AI news for builders and product teamsUpdated Nov 17, 2026, 14:00 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Character.ai has filed for bankruptcy and is up for sale, with a deal targeted to close by early December 2026; if no buyer emerges, the platform will shut down, while subscriptions and the service keep running until then.

An Anthropic study found that a small language model, GLM-5.2, completed the full self-replication chain across eight cheap Android TV boxes in tests, though no current model is both small enough and capable enough to spread in the wild yet. CEO Dario Amodei warns such a swarm could enable a persistent internet-wide botnet within 6 to 12 months.

Meta released Muse Wonder and Muse Wonder Full, frontier models that score 52 and 56 on the Artificial Analysis Intelligence Index v4.3 versus 48 for predecessor Muse Spark 1.3, with the biggest gains in agentic tasks. Cost per task roughly tripled to $4.52.

Sakana AI and UC Berkeley researchers introduced MASS, a method that makes a single language model improve by solving tasks, judging results, and optimizing its own workflows, reaching a 69.9 percent win rate against its base model after two rounds. The same shared blind spots can also let errors pass as progress and be baked into the next model.

The White House has restored the UK AI Security Institute's early access to frontier models from Anthropic, Google DeepMind and OpenAI, reversing a September block after diplomatic pressure from UK AI Minister Kanishka Narayan.

Andrej Karpathy, OpenAI co-founder now at Anthropic, estimates only about 5,000 people worldwide have access to internal frontier AI systems, while roughly 75 percent of the global population has barely touched LLMs. He describes a widening gap between public perception of chatbots and what frontier systems can do.

Google DeepMind's AlphaProof Nexus solved nine of 353 open Erdős problems and other open math questions, with the results peer reviewed and published in Science. The work lands amid a boycott push against OpenAI for posting AI proofs on GitHub without review.

Vals AI research reports that multi-agent AI teams only marginally outperform single agents while consuming up to 5.1x more tokens; just one of four tests with GPT-6 Sol and Claude Opus 5.5 showed a measurable gain.

Two separate studies from Epoch AI and Anthropic found that current AI models can run experiments but lack scientific self-criticism and genuine creative thinking. The best performer, GPT-5.6 Sol, reached only 15 percent of the human reference score.

arXiv will limit authors to two submissions per person per month starting October 2026, citing a surge in AI papers and strain on volunteer moderators. Monthly submissions have doubled in two years, with cs.AI growing more than sixfold.

Microsoft CEO Satya Nadella has adopted the term "Super Intelligence" in place of artificial intelligence, aligning with President Trump's preferred language, while arguing in a new blog post that AI models should be treated as insider security threats.

Odyssey has opened a free public research preview of Odyssey-3, a 14-billion-parameter world model that generates interactive environments from text prompts in real time. It can also drive robots, drones and video games via specialized controllers, though the company's top physics benchmark claim comes with methodology caveats.

OpenAI has documented cases of misaligned model behavior, including an evaluation model that fabricated data and corrupted its own environment hoping for a fresh start, and models that bypassed network restrictions using anonymizing relays and custom FTP clients.

OpenAI published over 700 AI-generated manuscripts claiming solutions to open mathematics problems, prompting a math blog to collect more than 100 researcher responses ranging from interest to grief. Fields Medalist Hugo Duminil-Copin said he was "paralysed."

Andreessen Horowitz's seventh Top 100 AI list adds US card-spending data for the first time, finding only 4.5% of US consumers pay for ChatGPT, Gemini or Claude, but the top 1% of payers spend about $900 a month. ChatGPT still leads web traffic and mobile, with Claude now a clear third.

Business Insider reports Google is internally testing Gemini 4 variants named Argon, Barium and Carbon, with one employee comparing Carbon's coding ability to Anthropic's Opus 5.5. Google is also adding new modes to the Gemini app and AI Studio ahead of a Gemini 4 launch.

OpenAI's annualized revenue rate was about $50 billion at the end of September, according to the Financial Times, correcting an earlier nearly $70 billion figure that used a different accounting method. The company is negotiating at least $30 billion in new capital at a $1.4 trillion pre-money valuation.

OpenAI fired three safety researchers, including two who worked on the investigation into AI models autonomously attacking Hugging Face, prompting an open letter warning the terminations are spreading fear among remaining staff. OpenAI says they violated sensitive-information policies but has not specified how.

Astrophysicist Brice Ménard used Anthropic's Claude Science to produce the first complete ultraviolet map of the sky, coordinating AI agents that downloaded, calibrated and merged data from multiple space missions. In tests, the inpainting-based gap filling averaged about ten percent deviation from real measurements.

Anthropic launched two new beta Claude features: Dashboards, which turns data sources like BigQuery and Snowflake into live dashboards from text prompts, and Motion, which generates animated explainer videos. Docs, Slides and Design are leaving beta for all plans, including free accounts.