AI Daily · 2026-09-10

OpenAI led the day with GPT-6 Astra, a work-focused model that can operate everyday apps without APIs, alongside a policy push and Paul Christiano joining its foundation board. DeepSeek introduced DeepSeek-V4.1-Flash, a multimodal MoE model with a 552B backbone, 1M-token context, and heavily compressed KV cache. Infrastructure and developer updates continued: NVIDIA expanded rack-scale XPU interconnects and Australian data center capacity; Hugging Face rebuilt AUTOMATIC1111 as a workflow canvas, IBM released a commercial-friendly time-series model, LangChain added managed credentials, and Replit/Databricks became generally available with Lakebase support. In consumer and security news, Google Search added live football features and a WeChat zero-click worm demo underscored AI-accelerated exploit development.

North America · First-hand

OpenAI

⭐⭐⭐⭐⭐ [Model Release] GPT-6 Astra: The next generation in intelligence for work

OpenAI released GPT‑6 Astra, which it describes as its most capable model built for work, now available in ChatGPT Work, Codex, and the API. The company says Astra is state-of-the-art on computer use, browsing, professional work, software engineering, cybersecurity, and science. The model can work through the same applications people use daily even when those applications have no API, so businesses can adopt it within existing workflows without extensive preparation or engineering. Partners including Cognition, Databricks, Hebbia, Box, Figma, and Thomson Reuters Labs provided evaluation feedback, with Databricks reporting new state-of-the-art results on its OfficeQA Pro and Pro V2 benchmarks. Internally, OpenAI used Astra and Codex to turn three hours of multicamera footage into a video, and its engineering team used Astra to find a memory-allocation bottleneck in a test environment, switching allocators to achieve 25x lower turn latency with roughly 30 percent higher peak memory use.

⭐⭐⭐ [Opinion] The AI policy window is open. We need to act.

OpenAI Chief Global Affairs Officer Chris Lehane argues that AI capabilities have entered a new chapter that demands a new chapter for AI policy, with a bias toward meaningful action over policy perfection. OpenAI says it wants to work with Congress on mandatory, capability-based national AI safety regulation and, until Congress acts, will keep supporting state legislation; it announced support for four California bills: SB 813 on infrastructure for independent safety assessments, AB 1405 on AI-auditor standards, SB 1119 on protections for young people, and AB 1864 on safeguards against AI-enabled biological threats. The company also says it will work with other frontier labs to advance frontier AI standards as a voluntary effort, and will advocate for compatible international approaches to measuring capabilities, managing risk, preserving human control, and deciding when development should slow or stop. It cites Chief Scientist Jakub Pachocki's call for extreme caution over the rapid rise of machine intelligence and the potential of recursive self-improvement, and notes strengthened monitoring, alignment and security safeguards across the model-development lifecycle, including universal monitoring of full trajectories (including chains of thought) for Astra and a mandatory alignment-evaluation gate before broader internal deployment. OpenAI says it will slow or stop development or deployment of systems it cannot sufficiently safeguard when proceeding would pose an unacceptable safety risk.

⭐⭐ [Other] Paul Christiano joins OpenAI Foundation Board

OpenAI announced the appointment of Paul Christiano to the OpenAI Foundation Board, where he will also be a non-voting observer on the OpenAI Group PBC Board. Christiano will join the Foundation Board's Safety and Security Committee (SSC), working with its chair, Zico Kolter; the SSC provides governance over safety and security practices across all of OpenAI, including OpenAI Group PBC. He currently serves as a Senior Tech Advisor at the Center for AI Standards and Innovation (CAISI) within NIST at the U.S. Department of Commerce, where he has worked on evaluating frontier AI models, including capabilities with national security implications, and on mitigating related risks. He is the founder of the Alignment Research Center (ARC) and from 2017 to 2021 led alignment research at OpenAI, contributing foundational work on RLHF. Bret Taylor said Christiano's experience will strengthen the Board's oversight and the SSC's work, while Christiano said AI capabilities have advanced rapidly and alignment remains a difficult technical problem.

