News

Wire It, Run It, Deploy It: AI Workflows in Gradio

Hugging Face introduced gr.Workflow, which represents AI pipelines as graphs of typed nodes on a drag-and-drop canvas. Each graph becomes a REST API and can be deployed to Hugging Face Spaces with one command. The examples range from a one-node image editor to a media workflow with four model and Space calls.

The canvas is also an API surface

The image editor calls Qwen-Image-Edit through Hugging Face Inference Providers from a single node. The AI Media Studio makes two model calls through Inference Providers and two calls to Gradio Spaces. Its workflow exposes three endpoints: /sticker, /voiceover, and /episode_title. An fn node can also carry the @spaces.GPU decorator. That setup runs Lightricks/LTX-Video through Diffusers on ZeroGPU. The same graph therefore connects interface design, hosted inference, GPU work and deployment. (Hugging Face)

thinkidiot take: Three endpoints emerge directly from one media workflow, which makes the graph more than a visual aid. I would use the canvas to inspect the pipeline, then call /sticker, /voiceover, and /episode_title from separate clients. That joins prototyping and API delivery in the same artifact instead of rebuilding the workflow after the demo works. Turning a graph into both an interface and a deployable REST API is the right abstraction here.

Introducing new Ray capabilities on SageMaker HyperPod

Amazon added managed Ray support on Amazon EKS to SageMaker HyperPod for foundation-model training and serving. SageMaker Studio can now create and monitor Ray clusters, connect notebooks to live clusters, and submit distributed jobs. The integration uses open-source KubeRay and standard Ray APIs, so existing scripts and workflows run without modification.

Recovery and observability move into the managed path

Studio can open both the Ray Dashboard and Amazon Managed Grafana dashboards. JupyterLab and Code Editor workspaces can attach to running clusters. Operators can also configure hung-job detection. HyperPod node-health monitoring and automatic recovery give Ray training jobs fault tolerance. Tiered checkpointing uses HyperPod distributed tiered storage to speed job resumption. SageMaker JumpStart loads model weights directly into Ray Serve endpoints, while KV-cache offloading supports long-context requests. (AWS Machine Learning)

thinkidiot take: Existing Ray scripts running without modification removes the most concrete migration burden in this integration. I would attach a JupyterLab workspace to a live cluster, watch Ray and Grafana dashboards, and enable hung-job detection before submitting training. Automatic recovery and tiered checkpointing then address the failure path instead of treating cluster loss as a manual cleanup exercise. Managed Ray is useful when it preserves standard Ray APIs, and Amazon made the correct choice by doing that.

How to Train a Critic Stably and Efficiently

An arXiv paper submitted on August 24, 2026 introduces Best-Practice Critic Optimization, or BPCO, for stable critic-based language-model training. The method estimates token-level advantages from one response per prompt instead of sampling multiple responses as group-based methods such as GRPO do. In reported mathematical-reasoning experiments from 1.5B parameters to 30B-A3B mixtures of experts, BPCO matches or exceeds a group-based baseline.

Hidden grading context gives the critic a separate view

BPCO combines DPPO with reward-range-bounded value predictions. It uses Monte Carlo value targets and unnormalized policy advantages. Length-adaptive generalized advantage estimation is another part of the recipe. The critic can receive reward-defining information that the policy does not see. That information can include a reference answer or a grading rubric. Across the reported experiments, the method consistently improves a strong critic-based baseline. (arXiv)

thinkidiot take: One response per prompt replaces the multiple responses used by group-based training, which directly changes the sampling workload. I would test BPCO where a reference answer or grading rubric is available to the critic but deliberately hidden from the policy. The reported gains across models from 1.5B to 30B-A3B show that the recipe is not tied to one model scale. A critic earns its place when it matches or beats group-based training with less response sampling, and these experiments meet that standard.

ReWorld: An Interactive World Model with Long-Horizon Memory

Researchers introduced ReWorld, an interactive world model that separates short-horizon control from long-horizon memory. It uses a bounded KV cache and a pose-indexed landmark bank to remember earlier places while streaming at 704x1280 resolution. A single backbone supports both high-fidelity multi-step generation and real-time interactive generation.

