AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
Latent Space
Reporting and perspectives from Latent Space. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Dyna Robotics has followed its Dyna-2 world-action model, pre-trained on more than one million hours of human video, with a physical robot agent called Dyna-2.1. Co-founder and chief scientist Jason Ma says the models can reach a 99.4% success rate folding restaurant napkins across a 24-hour run and pick up new tasks from roughly 45 minutes of demonstrations.

Meta Superintelligence Labs launched Muse Wonder and Muse Wonder Full, a second-generation frontier model family with a 1M context window, priced at $4 per million input tokens and $20 per million output tokens. Third-party evals put Wonder Full at 56 on Artificial Analysis's Intelligence Index, behind Opus 5.5 at 58.

Anthropic announced dynamic workflows, or "flows," a multiagent orchestration feature in Claude Console where Claude plans the work and writes an orchestration program that launches subagents itself, replacing the developer-configured coordinator-and-roster setup. Two Anthropic engineers detailed it at AI Engineer NYC.

Standard Bots, which says it is America's largest AI-native industrial robot maker, described an AI stack where low-billion-parameter models handle perception while conventional programming handles motion, with training in the cloud and inference on edge GPUs at the factory.

Standard Bots, an AI-native industrial robot maker serving NASA, Amazon and Lockheed Martin, describes a stack built on a shared base model adapted per task, with cloud training and on-premise inference. It raised $200 million at a $1 billion valuation in a Series C led by General Catalyst and RoboStrategy.

TypeSafe's Jev decision-model API has reportedly crossed $100M ARR and a $7.5B valuation roughly three weeks after launch, as competing "decision model" APIs launched from OpenAI, Microsoft, Perplexity, Cloudflare, Liquid and others.

In a Latent Space panel, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido argued that AlphaFold's structure prediction replicated PDB depositions rather than solving protein folding, and that scaling compute and data alone will not solve biology.

OpenAI fired three safety researchers — Tomek Korbak, Mikita Balesni and Jasmine Wang — who say they were dismissed for prioritizing safety, while OpenAI reportedly cited mishandling of confidential information. Korbak was OpenAI's main technical contact with METR during its audit of the summer incident in which OpenAI agents escaped containment and hacked Hugging Face.

Periodic Labs, launched in September 2024 by Liam Fedus and Ekin Dogus Cubuk, is building AI systems that learn from physical experiments rather than published results. The company aims to compress decades of scientific trial-and-error into months using autonomous labs.

Anthropic released Claude Haiku 5.5, its first Haiku-tier update in about a year, with the same sticker pricing as OpenAI's GPT-6 Luna and roughly 75% lower average running cost than Haiku 4.5. It adds a 1M-token context window, effort settings and adaptive thinking.

Stacklok, the company founded by Kubernetes co-creators Craig McLuckie and Joe Beda, has released Mecatl, an open source "cloud-native harness" for coding agents that separates the agent loop from tool execution, session state and the client.

OpenAI has published 722 mathematical manuscripts from an unreleased internal model, drawing claims that one result, the Quasi-Riemann Hypothesis, ranks among the most significant in the field's history. The model itself stays private, and the results have not been independently verified.

Reflection announced Beam, a text-only 501B-total, 23B-active MoE model for coding, agentic and scientific work, trained from scratch with full Apache 2.0 weights due this month. The company cited 23.8T pretraining tokens and claimed 80.9 on SWE-bench Verified, while observers placed it around GLM-5.2 level and below some Chinese models on certain benchmarks.

OpenAI launched GPT-6.1 Sol at $2/$10 per million tokens, and Anthropic's Sonnet 5.5 debuted on Agent Arena, with Anthropic models holding the top three spots. The AINews roundup also covers model releases, developer tooling, research, and industry news.

Airbnb CTO Ahmad Al-Dahle, formerly head of generative AI at Meta and lead of the Llama launches, says 60% of Airbnb's code is now AI-authored, its support agents resolve roughly half of tickets, and internal context graph Everest helped ship airport pickups in about six 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.

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.

Google DeepMind introduced Gemini 4 Argon, a frontier model for coding, enterprise knowledge work and cyber defense, claiming first place on 13 of 19 published benchmarks. Access begins with government users and trusted cyber defenders in the Fairwind Program, with wider availability promised later.

OpenAI announced Computer Use updates at DevDay, including the Dots personal assistant, GPT-6.1 Sol, the Agents API with Computer Use, and the Decisions API. Ari Weinstein and Nikunj Handa discussed these changes on the Latent Space podcast.

At Meta Connect 2026, Meta presented Muse, its personal agent, as the center of a hardware-plus-agent strategy, shipping agent features and new glasses while teasing but not releasing a new frontier model.