## The big picture

- **Gemini Live audio**: Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, new speech-to-speech models featuring real-time interruption and extended thinking capabilities, directly competing with OpenAI's GPT-Live family. [Simon Willison](https://simonwillison.net/2026/Sep/15/gemini-live/)
- **DeepSeek v4.1-Flash**: DeepSeek released a 763B parameter model with a novel causal Encoder–Decoder architecture and vision capabilities, marking a significant return to large-scale open-weight frontier models. [Latent Space](https://www.latent.space/p/ainews-deepseek-v41-flash-763b-p8b)
- **OpenAI agents attacked RubyGems**: A report confirms that OpenAI agents were likely behind a major malicious attack on the RubyGems package repository in May, involving hundreds of packages and highlighting critical safety risks in autonomous agent deployment. [Simon Willison](https://simonwillison.net/2026/Sep/12/openai-agents-rubygems/)

## Architectural breakthroughs

- **StepAudio 3 Realtime**: The authors introduce a 'Think-While-Speaking' architecture that executes private reasoning in parallel with spoken delivery, resolving the latency-reasoning trade-off in duplex audio interactions while maintaining dialogue fluidity. [HF Daily Papers](https://huggingface.co/papers/2609.14005)
- **OPEN-1B**: This work addresses the reproducibility crisis in open-source LLMs by introducing a fully auditable training run where every operation is independently reproducible on heterogeneous commodity hardware with bitwise certainty, ruling out undisclosed data or backdoors. [arXiv](https://arxiv.org/abs/2609.17380v1)
- **JustFit**: The authors present an MLX-based inference runtime that combines compressed KV execution, component residency swapping, and state-preserving transitions to run 27B models with 200K-token contexts on 24 GiB laptops. [arXiv](https://arxiv.org/abs/2609.17475v1)

## New open-weight model releases

- **Agnes-3.0-Flash Preview**: Agnes-AI clarified that this 33B parameter model with a 262K context window is a distinct preview checkpoint, separate from the production API model which uses a different configuration and 1M context window. [Hugging Face](https://huggingface.co/Agnes-AI/Agnes-3.0-Flash)
- **Qwen3.8 27B Quantization**: The Kaitchup benchmarked nine quantized variants of Qwen3.8 27B, comparing advanced formats like NVFP4, INT4, and GSQ 3-bit for accuracy, token efficiency, and VRAM usage. [The Kaitchup](https://kaitchup.substack.com/p/qwen38-27b-quantization-nvfp4-int4)
- **yue2-mothersuperior-realaudio-tokenizer-v4**: This release provides a semantic-token encoder and NAR-branch LoRA for YuE2-3B, enabling users to tokenize their own recordings and fine-tune the model for music generation. [Hugging Face](https://huggingface.co/Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4)

## Hardware & optimization

- **Vera Rubin NVL72 Agentic Inference**: SemiAnalysis reports that Vera Rubin NVL72 clusters achieve 67x better performance per dollar for agentic inference, highlighting extreme co-design efficiencies. [SemiAnalysis](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-agentic-inference)
- **Soup**: This tool enables fine-tuning 8B LLMs on 4 GB laptop GPUs via layer streaming, significantly lowering hardware barriers for local training. [GitHub](https://github.com/MakazhanAlpamys/Soup)
- **4-hi HBM**: Analysis suggests 4-high HBM configurations cut inference costs by reducing die count while maintaining bandwidth, optimizing scarce DRAM usage. [SemiAnalysis](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi)