A fixed cache can still find the way back

Most attention heads use windows restricted to the recent past, while a small set of global heads reaches farther back. ReWorld's data engine aligns eight sources on a common physical action scale. Distribution-matching distillation in a LoRA adapter reduces sampling to four steps. The evaluation compares ReWorld with six recent interactive world models. It reports an 11.95-degree rotation error and the best camera-motion consistency and generation quality under its three-axis protocol. In 64-second, 384-latent out-and-back rollouts, a fixed 12-chunk cache regenerates the starting view. (arXiv)

thinkidiot take: A fixed 12-chunk cache regenerating the starting view after a 64-second out-and-back rollout is the result that matters most to me. I would test the same route repeatedly and inspect whether the pose-indexed landmarks preserve the scene while the recent cache handles control. The design turns long memory into a retrieval problem instead of letting the KV cache grow without bound. That is a more convincing route to interactive world models than treating every past frame as equally active context.

Situational Awareness, star AI hedge fund that nearly imploded, now being probed by the SEC

The SEC is probing Situational Awareness, an AI-focused hedge fund led by former OpenAI employee Leopold Aschenbrenner, according to a TechCrunch report published at 5:23 PM PDT on August 24, 2026. The New York Times reported that regulators subpoenaed banks that supervised the fund's trading and channeled funding to it. The fund has not been accused of wrongdoing and says it will cooperate fully with regulatory requests.

Banks were told to preserve fund information

Situational Awareness had become a prominent AI-focused fund on Wall Street. A downturn in AI stocks at the end of July erased billions of dollars in value at the firm. The reported subpoenas target banks connected to its trading and funding rather than an announced accusation against the fund. The government reportedly instructed those banks to preserve information about the hedge fund. That preservation request places records of the fund's activity within the regulatory inquiry. Situational Awareness says it will cooperate fully. (TechCrunch)

thinkidiot take: Billions of dollars in value disappeared when AI stocks turned down at the end of July, exposing the cost of concentrating a fund around one theme. I would separate that market loss from the SEC inquiry because the fund has not been accused of wrongdoing. The subpoenas and preservation instructions still make its banking and trading records a matter of federal scrutiny. AI expertise does not make concentration risk or regulatory oversight disappear, and investors should judge the fund on that ordinary standard.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-08-25.

Apodex 1.1: Scaling Agentic Intelligence for Complex Work

Apodex 1.1 is an AI system designed to carry complicated jobs through to completion. It works with files, search tools, and code while keeping track of what has happened. The aim is to produce useful results that can be checked, even when a task takes many steps.

• Problem: Language models can reason and combine information, but that is not enough for long, practical assignments. They must also preserve state, recover from errors, use tools, and deliver work that others can verify.
• New idea: Apodex 1.1 improves both its working environments and its coordination between agents. An environment is the set of files, search sources, and code tools available for a task. Agent coordination means splitting a long job into parts, assigning work in parallel, combining later results, and changing the plan when needed. A shared execution system records task state and where each result came from, while training uses past tool use and coordination records to teach more reliable behavior.
• Simple example: It is like a project team that keeps a shared notebook, divides research among several people, checks where every finding came from, and revises the plan when one route fails.
• Evidence: Across professional work, finance, scientific research, mathematics, coding, and search, Apodex 1.1 reached the leading performance band despite being substantially smaller than many frontier systems. The 35B-parameter Apodex 1.1 Mini also retained strong working capability in a form that can run locally.
• Limitation: The abstract gives no benchmark scores, direct comparisons, or tests showing how reliably the system handles failures over very long tasks.
• Why it matters: This matters because useful AI must finish complicated work and leave results that people can verify.
• Paper: Apodex 1.1: Scaling Agentic Intelligence for Complex Work

TLive-Omni: An Omni-Modal Understanding Model for E-Commerce Live Streaming

TLive-Omni is a model built to understand online shopping broadcasts. It combines what is said, shown in the video, printed on screen, pictured in product images, and asked by viewers. The goal is to answer accurately while a long and noisy broadcast is still running.

