OpenAI Unveils GPT-6 Astra for Enterprise with Advanced Reasoning
OpenAI Unveils GPT-6 Astra for Enterprise with Advanced Reasoning
OpenAI has officially launched GPT-6 Astra, its next-generation flagship model designed for enterprise work. The model features significant improvements in advanced reasoning and computer use capabilities, along with stronger writing and design judgment. GPT-6 Astra is built to handle complex, multi-step business tasks, assisting enterprise users with data analysis, report generation, and automated workflows, marking a key product update for OpenAI’s commercial market.
OpenAI Claims Navier-Stokes Singularity Find in 88 Hours Using Astra-next, 10K Agents
OpenAI reported that Astra-next, using roughly 10,000 agents and 130B tokens (over $40M in cost), found a Navier-Stokes singularity in 88 hours — a potential contender for the second Millennium Prize. This overshadowed Cognition’s $48B Series E, Mistral’s $24B Series D, and other major announcements, marking one of the most eventful days in AI history.
Paul Christiano Joins OpenAI Board Safety Committee
OpenAI announced that Paul Christiano, a leading expert in AI alignment, has joined its Foundation Board and its Safety and Security Committee. With deep experience in AI alignment, safety, and standards, Christiano’s addition will strengthen OpenAI’s capabilities in model safety and long-term risk governance, reflecting the company’s continued focus on safeguarding frontier AI systems.
Anthropic Scientist Warns of >10% AI Extinction Risk
A researcher at Anthropic stated in interviews with CBS News and the BBC that there is a more than 10% chance AI could lead to the extinction of humanity. The claim has triggered widespread debate within the AI and high-risk technology communities. The assessment likely reflects internal concerns about frontier model risks, alignment challenges, and arms race dynamics, underscoring significant worry within the AI safety field.
Vercel Adds Per-Project Password Protection for Pro Users at $20/Month
Vercel announced that Pro teams can now enable Password Protection for individual projects at $20 per project per month. When enabled, visitors must enter a password you set before viewing the project’s deployments. You can activate it via Security > Deployment Protection in the sidebar, or configure it through the API.
Vercel Makes Production Deployment Protection Free on All Plans
Vercel announced that its Authentication service can now protect all deployments in a project, including production, at no additional cost on every plan. Previously, protecting production domains required the $150-per-month Advanced Deployment Protection add-on. When enabled, visitors must sign in with a Vercel account to access.
Google DeepMind AI Recreates 70-Year Love Story Frame by Frame
In a short film titled ‘Love, Rendered,’ filmmakers collaborated with Google DeepMind to use AI to recreate frame by frame a couple’s 70-year love story, filling in the gaps where no visual record existed. The film showcases the potential of generative AI in emotional storytelling and visual restoration, offering a new way for the creative industry and documentary filmmaking to bring precious but unrecorded historical memories back to life.
SCAFFOLD Enables Self-Improving Web Agents via Recursive Parametric Skill Abstraction
A new arXiv paper proposes SCAFFOLD, a framework that enables web agents to avoid learning tasks in isolation and discarding procedural knowledge. Instead, it uses recursive parametric skill abstraction to continuously accumulate and reuse skills, targeting visually rich, long-horizon interfaces that change across websites.
Study: Safety Monitors Alone Don't Prevent Model Compliance
A new arXiv study challenges current evaluation methods for safety monitors in large language models by pointing out a fundamental flaw: they are assessed on recall against harmfulness labels without accounting for whether the model would actually comply with the flagged request. If a model refuses harmful behavior anyway, intercepting it is unnecessary. The researchers argue monitors should be evaluated in conjunction with model compliance to measure real-world harm prevention.
PAC Privacy: Calibrating Noise to Output Disagreement in Private LLMs
When language models are trained on private text, privacy leaks through generated outputs, not exposed weights. This paper introduces PAC-Private Autoregressive Generation, which calibrates noise to ensemble disagreement, offering a more efficient privacy protection alternative to methods like PMixED. The approach enables safer deployment of private models via API without drastically sacrificing output quality.
