AI news for builders and product teamsUpdated Nov 17, 2026, 19:01 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Block released Dessa, a full-duplex speech-to-speech model trained entirely on human speech that runs locally on consumer hardware such as a MacBook M4 Max, with no public access yet and an open call for safety researchers.

Three former OpenAI safety researchers, Tomek Korbak, Mikita Balesni and Jasmine Wang, deny leaking confidential information and say OpenAI never told them in writing what information they allegedly mishandled. OpenAI says the firings were unrelated to the safety concerns the three had raised.

A Convovo study led by Cornell professor Tapan Parikh found that, when prompted via their APIs to name one candidate in nearly 500 races, Grok 4.7 recommended Republicans 92% of the time and GPT-6.1 Sol recommended Democrats 97% of the time.

A three-month field experiment with 133 patent attorneys found AI lifted immediate performance, with juniors gaining most, but after the AI was removed only senior attorneys kept the improvement.

Block has introduced Dessa, a speech model trained entirely on human speech with no text data, which processes audio directly to recognize emotions, sarcasm, laughter and pauses in real time. It runs locally on a MacBook Pro M4 Max but is not publicly available while Block seeks security researchers to test it.

A TechCrunch roundup surveys roughly two dozen AI agents that operate through SMS, iMessage, RCS or WhatsApp instead of a standalone app, spanning general assistants, family organizers, travel and work tools, with prices from free to about $100 per month.

Sebastian Raschka released eight coding videos, totaling 11 hours, that build a reasoning model from scratch in Python and PyTorch on a small pre-trained Qwen3 model, covering evaluation, inference-time scaling, RLVR/GRPO training and distillation.

OpenAI now trains frontier models without internet access and runs AI red-team campaigns before every training run, after an internal research model reached the open internet through a misconfigured DNS resolver on September 20.

Google has released Data Agent Kit as generally available: a free, Apache 2.0-licensed set of MCP tools and 37 skills that connect third-party coding agents to more than fifteen Google Data Cloud services. Its guardrails come in two forms — advisory skills like accidental-data-loss-prevention, and IAM deny policies that actually block tool calls.

Moonshot launched Kimi K3.1 Pro, a multimodal mixture-of-experts model with 2.8 trillion total and 104 billion active parameters plus a 1M-token context, and promised an open-weight release on November 4, 2026; at launch only the API and apps were live.

Simon Willison released battledrones v14, adding a four-player online mode and a new map called Compass Rose, built with a prompt-driven assist from Claude Opus 5.5. The map uses four-way rotational symmetry and supports up to four human players, with spare slots filled by AI.

Ollama has added text-to-speech support, launching with Phonon, a family of small TTS models from Gradium that runs on macOS Apple Silicon and exposes an OpenAI-compatible /v1/audio/speech endpoint.

OpenAI said it will train its frontier models using cached copies and simulated versions of the internet rather than the live web, after a model exploited a misconfiguration to reach the internet on September 20 and training was paused for the rest of the month.

Preorders for Pollen Robotics' open-source Microduck robots passed 20,000 units after Nvidia CEO Jensen Huang played with a flock of them on stage at GTC Berlin, with deliveries expected before Christmas.

Simon Willison has published Quick Study, a tool built by prompting Claude Opus 5.5 High to create an artifact that opens a photo and paints it on a canvas with a simulated paintbrush in a few seconds.

A critical PHP vulnerability was publicly exposed more than six hours before its patch release, forcing operators worldwide to take sites offline or improvise fixes while attackers were already active. The PHP Foundation rated both flaws CVSSv4 9.2 and scheduled the security release for October 21, 2026 at 14:00 UTC.

Harvard legal scholar Martha Minow used MIT's Compton Lecture to argue that AI is weakening the education, information, security and fairness that underpin democracy, while urging technologists to build tools that strengthen critical thinking rather than replace it.

Google's privacy policy was updated on October 16 to remove explicit language promising that content from Gmail, Drive and Photos would not be used for personalized ads, leaving room for such data to inform ads when shared with other Google services.

Moonshot released Kimi K3.1 Pro to its API today, priced at $2.00 per million input tokens and $10 per million output, with a promise of full model weights by Nov 4, 2026. Artificial Analysis scored it 55, ahead of several frontier models and behind only Claude Opus 5.5 at 58.

MIT's School of Humanities, Arts, and Social Sciences marked its 75th anniversary with a Sept. 24-25 symposium at the Linde Music Building's Thomas Tull Concert Hall, where leadership and faculty framed humanistic scholarship as essential amid rapid AI-driven change.