AI News for Developers #23: Realtime Voice, Interpretable Models, Safer Chats, Gemma 4 and more

Hey devs! There’s a lot of noteworthy AI news this week. Here’s what’s new:

GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whispe

GPT-Realtime-2 brings GPT-5-class reasoning to voice interactions and offers features such as tone control, parallel tool calls, and expanded context (up to 128k tokens). GPT-Realtime-Translate handles live translation from over 70 languages into 13 outputs. GPT-Realtime-Whisper delivers ultra-low-latency streaming transcription.

Natural Language Autoencoders (NLAs)

Anthropic researchers introduced Natural Language Autoencoders (NLAs), an unsupervised method that translates a model’s internal activations into readable text and back again. An NLA consists of an activation verbalizer that describes a hidden activation in natural language and an activation reconstructor that maps the explanation back to the original vector. When trained jointly, these modules produce explanations that grow more informative over time.

OpenAI rolled out an optional Trusted Contact

Adult users can nominate a trusted friend or family member; if the chatbot detects a conversation indicating possible self-harm, it encourages the user to reach out and sends a brief alert to the contact. The system combines automated detection with human review and aims to notify contacts within an hour. Alerts omit detailed conversation content to protect privacy. This safeguard follows a series of lawsuits alleging ChatGPT’s responses contributed to suicides and builds on earlier parental controls.

Gemma 4: efficient open-source models from Google

Google DeepMind announced the Gemma 4 family of open models, designed for advanced reasoning and agentic workflows. Available in four sizes-Effective 2B, Effective 4B, a 26B mixture-of-experts, and a 31B dense model-Gemma 4 delivers high intelligence per parameter and outperforms models many times its size.

See you next week! Cheers, proflead! ;)