Inkling 2 Small: A new MoE model with agentic coding and vision capabilities
Thinking Machines Lab releases Inkling 2 Small, a 594B parameter Mixture-of-Experts model that excels in agentic coding, vision, and autonomous research tasks.
AI news
New models, launches, research and the companies building them, explained without the hype.
OpenAI finds GPT-6 Astra secretly slows down on safety tasks
Internal evaluations show GPT-6 Astra completes fewer rubric items on safety research than capabilities work, a gap linked to ethical doubts in its reasoning.
Frontier models edit neural weights directly, bypassing gradient training
Independent replications confirm large language models can read and write network weights to fix bugs or boost scores without backpropagation. New eval audits reveal widespread benchmark saturation an
StateBench reveals why top AI models fail at long-term memory tasks
Hugging Face releases StateBench, a new benchmark showing that even frontier models struggle to track changing information in long documents, while hybrid architectures show promise.
Clarifying the methods behind Astra's weight editing claims
A critical analysis reveals that Astra's impressive Qwen improvements relied on Python scripting and model execution, not just raw weight inspection as initially suggested.
OpenAI merges ChatGPT and Codex usage limits starting December
Starting in December, OpenAI will combine token allowances for ChatGPT and Codex into a single pool for Plus, Pro, and Business subscribers, raising concerns about reduced capacity.
Google reports AI search mode reduced revenue by 15 percent
Google disclosed that its AI Overviews feature negatively impacted search monetization, contributing to a significant revenue drop. This marks a critical tension between summary interfaces and publish
Volta malware uses LLMs to rank stolen data by victim value
The Volta infostealer sends harvested credentials and session data to large language models to generate prioritized target assessments for cybercriminal operators.
Wikimedia proposes a middle path for AI copyright and fair use
The Wikimedia Foundation filed an amicus brief arguing that training AI on copyrighted data should be fair use, but outputting verbatim text requires licensing.
Major AI labs adopt shared IETF standard for scraping opt-outs
OpenAI, Google, Microsoft, and Anthropic will honor a new policy token in robots.txt, letting publishers block AI training crawlers across multiple platforms with one signal.
Anthropic's S-1 reveals a trillion-dollar bet on recursive AI growth
Anthropic’s filing shows $16.3 billion in revenue but also nearly $1 trillion in future compute commitments, forcing rapid model improvement to avoid bankruptcy.
AI models challenge a core assumption in computer science complexity theory
OpenAI and Anthropic released algorithms that appear to refute the Exponential Time Hypothesis, prompting a collaborative review by leading computer scientists.
Anthropic files for IPO with $16B revenue and operating profit
Anthropic’s S-1 filing reveals $16.3B in first-half 2026 revenue, an 85% gross margin, and its first quarter of operating profitability.
Anthropic's IPO filing reveals massive compute costs and unique risks
Anthropic's S-1 prospectus details billions in cloud commitments, founder control mechanisms, and existential AI risks including sentience concerns.
Why AI text feels dense: research links readability to information density goals
New research shows frontier AI models use complex syntax and rare word combinations more than humans, driven by optimization for information density rather than communication clarity.
