AI news for builders and product teamsUpdated Nov 18, 2026, 07:00 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Microsoft AI released MAI-Transcribe-2-Streaming for real-time transcription and two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Microsoft says the transcription model ranks first for accuracy on Artificial Analysis and handles 60 languages.

US companies are using more AI but spending less on it, according to the Ramp AI Index from Ramp economist Ara Kharazian. He attributes the drop almost entirely to competition between OpenAI and Anthropic, alongside price cuts on top models and cheaper standard and lite models.

Docker announced it is bringing the Sandbox Kit Specification, now at v3 under Apache 2.0, to the CNCF, packaging an AI agent, its tools, and requested permissions into an OCI image. Docker announced the move at WeAreDevelopers on September 24.

DigitalOcean launched DigitalOcean Managed Agents in public preview, a managed cloud infrastructure layer for AI agents featuring isolated microVM runtimes, governed tool access, and serverless AI inference. The offering combines a Harness Runtime and an Action Gateway to handle execution, persistence, tool access, and underlying infrastructure.

Thore Graepel, formerly of Google DeepMind and a core member of the AlphaGo team, argues that today's large language models do not genuinely reason. In an MIT Technology Review piece, he says chain-of-thought processing is not a separate reasoning mechanism and calls for systems built on AlphaGo's search architecture.

Black Forest Labs has released Flux 3 Image, the image component of its Flux 3 model family, supporting multi-step edits that leave the rest of an image unchanged. A free demo is available, API access is 50 percent off through October 8, and an open-weight version is expected in the coming weeks.

Pi 1.0 and Pi Durable from Earendil both reached the front page of Hacker News, with Pi Durable porting Pi to TypeScript and externalizing stateful components. The same AINews roundup also covers Gemini 4 Argon, GPT-6.1 Sol, Upstage's Solar Mini 4, and Black Forest Labs' FLUX 3 Image.

ServiceNow CoreAI introduced AutoSynthData, a pipeline that uses a target model's failures and a stronger teacher's successes to generate validated training tasks for enterprise agents. In EnterpriseOps Gym experiments, synthetic supervised fine-tuning improved Gemma-4-26B-A4B-it mean Pass@1 by 7.2 percentage points in Hybrid and from 18.77% to 27.18% in ITSM.

MIT's Martin Trust Center has launched Dear Dreamer, a free AI-powered entrepreneurship platform for middle and high school students, built on Bill Aulet's 24-step Disciplined Entrepreneurship framework and funded by the Frank & Eileen Foundation. Aulet aims to reach 50,000 young entrepreneurs by 2030.

Latent Space featured MIT researcher Alex Zhang in a podcast episode about Recursive Language Models, GPU kernels, KernelBench, multi-agent swarms, and research taste. The episode covered RLMs, harness design, capability overhang, and alternatives to standard autoregressive language models.

Apple researchers show that discrete diffusion samplers match the training distribution only when the positions written per step are conditionally independent given already-fixed tokens, and that no product of per-position distributions can match a dependent group.

Apple researchers report that strengthening language discrimination during pretraining narrows the multilingual gap in English/French HuBERT speech models, improving phone discrimination, lexical and prosodic measures. Gains were largest when the intervention was introduced in the first training iteration.

Chatham Financial is using Codex and GPT-5.6 to build technology and redesign workflows, cutting trade validation time from 30 minutes to under 4, according to OpenAI.

GPT-6 Astra Ultrafast is now available in the OpenAI API and to eligible ChatGPT Work and Codex users, running on NVIDIA Blackwell GPUs with up to 8x faster token generation than Astra Standard mode.

Databricks published a piece describing how Genie One changes work for finance teams. The available item contains only a headline and a truncated excerpt, with no substantive detail.

Meta's TLX-based Jagged Flash Attention kernel for GEM on NVIDIA B200 uses about 3.2K lines of Triton code, roughly 3x less than FA4's ~10K-line CuteDSL kernels, while beating FA4 by about 13% on the forward pass and 50% on the backward for jagged shapes.

An AWS Machine Learning Blog post describes a four-agent pattern built with the Strands Agents SDK on Amazon Bedrock AgentCore for enterprise cloud migrations. The post reports that on a 300+ application program, the pattern cut infrastructure as code development from 3 to 4 weeks per application to minutes, based on internal project tracking data.

Researchers at MIT CSAIL, Google, and Northeastern University developed InstructMesh, an AI-driven interface that lets users generate and refine 3D designs for 3D printing. It combines Microsoft's TRELLIS with GPT-4, and in a study, novices identified and fixed flaws in generated models about 90 percent of the time.

Time magazine reports that President Trump met secretly with Elon Musk about a month before the U.S. invaded Venezuela and captured Nicolás Maduro, and spent hours talking to Musk's Grok chatbot. A source said Grok told Trump Maduro was deeply unpopular and many Venezuelans would likely celebrate his downfall.

Databricks described read restrictions and catalog labels as a way to unify governance across engines and catalogs, continuing its series on open table formats, open APIs and unified governance. No specific version, date, pricing or availability details were stated.