News

Apparently, OpenAI isn't trying to build "magic intelligence in the sky" anymore

OpenAI CEO Sam Altman called attributing religious power to AI models a "real safety issue." He also warned against surrendering human judgment to them. In 2023 and 2024, Altman described OpenAI's goal as building "magic intelligence in the sky."

Humanized models sit beside warnings about belief

Altman's comments follow reports about Anthropic's meetings with religious thinkers. Pope Leo XIV also issued a statement on AI. OpenAI has held discussions with religious leaders itself. Altman has repeatedly said he wants AI on the level depicted in the movie "Her," meaning deeply humanized systems. As recently as 2024, he spoke of feeling "on the side of the angels" while working on AI. (The Decoder)

thinkidiot take: Altman now calls surrendering human judgment to AI a safety issue after describing his own ambition in terms of magic. I would take that warning as a reason to keep a model's human qualities separate from the authority I give its answers. His continued interest in AI like "Her" makes that distinction central to his own ambitions. Dropping the religious language is overdue, and the earlier rhetoric deserves scrutiny.

Amazon responds to data center backlash, says it no longer uses NDAs

AWS CEO Matt Garman said Amazon has stopped using nondisclosure agreements with government agencies on its data center projects. His response to data center backlash comes as New York announced a one-year moratorium on permits for large data centers.

Water figures leave power and emissions in the argument

Garman said more than 100 data center moratoriums are being considered across the United States. Citing an Amazon report, he said direct data center water consumption accounts for 0.5% of US industrial water usage. That figure concerns direct water consumption. An independent watchdog attributed a 76% year-over-year price increase on America's largest electrical grid primarily to data centers. A planned Amazon data center in Texas is permitted to release 33 million tons of carbon dioxide per year. (TechCrunch)

thinkidiot take: A watchdog attributed a 76% annual price increase on America's largest electrical grid primarily to data centers. I would put that finding alongside Amazon's water figure when assessing its response to local opposition. A share of industrial water usage does not answer a finding about electricity prices or a permit for carbon emissions. Ending NDAs is a useful change, but it is an inadequate answer to those concerns.

Capcom is preparing for a ‘future where we create games together with AI’

Capcom programmer Satoshi Ishida presented plans for the REX project and RE Engine at Capcom Open Conference RE: 2026. According to IGN's translation, he outlined gradually turning RE Engine into an "AI-generation game engine." The presentation described a planned direction for the engine, rather than announcing a completed transformation.

The asset policy sets a boundary for engine ambitions

Ishida advocated integrating AI into development workflows. He linked that approach to time-consuming tasks in games at Resident Evil's scale. Capcom previously said it would not use AI-generated assets in its games. Its stated use for AI was to improve development efficiency. The studio's Pragmata also explores the horrors of AI. (The Verge)

thinkidiot take: Capcom's proposed engine direction sits alongside its stated refusal to use AI-generated assets in games. I would judge the plan by whether it helps with the time-consuming development tasks Ishida identified while respecting that boundary. The phrase "AI-generation game engine" says less about that distinction than the existing asset policy does. Capcom should make the workflow changes the substance of its pitch.

Splice CEO Kakul Srivastava thinks AI emails are killing conversations

Splice CEO Kakul Srivastava said AI-written documents worsen discussion when colleagues cannot tell whether a person is behind the words. She described remote-first Splice's essential tools as a well-written document and a real conversation about it. Her criticism puts the connection between writing and discussion at the center of the company's working practices.

A music AI strategy built around human participation

Srivastava previously held executive roles at Flickr, Yahoo, GitHub, and Adobe. Splice acquired Spitfire Audio under her leadership. The sample platform supplies producers with one-shots and melodic loops. Samples from the service appear in Lisa's "Money" and Sabrina Carpenter's "Espresso." Splice has sought to offer AI products that keep humans central to the creative process. On Decoder last year, Srivastava criticized "push-button, get-song" AI. (The Verge)

thinkidiot take: Srivastava says discussion gets worse when colleagues cannot tell whether a person is behind a document. I would make the ability to explain and defend my own text a condition of using AI to help write it. That fits her emphasis on a real conversation about the document. A polished page is a poor trade if its author cannot carry the discussion it starts.

