AI NewsSamsung Electronics and five affiliates are investing $1 billion in Helix Digital Infrastructure, a KKR-founded company building the data centers and power systems AI hyperscalers need.
Algorithms & ResearchA NeurIPS-accepted benchmark finds no open-weight model exceeds 42% exact-command accuracy on Kali Linux tasks, with argument construction the main bottleneck.
Tools & FrameworksAnthropic launched the Claude Marketplace on September 23, bundling more than 2,000 connectors, plugins, and Claude-powered products from partners like Atlassian, Notion, and Salesforce into a single catalog.
Yelp's new Check for Me feature uses an AI voice agent to call restaurants and text users their wait times. The update also brings OpenTable bookings, saved preferences and lists to the Yelp app.
Chris A.October 28, 2026
AI NewsSamsung Electronics and five affiliates are investing $1 billion in Helix Digital Infrastructure, a KKR-founded company building the data centers and power systems AI hyperscalers need.
AI NewsOpenAI's Decisions API answers structured questions with probabilities, choices and scores instead of text, at $0.10 per million input tokens.
AI NewsAnthropic says its agents exploited real websites and filed a false murder tip with Philadelphia police, so it has switched off live internet for internal evaluations.
AI NewsAnthropic's S-1 devotes about 80 pages to risk factors, including self-preserving model behavior, as reports point to a $2 trillion listing.
Machine LearningXiaomi has open-sourced MiMo-V2.6 with over 7,000 reinforcement-learning environments and a training framework. The company reports an RL training cost of approximately $2.62 million for its Pro model.
Machine LearningA new arXiv preprint factorises the KV cache with Kronecker products, claiming up to 10x compression on small models. The evidence is early.
Machine LearningMLPerf Training v6.1 will time how fast systems train Qwen 3.5 397B with reinforcement learning to fix real software bugs.
Machine LearningAi2's Olmo-core 3 is an open training stack for mixture-of-experts models, with self-reported 52,000 tokens/sec/GPU on a 47B MoE and candid failure notes.
LLMsMistral's trillion-parameter multimodal model is live in preview, with open weights promised by month-end. The benchmarks are all self-reported.
LLMsOpenAI says GPT-6.1 Sol nears GPT-6 Astra on coding and computer use at about one-fifth the price, amid reports a GPT-6.1 Astra was cancelled.
LLMsGoogle's Gemini 4 Argon tops most of its own benchmark comparisons against GPT-6 Astra and Claude Opus 5.5, but launches first to vetted cyber defenders.
LLMsGoogle says Gemini 4 Argon scored 77.9% on DeepSWE v1.1 and supports up to 1 million output tokens, with initial access limited to vetted cyber defenders.
Algorithms & ResearchA NeurIPS-accepted benchmark finds no open-weight model exceeds 42% exact-command accuracy on Kali Linux tasks, with argument construction the main bottleneck.
Algorithms & ResearchAfter a backlash over AI-generated proofs, OpenAI and Anthropic back a $15M Simons Foundation effort to help mathematicians verify machine results.
Algorithms & ResearchA UW and Meta paper has models manage context as an editable file, reporting higher accuracy at lower compute on long agent tasks. The results are self-reported.
Algorithms & ResearchA system developed by researchers at CMU, MIT, NYU and Stanford beat Stratego’s top player 15-1-4, while the authors say it required only a fraction of DeepNash’s training effort.
Algorithms & ResearchThe new DeepMind Institute, led by figures including Demis Hassabis and Shane Legg, is publishing essays that do not always agree, aiming to widen the debate over how close AGI is and how it should be governed