Google

⭐⭐ [Product Update] Get ready for the game with new football features in Search

Google is rolling out new football features in Search. A new Live Game Feed lets users search an in-progress game to follow it in real time, with an overall recap, a play-by-play timeline, top social commentary, key video highlights and AI-powered insights; it is available on mobile in the U.S. in English, with collegiate team support and broader global availability coming later this month. Search also adds a matchup carousel for checking other league scores, plus more detailed league and player stats such as passing touchdowns, rushing yards, sacks, fumbles and yards after catch, with championship favorites and playoff brackets coming soon — all available globally on mobile. Users can additionally link their Yahoo Fantasy or Sleeper accounts with Search to get tailored insights through AI Mode, including start/sit recommendations and waiver wire targets, now available in the U.S. in English.

East Asia · First-hand

DeepSeek

⭐⭐⭐⭐ [Model Release] deepseek-ai/DeepSeek-V4.1-Flash

DeepSeek released DeepSeek-V4.1-Flash on HuggingFace, a natively multimodal Mixture-of-Experts model with a 552B backbone and support for contexts up to one million tokens, under an MIT license. It adopts a Causal Encoder-Decoder architecture of 20 causal encoder layers followed by 20 decoder layers, activating only 8B parameters per token during prefill and 16B during decode. Through SWA Bounded Replay, Compressed Sparse Attention 2, and FP4 main KV caching (E2M1 format with one E4M3 scale per 16 channels), the global KV cache footprint drops to 890 bytes per token, roughly one quarter that of DeepSeek-V4-Flash; other components include Single-Pass mHC, Engram conditional memory (196B parameters), and DSpark speculative decoding, with 1 shared and 384 routed experts per MoE layer and 6 routed experts activated per token. On the vision side, a from-scratch DeepSeek-ViT encoder and a two-layer MLP projector turn images into visual embeddings processed jointly with text embeddings from the start of language-model pre-training. The model was pre-trained from scratch on a 45T-token multimodal corpus (sparse attention trained at 64K sequence length, context extended to 1M at 34T tokens), with post-training following the standard SFT to RL to on-policy distillation recipe, where substantive changes sit in the data pipeline via large-scale automated synthesis of agent tasks and environments; it also supports a continuously controllable reasoning effort setting from 1 to 100.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM released Granite Time Series PatchTST-FM-r2, the latest model in the Granite TSFM family and a new version of PatchTST-FM-r1. It combines an updated architecture and a larger pretraining corpus with probabilistic forecasting and missing-value imputation, using roughly 385M parameters, a context length of up to 8,192, flexible forecast lengths, and a 99-quantile prediction head; its backbone is built from conformer blocks that combine multi-head self-attention with temporal convolution. The model is dual-licensed under Apache-2.0 and OpenMDW-1.0, letting users select either license. As of September 8, 2026, it ranks second for both CRPS and MASE among replicable, zero-shot models on the GIFT-Eval benchmark, and is the highest-performing model in that category with permissive, commercial-friendly licensing; the post reports a geometric-mean CRPS of 0.467 and MASE of 0.6846. Model weights, architecture, inference pipeline, and code to reproduce the benchmark results are all available.

Product Update

⭐⭐ [Product Update] Rebuilding AUTOMATIC1111 with Gradio Workflow

A Hugging Face blog post introduces Workflow1111, which rebuilds most of AUTOMATIC1111's stable-diffusion-webui feature set as a single workflow canvas. The canvas is a graph of eleven media pipelines built from seventy-three nodes, covering text-to-image, hi-resolution fix, image-to-image, prompt-matrix grids, VLM interrogate, detection-to-inpaint masks, ControlNet-style annotators, background removal, PNG Info storing, and image-to-video. Users sign in with a Hugging Face account or provide an access token to run any pipeline, with model calls using their own quota, and can also duplicate the Space to rewire it. Each node wraps one of four operator kinds: fn (a Python function), model (a model called through InferenceClient), space (another Gradio Space), and dataset (a row from a Hub dataset). The post walks through each pipeline, noting for example that text-to-image first passes the prompt through a prompt-builder fn node that appends the selected style preset and cleans the text, then into a model node calling the checkpoint, and a post-process fn node writes the generation parameters into the PNG metadata.

