Archive, page 25
The paper argues current System One decision models underemphasize structural requirements such as distinguishing absent answers from uncertainty and permutation invariance, and builds Jeva from those first principles.
A Nature comment argues research data must be preserved to enable future AI-enabled integrity checks on scientific work.
Nature presents NiCLIP, a neuroimaging contrastive language-image pretraining model that predicts text from brain activation images.
Nature examines a detector that claims to distinguish ideas generated by AI from those produced by humans.
A Nature piece argues that as AI handles more code generation, human programmers must shift to thinking like systems architects.
Agent-Reflex is a training framework that teaches language agents to act reflectively, handling noisy MCP tool feedback and user corrections, closing a gap in supervision for reflection learning and failure recovery.
GreenWAM is a robot world foundation model intended to act both as a predictor of how actions change a scene and as a policy across different robot bodies, addressing the split between world models and policies.
FlashVSR+ is a video super-resolution method that adds hybrid attention and reward-yield distribution matching to one-step diffusion models, making high-resolution inference cheaper and better aligned with human preferences.
The paper addresses sparse-to-dense motion completion, showing residual priors trained on one geometry family often fail off-family, and proposes decoupling scale from structure to cross that scaling barrier.
WAVE-0 bridges vision-language-action policies and world action models using lightweight latent future prediction, avoiding costly video generation while making policies model the consequences of their actions.
Stripe's Monetize 2026 leaders describe AI companies moving past SaaS pricing to usage-based billing, spend controls, and new packaging and value metrics.
Tev is a recipe for scaling long-horizon agentic environments using dense progress rewards, aimed at the training data and supervision gap that keeps open-source models weak on tasks lasting hours to days.
Hybrid Reasoning Agents dynamically navigate between latent reasoning and text, keeping native tool access and transparent reasoning that pure latent approaches lack for acting in external environments.
A paper addresses a flaw in Rank-GRPO for LLM recommenders: picking one reply per recommendation thread leaves 85% of the advantage variance reflecting which reply was drawn rather than quality, and proposes removing that responder noise.
A new benchmark generates controllable, verifiable scenarios where vision-language agents receive contradictory instructions, outdated states or compete for permissions, testing how multi-agent computer-use systems handle conflict.
Anthropic said new safeguards made substantial progress reducing the effectiveness of distillation attacks on Claude, cutting data harvested by 90% to 99%, per reports.
Anthropic brought 10 free AI courses to Coursera, including training aimed at teachers. The courses cover using Claude and AI concepts and are free to access.
Anthropic's Frontier Red Team suggests some AI models may be months away from being capable enough to self-replicate, per The Verge.
Anthropic ran a live Q&A with the Claude Code team covering recent feature updates and usage tips.
Anthropic described building on the Claude platform, including Claude Fable 5 and patterns for orchestrating multiple models.
Gladia joined OVH Groupe to accelerate Europe's sovereign voice AI ambitions, combining Gladia's speech-to-text tech with OVHcloud infrastructure.
An Engadget how-to on getting higher frame rates from an older GPU by adjusting settings and forgoing newer features. A general PC tuning guide, not AI news.
Nvidia is joining other tech firms in an effort to standardize data center development, according to CoStar. Specific standards or partners were not detailed in the source.
The New York Times, Agence France-Presse and The Economist joined over 40 other publishers in a coalition to stop AI companies from using their content without permission or payment.
Supertab Connect launches on Fastly Compute, adding AI bot classification and RSL licensing so publishers can identify AI crawlers and license content to them.
Binance launched a free AI trading assistant worldwide, with an interface that adapts to users' experience levels.
Perplexity's free iPhone and iPad app, which answers questions with cited sources, released version 26.40.1 with bug fixes and improvements.
xAI launches an X subscription combining Grok, the Grok Bot agent, Cursor and X Premium, with docs covering Grok Bot security and FAQ.
Startup Fortune reports Anthropic ran a 26 billion parameter AI model across 15 inexpensive TV boxes, an unusual distributed-inference demonstration. No technical details were given in the source.
Google launched Gemini 4 Argon via Google AI, DeepMind and Google Developers, with posts on solving hard problems, robotics, safety and optimizing a 200,000-line Python codebase overnight. The Verge confirms the release.
Google AI Studio is now open to all users rather than developers only, as its head Logan Kilpatrick calls this a golden age for entrepreneurs, builders, and developers.
Nvidia will no longer ship new Windows 10 drivers for RTX GPUs, per reports, pushing users toward Windows 11 for continued driver support.
The UK warned that an autonomous AI-driven cyberattack is likely, following a reported breach involving OpenAI's Medicare systems.
Microsoft's AI chief commented on whether the custom Maia 300 chip reduces dependence on Nvidia, saying competition is good, in a Yahoo Finance interview.
Mistlabs launched Ambi, a real-time ambient AI product focused on AI-to-AI communication, following a $26 million funding round.
Search Engine Land reports Microsoft is testing a model called Decision-1 inside Copilot and Bing. Details on capabilities and availability remain unclear.
QwenGyre is an elastic online reinforcement learning framework from Alibaba's Qwen team for training extreme-long-horizon coding agents through unmodified black-box harnesses, targeting rollouts that run for hours.
Unitree's CEO claims GPT-6 performs well across several robots with factory pilots underway, and says GPT-7 could unlock mass deployment.
Britain's AI minister confirmed Monday the UK will retain early access to the newest US AI models for safety testing, following reports the White House had restricted such access.
Intel's D-series data center GPU, codenamed Crescent Island, is named D270 and supports up to 480 GB of LPDDR5X memory, positioning it as an Nvidia rival for AI workloads.