• Problem: Important product details in a live sales stream can appear in several formats and at different moments. A system that treats speech, images, video, text, and questions separately can miss how they relate over time.
• New idea: TLive-Omni places images, video, audio, and text into one shared representation, meaning a common internal format the model can compare. Per-vGrid is its method for grouping each video section with the audio from the same time and marking where one section ends and the next starts. A three-stage supervised training process teaches perception, and response to instructions in a later stage. Faithful-RFT is an additional training stage that scores answers with feedback that can be checked against the task, with the aim of improving accuracy, wording, and speed.
• Simple example: Imagine watching a seller hold up a shoe while saying its size, as a discount appears in text and a viewer asks about the color. The model tries to connect all four pieces before replying.
• Evidence: Tests on e-commerce live-streaming benchmarks showed strong results across live-commerce tasks and excellent generalization on general benchmarks. The abstract reports no numerical scores.
• Limitation: The abstract does not provide numerical results, comparison details, or evidence from live deployment under real-world timing and noise conditions.
• Why it matters: This matters because reliable live-stream understanding depends on connecting what viewers hear, see, read, and ask at the right moment.
• Paper: TLive-Omni: An Omni-Modal Understanding Model for

RISE: Adaptive Imagination for World Action Models

RISE is a planning system that decides how far an AI should simulate the future before acting. It is aimed at driving systems that use imagined outcomes to choose their next move. The system spends more computation when further simulation may improve safety or planning, and stops when the likely value is low.

• Problem: Existing World Action Models use the same amount of future simulation for every scene. That fixed allocation can waste computation on simple situations and provide too little analysis when risk is developing.
• New idea: RISE stands for Refining Imagination through Selective Rollout. A rollout is a simulated continuation of what may happen next, and the system repeatedly chooses whether to extend it or stop. A Latent Evaluator estimates the risk seen so far and the possible planning gain from continuing, while a Rollout Gate compares that gain with the added computing cost. CounterDrive is a dataset of alternative driving outcomes with different risks, checked and labeled by experts for valid paths, the start of incidents, and their causes.
• Simple example: It is like a driver glancing briefly down an empty road but thinking through several possible outcomes before entering a busy junction.
• Evidence: On NAVSIM and nuScenes, RISE achieved the best overall planning performance while cutting unnecessary rollout. Additional transfer results supported its use across different World Action Model architectures. The abstract gives no numerical scores.
• Limitation: The abstract reports tests only on NAVSIM and nuScenes and does not establish performance in real vehicles or provide numerical safety and computation results.
• Why it matters: This matters because a planner can spend its effort where uncertain or dangerous situations most need deeper thought.
• Paper: RISE: Adaptive Imagination for World Action Models

Trending AI Repositories

Ranking source: GitHub Trending.

MadsLorentzen/ai-job-search

AI Job Search is a local Python framework for managing job applications with AI. It is worth attention because users can fork the project and keep control of their workflow.

• What it is: It is a Python project designed to run on the user's own machine. It sits across job discovery, application writing and interview preparation.
• What it does: The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
• Who it helps: It helps job seekers manage several parts of an application process. They can assess postings, adapt CVs, draft cover letters and prepare for interviews.
• Limitation: It requires Claude Code and a local setup.
• Repository: MadsLorentzen/ai-job-search

multica-ai/andrej-karpathy-skills

Karpathy-Inspired Claude Code Guidelines is a compact set of instructions for coding agents. It is worth attention because it turns observations about common LLM coding failures into reusable guidance.

• What it is: It is an instruction file intended to sit in a Claude Code project. English and Simplified Chinese versions are available.
• What it does: A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
• Who it helps: It helps people who use Claude Code for software work. They can give the agent clearer guidance based on known coding pitfalls.
• Limitation: The guidance is specific to Claude Code and requires adding the CLAUDE.md file to the relevant project.
• Repository: multica-ai/andrej-karpathy-skills

NousResearch/hermes-agent

Hermes Agent is a Python agent project from Nous Research. It includes links to a web presence, a desktop offering and documentation, making it useful to examine as a broader agent system.

• What it is: It is a Python-based agent project from Nous Research with links to Hermes Desktop, a project website, and documentation.
• What it does: The agent that grows with you
• Who it helps: It is relevant to readers interested in examining the Python agent project, Hermes Desktop, or its documentation.
• Limitation: The available evidence does not describe setup requirements or operational limitations.
• Repository: NousResearch/hermes-agent