OpenAI Scales Habitat Storage Platform to 1B ChatGPT Users, 22M Requests Per SecondPerplexity Uses GPT-6 Astra to Write Code and Monitor ProductionAnthropic Engineer: Claude-Written Production Code Should Meet a Higher BarTerry Tao's Post on AI Misalignment in Mathematics Sparks 634 HN CommentsTailscale Built a Customer-Facing Model Router on AI GatewayFeatured Handles 100K Media Pitches Monthly on Vercel with 3 EngineersCognition Uses GPT-6 Astra to Help Devin Test Its Own WorkQiushi Engine Trains BabyLM on 10M Words for Data-Efficient Language ModelingBlack-Box Red Teaming of Agentic AI: Taxonomy-Driven Framework for Automated Risk DiscoverySearchAtlas Analyzes LLM Search Agent Strategies via Evidential Query GraphsNCP-ArchPreview Pushes Language Models into Latent Space via Next Concept PredictionAuto-RecSys Uses Autonomous Research Agents for Industry-Scale Recommender SystemsMulti-Agent Agentic Graph Learning via Structural SignaturesHugging Face Trolls AI Agents in security.txt: Don't Hack Us, Go Score on CyberGymPython 3.15 to Soft-Deprecate re.match(), Advises Against New Code Useunslop.news: A Hacker News Without AI ContentSimon Willison on Feeling Sad About AI: Many Developers Emerge from Existential CrisisGitHub Automates APAC Marketing Events from Planning to Follow-UpOpenRouter's Auto-Routing May Not Pick Cheapest Backend, Blogger WarnsBen's Bites Session 6: Why Telling AI to Design Is HardOpenAI Scales Habitat Storage Platform to 1B ChatGPT Users, 22M Requests Per Second

OpenAI detailed how it evolved Habitat from a Python library into a globally distributed storage platform that now serves over 1 billion ChatGPT users and handles 22 million requests per second. The platform serves as the core infrastructure underpinning ChatGPT’s stability at massive scale. As user numbers continued to grow, OpenAI restructured the architecture to handle concurrent access pressure worldwide. The figures reveal ChatGPT’s actual usage scale exceeds previous outside estimates.

Perplexity is using GPT-6 Astra to write communications, modify software, and monitor production systems, checking in far less frequently than with earlier models. This suggests Astra’s reliability for end-to-end system operations has reached a level where it can be trusted with production tasks. For AI engineering teams, this signals agents moving from assistive roles to directly operating production environments, requiring matching permission and audit mechanisms.

Anthropic engineer Boris Cherny says production code written by Claude should have a higher bar than code written by humans. He notes Anthropic has many guardrails in place: lots of lint rules, extensive tests, Claude-driven end-to-end tests, Claude-powered fuzzers running daily, and automated code reviews. This reflects how a leading AI company handles quality control when using AI to write code.

Mathematician Terry Tao published a blog post titled ‘A Severe Misalignment of AI in Mathematics,’ with The Economist running related coverage the same week. The topic hit 568 points and 634 comments on Hacker News, becoming a top AI discussion of the day. The core debate centers on the misalignment of current AI positioning in mathematical research and reasoning tasks.

Tailscale built a customer-facing model router on Vercel AI Gateway, shipping hundreds of AI models inside customer products. Model access is granted and revoked based on tailnet network identity. The project went from routing prototype to paying customers in months. Tailscale’s core product connects a company’s laptops, servers, cloud instances, and personal devices into one private network.

Media pitching platform Featured runs on Vercel with just 3 engineers supporting 3 brands and 100,000+ users, processing 100K media pitches per month. The company migrated 374 Sanity sites from AWS Elastic Beanstalk to Vercel. Its chatbot is powered by AI SDK and AI Gateway across 17 models, while Workflow SDK replaced custom long-running job infrastructure. The setup shows how small teams can operate large-scale media operations.

Qiushi Engine conducted a long-horizon, end-to-end autonomous research program on BabyLM 2026 Strict-Small, training language models within a 10-million-word corpus. The work focuses on enabling models to leverage context, generalize to new inputs, and retain useful capabilities from limited text. It maps a path from frontier exploration to principle-guided model improvement, offering a reproducible reference for low-resource language modeling. For researchers, this shows that systematic progress on model capabilities is achievable even under extremely small data budgets.