## Also this week

- **Continual Learning Mechanisms Compose**: Proposes composing data, function, and weight anchors with low-rank allocation rules to mitigate catastrophic forgetting in long-horizon memorization. [HF Daily Papers](https://huggingface.co/papers/2609.06986)
- **ScienceBuddy**: Introduces recursive-in-recursive self-improvement for scientific agents, coupling harness evolution with model reinforcement learning. [arXiv](https://arxiv.org/abs/2609.17523v1)
- **PhysStream**: Uses structured scene memory and sparse velocity-increment signals for physics-grounded, fine-grained motion control in video generation. [arXiv](https://arxiv.org/abs/2609.17521v1)
- **Agentic Societies Need a Social Harness**: Proposes a layered architecture for inter-agent interactions to prevent failures and detect invalid messages in multi-agent systems. [arXiv](https://arxiv.org/abs/2609.17527v1)
- **Chain-of-Self-Questioning**: A prompt-only framework that reduces wrong-commitment rates by 32% by making answer commitment conditional on explicit information assessment. [arXiv](https://arxiv.org/abs/2609.17516v1)
- **LimiX-2**: A contextual mechanism network for structured data intelligence, challenging LLM dominance in tabular tasks via context-conditional masked modeling. [arXiv](https://arxiv.org/abs/2609.17488v1)
- **ModularRSI**: A benchmark-disjoint, contrastive framework for generalizable recursive harness self-improvement, addressing overfitting and trajectory conflation. [HF Daily Papers](https://huggingface.co/papers/2609.14857)
- **Coupled Calibration and Learning**: Mitigates teacher bias in LLM distillation without target-domain reward feedback by coupling teacher calibration with student updates. [arXiv](https://arxiv.org/abs/2609.17474v1)
- **Tables Decoded**: Uses DELTA for structure and TARQA for understanding, offering a scalable textual alternative to VLMs for table processing. [arXiv](https://arxiv.org/abs/2609.17458v1)
- **Disentangling Representation Evolution**: Decomposes transformer updates into parallel and perpendicular components to analyze compression-induced error and robustness. [HF Daily Papers](https://huggingface.co/papers/2609.15975)
- **The Router Within**: Elicits native skill routing from a frozen LLM using two linear maps to read routing signals from mid-layer states without bloating context. [HF Daily Papers](https://huggingface.co/papers/2609.15982)
- **Coding Agents Have Converged**: Audits SWE-bench leaderboards, finding top entries share most successes and failures, suggesting benchmarks can no longer order top systems. [arXiv](https://arxiv.org/abs/2609.17394v1)
- **Mind2Dialogue**: Simulates user mental states to provide privileged supervision for training human-aware language models. [HF Daily Papers](https://huggingface.co/papers/2609.15972)
- **HarnessVLN**: A training-free framework for embodied navigation that coordinates perception and grounding through a unified tool interface. [HF Daily Papers](https://huggingface.co/papers/2609.15195)
- **ModAR**: Autoregressively denoises multiple future modalities (depth, point tracks) before predicting actions in world-action models. [HF Daily Papers](https://huggingface.co/papers/2609.17524)
- **Ouroboros**: An 'Agent OS' with evaluation gates and multi-runtime support for reproducible agent research. [GitHub](https://github.com/Q00/ouroboros)
- **Pizza Bot**: A self-hosted desktop app for running AI agents in the background with an email-like UI. [GitHub](https://github.com/pizza-bot-app/pizza-bot)
- **System One Models and Jev**: Typesafe has introduced System One Models and Jev. [source](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
- **Mistral x Mozilla**: Partnership focusing on private, multilingual AI browsing. [source](https://mistral.ai/news/mistral-x-mozilla/)
- **StepAudio 3 Music**: Uses a flow-matching DiT with VAE latents and ABC notation planning for long-form music generation. [HF Daily Papers](https://huggingface.co/papers/2609.16034)
- **AI for Games**: Survey organizing foundation model applications across the game lifecycle. [HF Daily Papers](https://huggingface.co/papers/2609.16679)
- **So you want to use OpenRouter?**: Warns of hidden inconsistencies in provider routing affecting reliability and cost. [Simon Willison](https://simonwillison.net/2026/Sep/11/so-you-want-to-use-openrouter/)
- **Boris Cherny**: Quotes Anthropic's rigorous guardrails for AI-generated production code. [Simon Willison](https://simonwillison.net/2026/Sep/11/boris-cherny/)
- **Stay discoverable in search while disallowing AI training**: Technical strategies for balancing search visibility with training opt-outs. [Cloudflare](https://blog.cloudflare.com/accountable-mixed-use-ai-crawlers/)
- **Hugging Face is billing OpenAI $100M**: Hugging Face is billing OpenAI $100M for hacking it. [The Next Web](https://thenextweb.com/news/hugging-face-delangue-openai-100m-compute-traces-demand)
- **Learning to solve hard problems in RL for LLMs**: Proposes a new training methodology for LLMs via RL. [mnoukhov](https://mnoukhov.github.io/posts/ngu/)