Reasoning-Aware Compression Cuts LLM Energy Loss
A new arXiv study introduces a ‘reasoning-aware compression’ framework designed to address the high energy cost of deploying Large Reasoning Models (LRMs). Traditional compression applies uniform quantization across components, risking damage to critical reasoning circuits. The new method identifies and protects vulnerable circuits during compression, reducing energy consumption while preserving reasoning capability, offering a new direction for efficient deployment of large AI models.
CONDUIT Framework Restores KV Cache for VLMs
A new arXiv paper introduces CONDUIT, a unified residual-stream restoration framework designed to address the failure of KV cache reuse in vision-language models (VLMs) when recurring visual content appears with slight variations. The framework enables efficient reuse of historical KV caches by restoring the residual stream, avoiding expensive re-encoding of visual prefixes. This improves inference efficiency in applications like video understanding and multi-turn dialogue.
Geiger: See Every AI Agent on Your Machine and Its Access
A developer has released an open-source tool called Geiger on Hacker News, allowing users to see every AI agent running on their personal computer and what files or resources these agents can access. The tool aims to provide transparent, secure visibility and management of local AI agents, helping users understand and control the scope and potential risks of AI usage. Such tools are increasingly important as local AI applications grow rapidly.
Jeff's Blog Defines 'Prolific AI Psychosis': Coherent Yet Context-Detached Outputs
Jeff’s Blog published Part 2 of “Defining AI Psychosis,” focusing on ‘Prolific AI Psychosis’ — a phenomenon where AI systems continuously generate well-structured, plausible-sounding outputs that detach from the user’s true intent. The post examines how this manifests in long conversations and complex tasks, sparking discussion on Hacker News (60 points, 43 comments).
Google Search Adds Football Features: Live Scores, Detailed Stats, Fantasy Picks
Google launched new football features in Search, letting users track live game feeds, explore detailed stats, and receive custom fantasy football recommendations directly in search results. The features arrive ahead of the new season, requiring no separate app — just search.
Claude Code v2.1.267 Adds MaxEffortLevel Setting Across All Providers
Claude Code released v2.1.267, adding a maxEffortLevel setting (top-level or per-model under modelSettings) that caps effort level across all providers including Bedrock, Vertex, and Foundry. Users can still select lower levels. The update also adds –system-prompt-snapshot off to render the system prompt fresh on every request.
OpenAI Codex 0.154.0 Adds GPT-6-Astra and Experimental Worktree Support
OpenAI Codex shipped version 0.154.0. New features include GPT-6-Astra now available in the model picker and Amazon Bedrock catalogs, plus experimental worktree support that lets users create isolated checkouts for new or forked sessions via –worktree or /worktree, with the ability to browse and resume them. The release also includes several 0.154.0-alpha pre-release iterations.
Data Scout: Targeted Web Crawling Solves Sparse Domain Pretraining Data
Building domain-specific pretraining corpora usually involves filtering massive archives like CommonCrawl, which fails for specialized domains with sparse content. Data Scout is a targeted web crawling approach that actively discovers relevant pages, offering a new path to construct high-quality datasets for niche domains.
Better AI Code Comment Detector
A technical article explores how to improve AI code comment detection. The author points out the limitations of current detectors and proposes improvements to more accurately identify AI-generated code comments, helping developers and managers understand the extent of AI involvement in their codebases. The post has sparked discussion in the developer community, offering a new tool direction for transparency and code review in AI-assisted programming.
OpenAI's Chris Lehane Calls for Stronger Safety Evidence and Policy Action
OpenAI’s policy chief Chris Lehane argues that as AI capabilities grow, stronger safety evidence, shared standards, and durable policy action are required. He believes the policy window remains open, urging the industry and regulators to act now rather than react after capabilities have advanced further.