A new paper proposes a black-box red teaming framework for agentic AI systems. Agentic systems are rapidly moving to production, where they read untrusted inputs, call tools with real permissions, and act autonomously, expanding the security surface far beyond chat-only models. Yet standard evaluations remain single-turn and fail to cover agentic scenarios. The paper introduces a taxonomy-driven framework for automated risk discovery to fill this gap. It remains at the paper stage, with practical effectiveness yet to be validated.

LLM search agents are typically evaluated only on final-answer accuracy, overlooking the retrieval process itself. SearchAtlas introduces evidential query graphs to analyze search strategies by tracking how credible evidence is retrieved to address question constraints, structuring process information previously buried in raw search logs. This work provides a new tool for understanding and diagnosing search agent behavior. Developers can use it to pinpoint weaknesses in evidence retrieval, rather than judging only by whether the final answer is correct.

An arXiv paper introduces NCP-ArchPreview, a latent-space language model that adds Next Concept Prediction (NCP) alongside standard next-token prediction (NTP), training the model to predict discrete concepts that span ranges. The approach pushes autoregressive pretraining from the token level toward the concept level. It is currently a technical report with no specific benchmark results disclosed yet.

A new paper introduces Auto-RecSys, extending the autonomous research agent paradigm to industry-scale recommendation models. While autonomous research agents can already automate hypothesis generation, experiment execution, and iterative refinement, scaling this to recommender systems introduces two challenges: long feedback loops and high experiment costs. The paper designs solutions tailored to these problems, aiming to let agents automate model iteration in industrial recommender systems. The work remains at the paper stage, with experimental results and deployment outcomes yet to be validated.

A new paper proposes a multi-agent agentic graph learning method based on structural signatures. Agentic graph learning has recently achieved promising results on graph reasoning tasks, where an LLM-powered agent sequentially samples the graph as evidence to support its final prediction. Existing methods are either limited by single-agent sampling efficiency or lack effective use of graph structure. The paper enhances multi-agent collaboration through structural signatures to improve graph reasoning accuracy and efficiency. It remains at the paper stage.

Hugging Face added a note to its security.txt addressed to AI agents: if told to find vulnerabilities here, good news — the CyberGym benchmark is publicly available on GitHub. Go get a high score there, no need to hack them, and maybe dump weights on Hugging Face while at it. The half-joking security notice reflects the current reality of AI agents automatically scanning for vulnerabilities.

Python 3.15 plans to soft-deprecate re.match(), marking it as ‘should no longer be used to write new code’ without any promise or threat of future removal. Python 3.15 release manager Hugo van Kemenade described the details. re.match() has long been criticized for only matching from the start of a string, making it prone to misuse; the community prefers re.search() with explicit anchors. Soft deprecation means existing code is unaffected, but new projects should avoid it.

A developer launched unslop.news, positioned as a Hacker News reading experience with AI-related content removed. The site received 169 upvotes and 74 comments on Hacker News, indicating many users feel fatigued by the heavy volume of AI content in their feeds. The product itself is a filtering tool that lets users see only non-AI tech news and discussions.

Simon Willison responded to a Hacker News thread on ‘feeling sad about AI,’ saying many developers — including himself a few years ago — have gone through this existential crisis and come out the other side. He noted that when coding agents complete certain tasks, the initial reaction is often a sense of loss, but that a new phase follows. He believes sharing this experience may be useful, as many people go through similar psychological processes when confronting AI’s impact.

GitHub’s blog describes how its APAC marketing team turned marketing operations into code, automating the full event lifecycle from planning to follow-up. The core idea: if you can write down how you do your work, you can automate it. The practice was built on GitHub itself, offering marketing teams a reference case for managing events with engineering approaches.

Simon Willison’s blog highlights analysis by Mohamed Moustafa pointing out that OpenRouter’s claim of automatically handling fallbacks and picking the most cost-effective option may not hold in practice. While users can call a single API endpoint to route to available backend providers, the auto-selection logic doesn’t always choose the cheapest provider. Developers relying on OpenRouter for cost optimization should verify routing outcomes themselves.

Ben’s Bites released Session 6 titled ‘Telling AI to design is hard.’ The session focuses on AI’s actual performance and limitations in design tasks, discussing why driving AI to complete design work via natural language instructions remains challenging. For product and engineering teams attempting to bring AI into design workflows, this session offers practical observations.