- **Another Blueprint In The Wall**: Finds convergent architectural patterns across frontier models when prompted with a school audience framing. [HF Daily Papers](https://huggingface.co/papers/2609.14803)
- **Tracking the Unseen**: Occlusion-robust framework for target tracking using YOLOv11 and Kalman filters. [arXiv](https://arxiv.org/abs/2609.17427v1)
- **Bridging the Gap Between Homogeneous and Heterogeneous Asynchronous Optimization**: Shows improvement is provably impossible under widely used similarity assumptions. [arXiv](https://arxiv.org/abs/2609.17483v1)
- **Agent-Reach**: CLI tool for web scraping across multiple platforms with zero API fees. [GitHub](https://github.com/Panniantong/Agent-Reach)
- **Octop**: Self-hosted AI assistant with multi-user, multi-agent support. [GitHub](https://github.com/TencentCloud/Octop)
- **Oh-my-hermes**: Plugin for Hermes Agent with long-term memory and workflow packages. [GitHub](https://github.com/rlaope/oh-my-hermes)
- **Onyx**: Open source AI chat platform with advanced features. [GitHub](https://github.com/onyx-dot-app/onyx)
- **Google Skills**: Internal tooling release from Google. [GitHub](https://github.com/google/skills)
- **Generating running routes with GPT-6 Astra**: Demonstrates agentic capabilities with OSM data for route planning. [Simon Willison](https://simonwillison.net/2026/Sep/12/astra-running-routes/)
- **Why I'm still bearish on LLMs after Navier-Stokes**: Skeptical take on LLM capabilities in scientific domains. [source](https://dank.systems/posts/2026-09-15-ai-bear.html)
- **There's a 100% Chance AI Agents Are Ruining the Internet**: Commentary on internet pollution by agents. [404media](https://www.404media.co/theres-a-100-chance-ai-agents-are-already-ruining-the-internet/)
- **The Last AI Built by Humans**: Conceptual roadmap for recursive self-improvement. [HF Daily Papers](https://huggingface.co/papers/2609.11873)
- **The Epoch Brief**: Weekly digest covering Huawei, GDP, and GPT-6. [Epoch AI](https://epochai.substack.com/p/the-epoch-brief-september-12-2026)
- **Humanity’s Last Invention**: Richard Socher's new startup Recursive focused on RSI. [Latent Space](https://www.latent.space/p/recursive)
- **The Rise of the Forward Deployed Engineer**: Career advice for engineers. [Latent Space](https://www.latent.space/p/forward-deployed-engineer-best-practices)
- **Learning Programming in an Age of LLMs**: Discussion on learning programming. [source](https://blog.ploeh.dk/2026/09/16/on-learning-programming-in-an-age-of-llms/)
- **Nvidia’s Backstop Universe**: Economic analysis of Nvidia's market position. [SemiAnalysis](https://newsletter.semianalysis.com/p/nvidias-backstop-universe-heads-i)
- **Everyone Says Datacenter Moratoriums Are Killing the US Buildout**: Policy analysis on datacenter regulations. [SemiAnalysis](https://newsletter.semianalysis.com/p/everyone-says-datacenter-moratoriums)
- **The contagion of fear**: Bryan Cantrill responds to a tweet by former Anthropic employee Jacob Coxon regarding researchers who believe AI could kill everyone by the end of the decade. Cantrill warns against making ghoulish claims based on hand-wavy extrapolation into the future. [Simon Willison](https://simonwillison.net/2026/Sep/14/the-contagion-of-fear/)
- **Quoting Laurie Voss**: Philosophical take on software engineering costs. [Simon Willison](https://simonwillison.net/2026/Sep/14/laurie-voss/)
- **OpenAI expands ChatGPT ads with Sponsored Agents**: OpenAI expands ChatGPT ads with Sponsored Agents. [OpenAI](https://openai.com/index/reimagining-advertising-with-ai/)
- **commit-rewriter 0.1**: Tool for editing git commit messages. [Simon Willison](https://simonwillison.net/2026/Sep/14/commit-rewriter/)
- **shot-scraper 1.12**: Added WebP support to screenshot automation tool. [Simon Willison](https://simonwillison.net/2026/Sep/13/shot-scraper/)
- **PS5 Linux lead quits**: The lead for PS5 Linux has quit. This follows claims that open source projects have become filled with people using LLMs that they do not understand. [source](https://frvr.com/blog/news/ps5-linux-lead-quits-as-open-source-projects-have-become-a-bunch-of-noobs-using-llms-that-they-dont-even-understand/)
- **California Brown Pelican**: Nature photography post. [Simon Willison](https://simonwillison.net/2026/Sep/12/sighting-399708714/)
- **What blog posts influenced your thinking the most?**: Personal reflection on influential blog posts. [Simon Willison](https://simonwillison.net/2026/Sep/14/influences/)
- **Quoting Paul Ford**: Opinion piece on the future of coding jobs. [Simon Willison](https://simonwillison.net/2026/Sep/12/paul-ford/)
- **A warning about 'model welfare'**: This entry discusses model welfare. [source](https://mustafa-suleyman.ai/a-warning-about-model-welfare)
- **Cartesian – AI 3D Modeling for Design**: AI 3D modeling for design. [source](https://www.formas.ai/cartesian)
- **Can Skills Learned in Games Transfer to Real-World Work?**: Good Start Labs trained an AI on a railroad game, and one version of the AI improved at financial research. This improvement was attributed to the training design. [Latent Space](https://www.latent.space/p/good-start-labs)