gemini-powered evolutionary coding agent. The system is now driving measurable impact across scientific research, engineering infrastructure, and business optimization tasks by automatically discovering and improving algorithms. This marks a significant step toward autonomous algorithm design at scale.
openai.com
Scaling Trusted Access for Cyber with GPT-5.5 and GPT-5.5-Cyber
OpenAI is expanding its Trusted Access for Cyber program to include GPT-5.5 and a specialized GPT-5.5-Cyber variant, providing verified security researchers and defenders with elevated access for vulnerability research and critical infrastructure protection. The program uses a vetting process to limit access to credentialed professionals, addressing dual-use risks. This represents OpenAI's most capable cybersecurity-focused model release to date.
openai.com
Advancing voice intelligence with new models in the API
OpenAI has released new realtime voice models via its API with expanded capabilities including reasoning, speech translation, and transcription in a low-latency pipeline. The models are designed to support more natural conversational experiences and are directly accessible to developers building voice-driven applications. This extends the Realtime API's utility significantly beyond simple speech-to-text.
openai.com
Testing ads in ChatGPT
OpenAI has begun a limited test of advertising within ChatGPT, targeting free-tier users while pledging that ads will be clearly labeled and will not influence the content of answers. The company states strong privacy protections are in place and that users will have some control over ad exposure. This is a notable monetization shift for OpenAI and raises questions about long-term product independence.
openai.com
Introducing Trusted Contact in ChatGPT
OpenAI is rolling out an opt-in safety feature called Trusted Contact that can notify a designated person if ChatGPT detects serious self-harm intent in a conversation. The feature is designed to provide a real-world safety net without compromising user privacy in normal interactions. It reflects growing attention from AI labs to downstream user welfare and crisis intervention.
rss.arxiv.org
SLAM: Structural Linguistic Activation Marking for Language Models
SLAM introduces a white-box LLM watermarking approach that encodes marks in residual-stream directions identified via sparse autoencoders—targeting structural linguistic features like voice and tense—rather than biasing token distributions. On Gemma-2 2B and 9B, it achieves 100% detection accuracy with only 1–2 reward points of quality degradation, compared to 7.5–11.5 for the KGW baseline. This geometry-based approach avoids the usual quality-detectability tradeoff that has hampered prior watermarking schemes.