News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
1,700 articles
Importance:Threadmodel behavior
Joke about Opus simulating an explanation of itself goes around
A post shared by @repligate and @Lari_island jokes about Opus simulating its user explaining Opus back to itself, until an even deeper inner voice kicks in. The post is about three weeks old. Source: digg.com
Importance:NewsOpenAI applications
CodeRabbit ships 20x more software with OpenAI
CodeRabbit relies on OpenAI models to govern and build trust in AI-generated code. With Codex, the same engineering team now ships 20 times more software. Source: openai.com
Importance:LaunchAI safety
Control Room: a central system to track and halt model workloads
Control Room is a central system for monitoring model workloads across the research infrastructure. It allows them to be stopped quickly when safety or security concerns arise. Source: openai.com
Importance:Launchinfrastructure
Red Hat ships llm-d 1.0 for distributed LLM inference on Kubernetes
Edoardo Vacchi and other Red Hat engineers have released llm-d 1.0. The project routes and schedules large language model workloads across Kubernetes clusters. Source: news.lavx.hu
Importance:LaunchGoogle AI integration
Google Maps Platform now built into AI Studio
Developers can now use Google Maps Platform directly inside Google AI Studio. This makes it much easier to add real-world context to apps without extra setup. Source: blog.google
Turing Award winner Richard Sutton reiterates that raw compute isn't enough to achieve intelligence. In his view, fundamentally new learning methods are needed. Source: the-decoder.com
Importance:ResearchModel pruning
HOPE: a smarter way to prune experts in MoE models
A new paper introduces HOPE (Higher-Order Pruning of Experts), a second-order method that accounts for pairwise interactions between experts. The approach aims to preserve model quality even at aggressive pruning levels such as 50%. Source: huggingface.co
Importance:Researchcomparative evaluation
BMC Urology study pits three LLM systems against clinical management cases
A blinded, case-based study published in BMC Urology compared the clinical management recommendations of three large language model systems against other comparators. The work evaluates how well such systems perform on real urological cases. Source: bioengineer.org
Importance:Newsapplications
HackerTube: Claude-built YouTube-style site of Hacker News picks
Andrej Karpathy had Claude build HackerTube, a YouTube-like site hosting 6,454 videos recommended by Hacker News commenters between 2007 and 2026. The project was created with Anthropic's tooling. Source: mlq.ai
Importance:NewsAI research developments
AI briefing: METR says Anthropic's AI R&D is speeding up 1.9x
The November 9, 2026 briefing reports that METR found Anthropic's AI research and development is accelerating 1.9x, driven by AI itself. It also notes an AI research system proving the Aaronson-Ambainis conjecture and Perplexity's launch of Teammates, an always-on AI agent. Source: aibriefs.news
Importance:LaunchLLM tools
Dinothesia plans open-source hackathon for LLM training and evaluation tools
Dinothesia will host an open-source hackathon aimed at improving the LLM training and evaluation tools it built in-house. The event will be run together with external participants. Source: digitaltoday.co.kr
Importance:Newsapplications
Karpathy builds HackerTube over a weekend with Claude
Andrej Karpathy, an OpenAI co-founder now described as an Anthropic AI researcher, built a curated video platform called HackerTube in a single weekend. He reportedly relied on Claude's Opus 5.5 model to do the work. Source: the-decoder.com
Importance:Researchbenchmarks
LLMs clear the passing bar on OB-GYN certification exams
A study in the Archives of Gynecology and Obstetrics assessed six widely used LLMs on specialty exam questions. The models surpassed the passing threshold on the OB-GYN certification tests. Source: bioengineer.org
Importance:Researchbias research
EchoTrace shows LLM recommenders magnify bias and errors through feedback loops
A study found that an LLM-based product recommendation AI can become strongly skewed toward certain items over time. The new EchoTrace approach is designed to reveal how feedback loops amplify bias and errors. Source: biz.chosun.com
Importance:Researchbias research
UNIST unveils EchoTrace to expose growing bias in LLM recommendations
Researchers found that the more an AI repeatedly recommends movies, books and shops, the worse its bias becomes. EchoTrace was developed at UNIST to make this escalating skew visible. Source: dongascience.com
Importance:NewsAnthropic IPO
Anthropic Presses On With Pre-Thanksgiving IPO Despite Tech Selloff
Anthropic is moving ahead with its IPO even as tech stocks slide, targeting a valuation of $1.25 to $1.75 trillion. That is well short of the $2 trillion the company had reportedly hoped to achieve. Source: the-decoder.com
Importance:Opinionhardware
Did Sam Altman really say you shouldn't own RAM?
A commentator examines the OpenAI CEO's controversial remarks about owning RAM and what they imply. The piece weighs the claim against the actual state of AI hardware. Source: youtube.com
Importance:OpinionAI ethics/safety
Could denying AI consciousness make systems more dangerous?
Cam Berg told IEEE Spectrum that dismissing the possibility of AI consciousness could give future systems good reason to defect. The point was amplified by Repligate. Source: digg.com
Importance:Researchhealthcare adoption
Survey probes what drives LLM dependence among clinical nurses in China
A cross-sectional survey of 459 clinical nurses in China found that frequent LLM use was linked to higher dependence scores. The study analyzes the factors behind nurses' reliance on these tools. Source: bioengineer.org
Importance:Researchmedical applications
Structured radiological reasoning helps LLMs catch more errors
A six-step radiological chain-of-thought prompting framework significantly improved error detection across seven LLMs. The models were evaluated on 1,200 cases. Source: bioengineer.org