"Muse Gadgets" turns AI hardware into an open-source DIY project

Meta announced Muse Gadgets, an open-source project for building AI hardware connected to its Muse agent. It gives hobbyists firmware for ESP32 boards and a Linux SDK. The project's code uses the Apache 2.0 license.

A limited device run brings Muse onto the home network

Meta also produced a USB-C device called Muse Home Link. It connects Muse to a home network to control TVs, speakers, and other devices with an HTTPS interface. Meta said it made 5,000 units. Those units are expected to ship within weeks. They are offered free to Muse subscribers while supplies last. The open-source approach also helps Meta learn which AI hardware form factors people want. (The Decoder)

thinkidiot take: Meta is offering 5,000 Muse Home Link units while also releasing code for people to build their own hardware. I would start with the ESP32 firmware and Linux SDK to explore a device of my own. That route also serves Meta's stated interest in learning which hardware forms people want. Giving hobbyists something they can build with is the more compelling part of this release.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-10-03.

OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction

An AI watching live video cannot know which details someone will ask about later. OneStreamer teaches it to keep written notes as events unfold and answer when it has enough information. The aim is to remember earlier events while still following what is happening now.

• Problem: A live video system must save useful details before it knows how they will matter. Keeping track of the past must not weaken its ability to understand the present. It also needs to learn when to answer, rather than letting repeated examples of waiting dominate its training.
• New idea: OneStreamer trains a model to write notes without needing a question first, alongside learning to answer tasks. Its caption memory combines descriptions of details tied to particular moments with summaries of finished events. When answering, it uses those notes together with recent video, without reopening older visual material. Its timing training keeps every output checkpoint but selects representative signals of whether to keep waiting or change state.
• Simple example: Think of someone taking notes during a live demonstration. Short notes preserve earlier steps for later questions, while the person keeps watching the current step. OneStreamer uses generated captions in a similar way.
• Evidence: The 4B model ranked best among the compared methods on all eight evaluated streaming video understanding benchmarks. Keeping generated captions improved answers about earlier events without reducing real-time perception performance. The timing method beat supervision using all annotated state tokens while using only 27.5% of them. The training dataset contains over one million records.
• Limitation: The abstract gives no measured response delays or memory costs as a video stream grows longer. Its benchmark results therefore leave the practical cost of sustained live use unresolved.
• Why it matters: A video assistant needs to remember earlier evidence while staying ready to respond to what happens next.
• Paper: OneStreamer: Unifying Perception, Memory

On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics

When a weaker AI model learns from a stronger one, whose answers should it practise on? This study separates that choice from other training settings that often change alongside it. The results suggest that how learning is scored and how large the updates are can matter more than who generated the practice answers.

• Problem: Earlier comparisons changed several training ingredients at once, making it hard to tell why one approach worked better. That weakens claims that practising on a model's own answers inherently preserves old skills or improves performance on unfamiliar problems.
• New idea: The study uses distillation, where a weaker student model learns from a stronger teacher model. It separately changes the rollout policy, meaning which model generates the practice answers; the KL direction, meaning which way the training measure compares student and teacher predictions; and the learning rate, which controls update size. It also tests mixtures of student and teacher answers. This setup helps distinguish the effects of those choices.
• Simple example: Imagine comparing a pupil who works through their own attempts with one who studies a tutor's worked answers. If the marking method and size of each correction also change, the comparison cannot isolate the value of either practice style. This study separates those ingredients.
• Evidence: Across the tested Llama3 and Qwen2.5 models and reasoning tasks, the direction of the prediction comparison shaped task results and the range of outputs. Learning rate governed forgetting and how concentrated parameter changes were. Forward KL stayed strong as the source of practice answers changed, while reverse KL was more sensitive and favoured student-generated answers. Student-generated practice also helped on harder Countdown problems under both directions.
• Limitation: The advantage on harder Countdown problems did not reliably survive a later stage of reinforcement learning. The abstract reports no numerical effect sizes, so it does not show how large the observed differences were.
• Why it matters: Training choices are easier to judge when their effects are tested separately.
• Paper: On-Policy or Off-Policy Learning? A Systematic Study of

GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis

GraphForge builds training exercises for AI assistants that work with files and tools. It starts with real documents and uses their contents to shape both the assignment and the checks on the finished work. The aim is to teach assistants through tasks whose answers can be checked against the material they were given.

• Problem: Training an assistant for file-based work needs realistic materials and a reliable way to judge the result. Existing methods either invent files that lack realism and variety or use real files without checks tailored to the task. That leaves a gap between having a plausible assignment and knowing whether it was completed correctly.
• New idea: GraphForge starts with prompts based on occupations and gathers real files into a workspace, the collection available for a task. It builds an evidence graph, a map of relationships among those files. That map supplies both the assignment and its rubric, the criteria used to judge the result, with each criterion linked to supporting files. A trial attempt checks whether the task can be carried out, then another AI revises the assignment and criteria against the original files before training examples are collected.
• Simple example: Think of a workplace exercise with a folder of source documents and an answer checklist. Each checklist item points to the document needed to check it. GraphForge builds that connection into the exercise itself.
• Evidence: Training Qwen3.6-27B on 2,169 GraphForge examples brought GDPVal to 1445.7, a gain of 65.7, under OpenHands. Under Claude Code, Workspace-Bench-Lite reached 63.7, a gain of 7.7, and SpreadsheetBench II reached 24.0, a gain of 13.7. Further training on the model's own attempts, selected using the file-backed criteria, improved all three benchmarks again. The data and models are available.
• Limitation: The abstract reports training results for Qwen3.6-27B but does not establish whether other models would benefit similarly. It also gives no numerical gains for the additional training on selected attempts.
• Why it matters: File-backed assignments and checks give working assistants a clearer basis for learning what counts as a correct result.
• Paper: GraphForge: Training Working Agents with Graph-Anchored

Trending AI Repositories

Ranking source: GitHub Trending.

pingdotgg/t3code

T3 Code is a control surface for agents running on your machine. It brings access to those agents to mobile, web and desktop apps.

• What it is: This TypeScript project sits alongside agent tools such as Claude Code, Codex and Cursor. Its desktop app is built with Electron.
• What it does:
• Who it helps: It helps people who want to control their local agents from different devices. They can use an iOS or Android app, a browser or a desktop app.
• Limitation: The agents it controls run on your machine.
• Repository: pingdotgg/t3code

thedotmack/claude-mem

Claude-Mem is a TypeScript project focused on memory for agents. Its support spans several agent tools, making it relevant beyond Claude Code.

• What it is: It captures session activity, compresses it with AI, and supplies relevant context in later sessions. It supports Claude Code and several other agent tools.
• What it does: Persistent Context Across Sessions for Every Agent , Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More
• Who it helps: It helps people whose agent work continues across sessions. They can draw on earlier session context when starting later work.
• Limitation: The supplied excerpt does not explain installation or configuration requirements.
• Repository: thedotmack/claude-mem

cloudflare/cloudflare-os

Cloudflare OS is an AI productivity environment originally developed inside Cloudflare. A large portion of the company's workforce uses it daily, giving readers an example of AI tooling used across job roles.

• What it is: This TypeScript project is a shared environment for AI-assisted work. Its README shows a planning workspace with an AI-generated slide deck.
• What it does: Agent workspace built on Cloudflare Workers for creating documents, building apps, and running agents with your company’s context and systems.
• Who it helps: It serves people across engineering, sales and other company roles. They can use agents to help create documents and build apps with company context.
• Limitation: Despite its name, it is not a traditional computer operating system.
• Repository: cloudflare/cloudflare-os