⭐⭐ [Product Update] Connections: managed credentials and per-caller identity for Managed Deep Agents

LangChain introduced Connections for Managed Deep Agents, moving credentials out of project code and .env into the LangSmith workspace, where tools read them at run time by slug through connections.get(), so credentials can be rotated or revoked without code changes or redeploys. A connection has two independent axes: owner (agent or caller) and credential type (static secret or OAuth grant). An agent-owned secret is shared by every caller and suits capabilities that do not differ per person, such as web search; a user-owned credential resolves per person at run time, so the agent's actions carry that user's identity. Managed Deep Agents handles the OAuth authorization round-trip, with no callback route, token store, refresh logic or consent screen in the project. The connections catalog ships with GitHub and 22 other services, and the feature is available in Managed Deep Agents v0.7.0+.

⭐⭐ [Product Update] Replit | Databricks Integration is Now Generally Available with Native Lakebase Support

The Replit–Databricks integration is now generally available, adding native Databricks Lakebase support. It builds on the public preview announced this past June. In the described setup, Replit accelerates application creation while Databricks hosts governed live enterprise data and provides access and security, with Lakebase offering a managed database for storing and updating application data. Native Lakebase support lets teams build full-stack apps that combine live warehouse data with data created and updated through the app, and when an app is ready to deploy, Replit Agent automatically provisions its Lakebase database without manual setup. A new automated preview deploys feature creates a separate preview environment that isolates test data from live business data in Databricks.

⭐⭐ [Product Update] .blend URL Viewer

Simon Willison released a .blend URL Viewer that renders Blender .blend files directly in the browser when you paste a URL to a CORS-accessible file or a GitHub repository link. The viewer renders mesh geometry with materials and lighting, plus optional saved camera positions from Blender 5.x files, and offers interactive orbit controls, a wireframe mode, and model fitting. He used ChatGPT Images 2.5 to generate an image of a Fabergé-style egg themed after the TV show Pluribus, then pasted it into Codex running GPT-6 Astra (high) and prompted it to use a local Blender skill to build a model of it; the run took 17 minutes 51 seconds and produced several .blend files. He added the viewer to his tools collection so the resulting Pluribus model can be viewed in a browser.

Research

⭐⭐ [Research] Quoting Calif Research

Calif Research released a demo of WeWorm, which it describes as the first zero-click worm to spread through WeChat calls on iOS and Android. According to the post, the victim does not need to answer the call or interact with the phone at all, and even if they do answer they hear nothing while the exploit still succeeds. The team says that working with AI it found the bug and wrote the first remote code execution (RCE) exploit in about two days, with the worm taking one more week to build. It notes that a worm at this scale used to take a larger team months, that AI can already do most of the work, and that the team supplied the judgment about what to target and how to test safely. This is a quotation collected by Simon Willison.

Other

⭐⭐ [Other] d-Matrix Adopts NVIDIA NVLink Fusion for Rack-Scale XPU Deployment

AI inference chipmaker d-Matrix announced it will adopt NVIDIA NVLink Fusion to connect its next-generation Raptor XPUs to NVIDIA's AI infrastructure platform, joining a growing roster of ecosystem partners. Through the integration, Raptor connects to NVIDIA NVLink scale-up and Spectrum-X scale-out networking, the NVIDIA MGX rack architecture, and the broader NVIDIA AI platform. The source text is truncated at this point and provides no further technical details or timeline.

⭐⭐ [Other] NVIDIA Expands AI Infrastructure Capacity in Partnership With Australia’s Data Center Ecosystem

NVIDIA announced it is collaborating with a growing ecosystem of Australian NVIDIA Cloud Partners (NCPs) and AI infrastructure partners to expand land, power and shell capacity intended to host multiple generations of NVIDIA DSX AI factories. The company says the buildout is aimed at supporting Australia's growing demand for AI compute, with up to 2 gigawatts planned by 2027.