NVIDIA's $12B Poolside Deal: Founders Stay, Employees ExitOra on Vercel Benchmarks Major AI Agents on Live WebsitesSimile AI's Joon Sung Park Builds 8 Billion Digital Twins with $80M VisionScan Rare Books Before AI Companies Destroy ThemAI Homework Help Boosts Scores, But Exam Scores Drop: StudyCross-Lingual Safety Gap: LLM Safety Filters Fail in Non-English LanguagesStop Making TUIs, Build Native GUIs NowDeveloper: I'm Becoming AI-BlindGoogle DeepMind Partners with Game Studios for Breakthrough AI GameplayTurning LLM Hallucinations into Scientific Hypotheses: A Multi-Agent ApproachDeltaMomentum: Anisotropic Momentum via Delta RuleMechanistic Tomography: New Method for InterpretabilityImproved Confidence Estimates for Black-Box LLMsMatt Webb: I Used ChatGPT as a Tutor to Learn MathVercel Deployment Storage Enables Instant Rollback for Production Deploysllm-openrouter 0.7 Adds LLM 0.32 Support, New ToolsBen's Bites Founder Shares How the Product Was BuiltNew Benchmark: MLLMs Under System Messagesllm 0.32.1 Fixes Install Breakage Caused by OpenAI Dependency ChangesOpenAI Codex Ships 0.150.0-alpha.6 After Rapid IterationsNVIDIA's $12B Poolside Deal: Founders Stay, Employees Exit

NVIDIA completed a $12 billion reverse acqui-hire of AI startup Poolside. Founders will stay and receive $1 billion, while employees get $6 billion. NVIDIA plans to leverage Poolside’s tech and team to build Infraco, a 7GW neocloud infrastructure project. The deal combines talent acquisition with infrastructure expansion, highlighting the intense competition for AI talent and compute resources.

Vercel released Ora, a platform that benchmarks major AI agents side-by-side on live websites. Agents are tasked with signing up for products, integrating, and making payments. The platform is maintained by a 16-person engineering team handling hundreds of commits daily. Developers can use Ora to compare agent performance in real-world scenarios.

Simile AI CEO Joon Sung Park discusses his journey from the viral Generative Agents project to founding a company aimed at creating 8 billion digital twins for every living human. He explains why the work has shifted from fun exploration to serious business. Digital twin technology could power personalized services and decision simulation across industries.

A post on Anna’s Archive warns that AI companies are physically destroying physical books (possibly by cutting spines for scanning), leading to permanent loss of rare books. The article calls for immediate digitization to prevent irreparable loss of human knowledge heritage. The post received 701 points on HN.

A study reveals that students using AI for homework saw improved assignment scores but lower exam scores. The research sparked discussion on Hacker News with 165 points. Findings suggest AI tools may help complete assignments without promoting knowledge retention, potentially harming learning outcomes.

Thomas Ptacek argues developers should stop building TUIs for even small tools and instead create native GUIs. He contends that coding agents have made building usable GUIs nearly cost-free. He used ‘vibe-coding’ to build his own macOS menu bar bandwidth and GPU monitor. The post sparked debate in the developer community about the future of tool interactions.

A popular Hacker News post (252 points, 258 comments) describes the author’s growing numbness to AI-generated content, dubbing it ‘AI-blind.’ The piece reflects on how developers can balance efficiency and vigilance as AI output becomes the norm. The discussion focuses on over-reliance on AI tools and its potential impact on critical judgment.

Google DeepMind published a blog post reviewing 15 years of AI research in games, from early Atari work to partnering with game studios on AI prototypes for titles like EVE Online. The post highlights the importance of games as AI testbeds and how DeepMind applies research to real gaming scenarios, advancing decision-making in complex environments.

A new arXiv paper (2608.19206) argues that LLMs’ heavy alignment to suppress hallucinations, while reducing misinformation, also limits their combinatorial creativity. To address this, the authors propose a multi-agent architecture that repurposes speculative model outputs (hallucinations) into testable scientific hypotheses. Instead of correcting errors, the system structures and filters divergent thinking for scientific validation. This offers a new path for AI in research—allowing scientists and developers to mine unrestricted model outputs for novel ideas and test them empirically, rather than relying solely on factual answers.

A new paper introduces DeltaMomentum, a Key-Value based optimizer that uses the Delta Rule for momentum updates. Unlike traditional optimizers (like Adam) that apply a fixed EMA across all gradient directions, DeltaMomentum dynamically adjusts updates to account for anisotropy in training data. The approach could better adapt to non-uniform data distributions in deep network training and potentially improve training efficiency and model performance.

A new paper introduces ‘Mechanistic Tomography,’ a method for control-oriented interpretability in large models. The approach involves designing measurements to identify internal states, component effects, and interactions. By combining techniques like patching, gradients, and Hessian-vector products, the method offers more precise tools for controlling model behavior, potentially aiding researchers in understanding and steering AI systems.

A new paper introduces an improved method for estimating confidence in black-box LLMs. Current methods, like verbalized confidence or multiple generations, are often zero-shot and lack accurate uncertainty quantification. The new approach refines confidence scores through iterative generations, providing users with a better gauge of output reliability—a practical step for safer LLM deployment.

Simon Willison quotes Matt Webb: after releasing v1.0, he needed to learn matrix rotations himself. Instead of asking ChatGPT to write the code, he used it as a patient, interactive tutor. Through this, he finally succeeded in learning a concept that books and mathematician friends couldn’t teach. The example highlights LLMs’ potential as personalized educational tools.

Vercel launched Deployment Storage, which keeps files from every deployment including pages, functions, and assets, enabling instant rollback in seconds. If a production deploy ships with issues, teams can revert to previous versions immediately. The feature removes manual backup hassle and provides a safety net for rapid iteration.

llm-openrouter 0.7 is out, now compatible with LLM 0.32, with improved support for reasoning models via OpenRouter. Models now use OpenRouter’s Responses API. Three new server-side tools were added: Shell, WebFetch, and one more. Developers can now more easily use reasoning models from OpenRouter within LLM.

Ben, founder of Ben’s Bites, shares in session three how he built the AI newsletter product. The talk covers the journey from idea to execution, tech stack choices, and content generation workflows. Valuable for independent developers interested in running AI-powered content products.

A new paper proposes a benchmark for evaluating Multimodal LLMs under system messages. Existing benchmarks either test text-only constraints or embed them in user turns, ignoring system-message impact. This research specifically assesses compliance, capability, and conflict in MLLMs under system instructions, revealing potential issues in production deployments.

llm command-line tool released version 0.32.1, fixing an installation breakage caused by the OpenAI Python library discontinuing httpx usage. Previously, llm depended on httpx through a transitive openai dependency, which broke fresh installs. This dot-release temporarily resolves the issue by pinning the openai version. Developers should upgrade to restore normal installation.

OpenAI Codex released version 0.150.0-alpha.6, following rapid iterations including 0.150.0-alpha.5 and rust-v0.150.0-alpha.4. The frequent alpha releases target code generation and execution for developers. Users can try the latest alpha for new features, though stability risks exist with pre-release builds.