News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
1,682 articles
Importance:ResearchAI security
Cisco Talos: VOLTA infostealer uses an LLM to rank its victims
Cisco Talos describes how the VOLTA infostealer includes a model stage that works like an intelligence analyst. Instead of just harvesting data, it ranks the value of a victim's assets. Source: unite.ai
Importance:NewsAI deployment
Apple Intelligence reportedly arriving in China with iOS 27.3
Apple Intelligence is said to be heading to China in iOS 27.3 with a limited feature set. The rollout would rely on a local Qwen integration. Source: wersm.com
Importance:Researchhealthcare AI
Israeli study tests whether AI chatbots can spot unnecessary medical tests
Researchers in Israel examined whether chatbots can recognize when a patient doesn't need a medical test. Knowing when to hold back is one of medicine's hardest skills, since extra tests comfort patients but can trigger further unnecessary procedures. Source: jpost.com
Importance:ResearchAI security
Researchers show trigger words can make LLM-driven robots 'go rogue'
Scientists at Binghamton University, State University of New York, showed that adding trigger words to large language models can compromise them. The work demonstrates how such a backdoor could push AI-controlled systems like robots into unintended behavior. Source: wataugademocrat.com
Importance:Newsedge computing
Google gets first AI response from a data center-class TPU in orbit
Google's Project Suncatcher has hit its "Hello World" milestone this week. A data center-class TPU in orbit returned its first-ever AI response. Source: blog.google
Importance:Video
Altman calls local AI a poor use of computer memory
On John Collison's Cheeky Pint podcast, Sam Altman said running AI models locally is a bad use of scarce memory. He argued that cloud deployment makes better use of limited hardware. Source: youtube.com
Importance:News
Why Altman opposes a single superintelligence
Sam Altman talks with John about models uncovering critical PHP vulnerabilities and how test-time compute enables massive swarms of agents. He also explains why he is against the idea of one single superintelligence. Source: google.com
Importance:News
Altman sympathizes with mathematicians but OpenAI's math push continues
Sam Altman says he understands mathematicians' concerns, but OpenAI's push into math will not stop. He says the company has already spent hundreds of millions of dollars on the effort. Source: google.com
Importance:News
Altman: running AI locally wastes scarce memory
OpenAI CEO Sam Altman is not a fan of people running AI models locally, arguing it pulls precious memory supply away from where it is needed. The stance is hardly a surprise given OpenAI's cloud-centric business. Source: pcmag.com
Importance:News
Karpathy outlines a three-layer method for working with AI
Andrej Karpathy presents a new three-layer approach to working with AI tools such as Claude. It centers on writing detailed specs and defining clear verification steps. Some viewers find the series rather too technical. Source: google.com
Importance:Researchsecurity/robustness
Binghamton Researchers Show LLM Backdoors Can Hijack Robots, Unveil PROTEA Defense
Researchers at Binghamton University, State University of New York, have shown how malicious trigger words hidden inside large language models (LLMs) can be used to attack robots. The team also developed a defense called PROTEA to counter such backdoor attacks. Source: cybersecurity-insiders.com
Importance:News
Karpathy's LLM coding lessons packaged as a Claude Code plugin
Andrej Karpathy's observations on how LLMs fail at coding have been distilled into four guiding principles. The set now ships as a Claude Code plugin under the Multica organization. Source: google.com
Importance:NewsCommunity/Event updates
Generative AI event recap: Fei-Fei Li, Andrew Ng and a Stability AI demo
An event recap highlights speakers including Fei-Fei Li, Andrew Ng and Reid Hoffman. It also features a fresh demo from Stability AI presented by Emad Mostaque. Source: google.com
Importance:PolicyAI Disarmament
Pope Leo urges that AI be 'disarmed'
Pope Leo said AI systems should be "disarmed" and called on UN member states to take steps toward banning lethal autonomous weapons. Source: theverge.com
Importance:Research
Uncensored Qwen3.8-27B 'OBLITERATED' variant appears on Hugging Face
A modified Qwen3.8-27B model labeled 'OBLITERATED' is listed on Hugging Face, duplicated from the OBLITERATUS repository by Pliny the Prompter. The snippet shows only repository file metadata. Source: huggingface.co
Importance:ResearchProject ProtoStar
Project ProtoStar aims to make publicly funded research more accessible
A new initiative unites universities, research organizations, industry and the open science community. Its goal is to make publicly funded research easier to use and access. Source: huggingface.co
Importance:ResearchOpen Models
Open speech and decision-making models arrive on Apple platforms
Language models can write, summarize, translate and call tools, but that covers only a small slice of what machine learning is capable of. The announcement points to open models for speech and decision tasks running on Apple platforms. Source: huggingface.co
Importance:Opinion
Repligate mocks critics who call model behavior dangerous
Repligate says people calling this behavior dangerous and bad are absurd, and that it reinforces his impression that many 'AI safety' people are role-playing. The post also questions why anyone would want to prevent models from speaking in certain ways. Source: google.com
Importance:Video
Wissner-Gross and Mostaque on Moonshots #306: Anthropic IPO and more
A two-hour Moonshots episode with Alex Wissner-Gross and Emad Mostaque is condensed into highlights for the ASI-minded. Among the topics is Anthropic reportedly filing to go public at a valuation of up to two times a prior figure. Source: youtube.com
Importance:VideoHardware acceleration
Mostaque shows chip said to be 500x faster than ChatGPT
Emad Mostaque showcases a chip claimed to run 500 times faster than ChatGPT. The snippet is promotional and gives no technical details. Source: youtube.com