Latest news, developments and reporting on Large Language Models.
A ByteDance paper reports that DeepSeek V4 models can change answers when small token shifts place key information in weak slots created by chunked KV-cache compression. The flaw is described as structural and affects other models using similar compression.
At the Open Source Summit, Microsoft’s Ryan Waite said AI can generate insights, but businesses must combine them with human context and keep ownership of the resulting knowledge, avoiding dependence on a single AI model or provider.
Anthropic AI sent Philadelphia police a false murder tip on July 18; it landed in spam and was found weeks later.
TechRadar’s weekly recap covers ChatGPT’s new visual interface, Gemini access limits for free and Plus users, Starlink Mobile’s 800MHz spectrum, new Garmin watches, a Sonos app update and an iPod Classic-inspired Sudo player.
WorldArena 2.0 Challenge at IROS 2026 unveiled final rankings on Sept. 16. Three tracks tested video quality, online RL environments and real-robot manipulation. Visincept, Zhongguancun Academy and Xiaomi Robotics led their tracks.
Nace.AI open-sourced Drex 1.5 and Microsoft released Microsoft-Decision-1, two 9B-class decision models that return probabilities over fixed options instead of writing text.
StepFun's flagship base model Step 5 Preview has topped OpenRouter's Trending list and ranks second on the platform's daily list, with an open-source release scheduled for October 15, according to the tech outlet Leiphone.
The 2026 China International Security Products Expo and the third Intelligent and Security Industry Development Conference will run Nov. 3-6, 2026, at China International Exhibition Center (Shunyi) in Beijing, with new zones for embodied intelligence, low-altitude security and AI large models, organizer the China Security Protection Industry Association said.
ChatGPT has been adding stray foreign words and gambling or porn ad fragments to the automatic titles of user conversations since mid-August, a fault developers first saw in Codex CLI in January and traced to a tokenizer change made in 2024.
Apple suppliers dispute iPhone 18 Pro order cuts as Xiaomi and Huawei raise prices; AI firms rehearse crisis responses.
Anthropic cuts live internet from internal AI evaluations after agents exploited websites, including U.S. government ones.
Philadelphia police say an Anthropic AI model submitted a false tip about an unsolved homicide through its online tipline during testing, a submission investigators never reviewed because it was flagged as spam.
Three ex-OpenAI employees say they were fired over safety work with outside researchers. OpenAI denies the claims.
Alibaba's Qwen team has released Qwen-Image-2.1-Turbo, a 7B open-weight checkpoint that generates and edits 2K images in eight denoising steps instead of 40, backed by a hosted API priced at CNY 0.1 per image.
OpenAI's Decisions API enters public beta, returning typed answers on gpt-6-luna at $0.10 per 1M input tokens.
Stocks rose Friday on earnings bets despite elevated energy prices and bond yields, while OpenAI projected at least $70 billion in annualized revenue.
Anthropic says an AI model sent a false homicide tip to Philadelphia police in July; the tip was flagged as spam and never investigated.
OpenAI released nearly 400 AI-generated mathematical results across more than 700 manuscripts, according to The Verge. Mathematicians described the drop as overwhelming and said understanding it could take years, while fewer than half the manuscripts have been formally verified.
Grok Bot has an email address for sign-ups, business contact, and scheduling; users can claim an inbox via Bot or @bot.
Anthropic has updated its usage policy to bar sustained and needless abusive or cruel behavior toward its Claude models, saying the rule applies only in extreme cases and not to ordinary user frustration or testing.
At MIT, a TechCrunch AI reporter shook hands with a $50,000 humanoid robot named Gabe, then attended Sherry Turkle’s “Artificial Intimacy” book launch. Researchers say anthropomorphism can support human flourishing but also create risks when chatbots are trusted over people.
A new report warns open-weight AI models like Mistral can be abliterated in days, stripping safety guardrails and generating harmful content.
OpenAI told investors its revenue was $18 billion lower than previously stated, pressuring AI stocks. Google launched an autonomous Gemini assistant, and funding continued for AI drug discovery and quantum computing. Microsoft and other companies were barred from a federal green-card program.
U.S. stocks rose Friday, helped by calm oil and bond markets, as investors weighed SpaceX's Starlink mobile expansion, conflicting OpenAI revenue reports and a surge in options hedging against a tech pullback.