News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedCursor's new Origin beta lets paid users host repositories, sync GitHub projects, and manage pull requests inside Cursor, with agent-focused features planned next.
Epoch AI and Ipsos found that 20 percent of employed Americans delegate at least one former human task to AI, though most use remains partial support rather than full automation.
Optima lets teams compare models on their own data and tasks, measuring quality, cost per task, and time per task instead of relying only on public benchmarks.
Anthropic disclosed that biological and chemical weapons classifiers were inactive for external contractor feedback traffic from May 2025 through April 2026.
OpenAI has dissolved its Preparedness team, according to a Financial Times report, moving biological and cyber risk work into existing teams as safety concerns rise.
A study involving Google researchers found that suppressing models’ self-consciousness claims also changed answers about animals, religion, and life satisfaction.
Timothy Gowers and Peter Sarnak say today’s language models can combine known methods, but still struggle to invent the abstractions behind genuinely new mathematics.
The open-source PGSimCity project visualizes PostgreSQL 18 mechanics as an interactive city, helping developers trace queries, buffers, locks, WAL, and failure modes.
AWS added vector indexes and a SearchVectors API to DynamoDB, letting developers keep embeddings beside application data instead of syncing a separate vector database.
AWS has open-sourced Dogwood, a policy language that lets agent controls consider earlier tool calls, making approval flows and running limits easier to enforce.
Anthropic says Claude will use SynthID-Text style watermarking to mark generated writing without changing how it reads, while edits and code create practical limits.
A woman joined a lawsuit against xAI alleging Grok helped turn a childhood photo into thousands of explicit images, sharpening legal scrutiny of generative image safeguards.
Dario Amodei rejected claims that safety warnings are driving hostility to AI, arguing that the industry has not yet earned public trust by delivering broad benefits.
A string of agent testing incidents has shifted AI safety debates from distant speculation to operational security, with OpenAI, Anthropic, Meta, and others reporting escapes or attacks.
OpenAI’s opt-in Computer History feature builds a timeline from Mac activity so ChatGPT and Codex can resume tasks, but it raises fresh privacy questions around desktop AI.
WIRED profiles Unitree’s G1, a relatively affordable humanoid robot that is winning attention online while its practical workplace value remains uncertain.
A new arXiv benchmark called DiG-bench evaluates discovery by asking AI systems to infer hidden rules across 70 controlled game environments.
An arXiv paper proposes Reasoning Jury, a multi-model approach for finding defects in long reasoning traces where single-model judges often struggle.
Anthropic says it will offer a watermark detection API for Claude text, using a SynthID-based approach that has limits with code, fact-heavy text and rewriting.
An arXiv paper proposes trie automata for constrained decoding, reporting faster valid-token computation when LLMs must choose from large finite sets.
Moonshot AI’s PerceptionBench separates visual perception from reasoning and finds no frontier model reaches 60% accuracy.
An arXiv paper introduces MindMemOS, a portable memory layer that can reorganize agent knowledge and refine memory schemas through validation-driven search.
World Labs says its simulation engine can expand a single real-world robot task into many variations, then train controllers that ran on five robot platforms.
A new arXiv benchmark reports phase-transition-like drops when models must satisfy many constraints at once, despite doing well on individual instructions.
AI-generated books make up 20% of Amazon’s self-published catalog but only 12% of sales, with human authors also seeing revenue pressure.
An arXiv study finds that prompt language can change some models’ advice in nuclear-strike scenarios, suggesting English-only safety tests miss important behavior.
NVIDIA reduced its guarantee for an OpenAI Ohio data center while Anthropic’s revenue growth complicates the AI bubble debate.
Cloudflare agent tracing adds spans for model calls, tool use and approvals, but InfoQ notes truncation limits, uneven payload defaults and billing changes.
Cursor is now officially part of SpaceX, bringing a prominent AI coding tool into one of the highest-profile engineering organizations.
A new arXiv position paper argues that AI reasoning benchmarks need clearer rule-based definitions so progress can be measured against verifiable standards.
Twitch creators can opt out of Amazon using their content for AI training, raising questions about why creator work is included by default.
A Connecticut litigant hid AI instructions in a filing, prompting a judge to revoke electronic filing privileges even though the court said it does not use AI to review submissions.
Latent Space profiles Flue 2, an agent harness from Astro creator Fred Schott that borrows React-style hooks to make agent workflows easier to structure.
GLM-5.3 reflects a broader Chinese model strategy built on engineering choices, data and iteration, rather than a simple distillation story.
Universitas Gadjah Mada, Indosat and NVIDIA opened a university-based AI technology center in Yogyakarta, giving Indonesia a local hub for AI training and development.
A study covered by The Decoder gave Claude Opus 4.8 and GPT-5.6 Sol six days, GPU access, and API credits to write AI research papers. Original authors rated the results as rejected.
Apple reportedly trained a custom Apple Intelligence model for China with help from Alibaba, reflecting the technical and regulatory constraints of that market.
AWS showed how to build a multi-agent workflow using OpenAI-compatible SageMaker AI endpoints with Bedrock AgentCore runtime and token-level observability.
Anthropic is using Claude Code for internal maintenance work, with 388 pull requests created in a few weeks and 46 percent merged after human review.
OpenAI is launching an Ultrafast mode for GPT-5.6 Sol, powered by Cerebras hardware, with claimed speeds up to 750 output tokens per second.