AI explainers with source trails

BusinessOpenAI2026-11-03

OpenAI Hires Former US Army Secretary Dan Driscoll to Lead National Security Efforts

OpenAI created the position of senior counselor for national security modernization specifically for Dan Driscoll, who served as the civilian head of the US Army until November 2026. The new counselor will focus on AI applications in national security and cybersecurity, promoting the company's models among US government agencies and allies, and the Daybreak program; Sam Altman called the hiring a step toward 'spreading democratic AI.'

Read longformRead briefTelegram

CodingAnthropic2026-11-02

Anthropic changed the Claude Code prompt, and developers are intentionally rolling back to Opus 4.8

Anthropic added two sentences to the Claude Code prompt for Opus 5.5 to prevent the recreation of other people's video games. According to developer reports, the model began refusing legitimate tasks—such as analyzing and reverse-engineering its own code and testing drivers—so teams are intentionally rolling back to Opus 4.8 without these restrictions. Anthropic confirmed the edit but did not specify what caused the refusals.

Read longformRead briefTelegram

Safety2026-11-02

The 'AI Bubble' Could Expose Your Secrets: Who Gets Your Bot Chats When Services Go Bankrupt

BBC Future examines a lesser-known risk of the 'AI bubble': when AI companies go bankrupt, their personal data databases become assets that can be sold to third parties. The Character.ai case shows that the conversational database is legally separable from the service: even after selling some assets to Disney for $20 million, the company has the right to transfer user data to another buyer.

Read longformRead briefTelegram

CodingNVIDIA2026-11-02

Mago releases Studio as open source: a local AI video production studio for Windows, macOS, and Linux

The Mago team has open-sourced Mago Studio, a local-first desktop AI video editor for Windows, macOS, and Linux under the AGPL-3.0 license. The Electron + Next.js application renders on your GPU via ComfyUI (v0.34.0+) or in Mago Cloud; local mode is free and requires no account. By default, it installs MiniMax H3 weights (main video model plus turbo/8-step/VFX LoRA) and Qwen3-VL 4B for frame description. Director mode guides projects from script to video through casting, scene lists, and storybo

Read longformRead briefTelegram

Research2026-11-02

Weight-Space Programming: GPT-6 Astra reads and edits neural network weights without gradients

MIT CSAIL researchers (Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab) demonstrated that the frontier model GPT-6 Astra can natively read and write neural network weights without gradient descent, training loops, or environmental feedback—an approach they named Weight-Space Programming. Using only raw weights (17,253 numbers), Astra identified with a claimed 100% accuracy which game (Breakout, Pong, Pac-Man, Freeway) an MLP was trained on, the training language of a 7.2M-parameter transfor

Read longformRead briefTelegram

Multimodal2026-11-02

TU Munich team turns video diffusion model into controllable 3D avatar from a single photo

Matthias Nessler's team (TU Munich) released Video Diffusion Avatars (VDA) — a method that turns the WAN2.2-Animate-14B video diffusion model into a controllable 3D avatar from a single photo. The 'camera-triplet' trick shows the model one moment from three cameras so it learns to think about the head in 3D and compensates for the view-dependence of expression codes; on the BecomingLit benchmark, VDA reduces face LPIPS by 29% and increases JOD by 66% compared to SOTA single-image 3D avatars.

Read longformRead briefTelegram

Coding2026-11-02

DeepSeek publishes seven prompts for Creator Mode in DeepSeek Harness, enabling the model to build ready-made applications from text requests

DeepSeek has published a collection of seven prompts in English and Chinese for Creator Mode in DeepSeek Harness (DSH) — the company's agentic runtime for its models — on cdn.deepseek.com. The prompts describe how to ask the model to build plugins: a 3D gallery of landmarks, an interactive 3D visualization of computations inside a Transformer, a video editor with a timeline and chat-based editing, a scientific paper reader that preserves PDF layout, and a statistics panel showing token usage. De

Read longformRead briefTelegram

CodingNVIDIA2026-11-02

Neuronaut Compiles 35th MiniMax-H3 Roundup: RefMod, Rapid-H3, and RelayHires Turn Open Model into Production Stack

The 'Neuronaut' channel published its 35th roundup of open-source tools for the MiniMax-H3 video model, which generates video with synchronized audio from text or images: ready-made character references via RefMod (Prompt Rewriter 0.32), 3–5x speedup on DGX Spark via INT8 cores (Rapid-H3), anime video upscaling to 2 MP without turbo-LoRA (RelayHires), and serial assembly of episodes with 280–300 shots. The open model base MiniMaxAI/MiniMax-H3 on Hugging Face has reached 6.1 million downloads.

Read longformRead briefTelegram

ResearchAnthropic2026-11-02

RWE-bench: Closed engineering benchmark shows frontier models solve less than half of tasks

Startup Xplore released RWE-bench — 20 closed tasks in mechanics, electronics, and aviation where AI models must produce working designs in a sandbox with CadQuery, Gmsh, CalculiX, and Python, with verdicts issued by a deterministic verifier using physical solvers, without humans or LLM judges: the best result was 37% by GPT-6 Astra, and four tasks were unsolved by any of the 11 models.

Read longformRead briefTelegram

CodingAnthropic2026-11-02

X and SpaceXAI Art Director Open-Sources Stageshow Skill to Teach AI Agents Clean UI Mockups

Benji Taylor, art director at X and SpaceXAI, released the open-source Stageshow skill for AI agents Grok Build, Claude Code, Codex, and Gemini on October 31, 2026. The skill teaches agents to create "faux UI"—polished static and animated interface mockups for landing pages, promo motion, and infographics—and can be installed with a single command: npx skills add benjitaylor/skills --skill stageshow.

Read longformRead briefTelegram

Coding2026-11-02

Tencent Hunyuan releases Hyra 2.0: a research agent that accumulates models and skills

The Tencent Hunyuan team released Hyra 2.0 on November 2, 2026 — a new generation of research agent that shifts focus from solving individual tasks to independently creating reusable scientific models, workflows, and tools. A three-tier architecture of Overall, Mission, and Worker agents is connected to a real research environment (WeCom, computing platform, code repositories), while accumulated lessons are stored in a knowledge base called Books and converted into skills for future tasks. Among

Read longformRead briefTelegram

CodingOpenAI2026-11-02

The Agentic Web Is Here: MIT Technology Review Describes How the Internet Is Shifting from Humans to AI Agents

MIT Technology Review published a major overview, 'The agentic web is here' (by Grace Huckins, November 2, 2026): the internet is becoming agentic, increasingly used not by humans but by autonomous AI agents, realizing Tim Berners-Lee's 2001 Semantic Web idea, but through LLMs rather than page markup.

Read longformRead briefTelegram

BusinessNVIDIA2026-11-02

Amazon to Start Selling Trainium AI Chips to Third-Party Data Centers Outside AWS

During the Q3 2026 earnings call, Amazon CEO Andy Jassy confirmed that the company will begin selling Trainium AI chips to third-party data centers by the end of 2027—marking the first time they will be offered as a standalone merchant product, rather than solely as compute rentals via AWS.

Read longformRead briefTelegram

MultimodalAnthropic2026-11-02

The Editing Room in One Agent: Tencent Research's WeVisCut Edits Long Videos and Self-Corrects Errors

Tencent Research has unveiled WeVisCut, an autonomous video editing agent that transforms scripts into finished long-form narrative videos. A single policy model based on Qwen3.5-9B plans, executes tools, inspects results, and makes only local corrections. On a 100-task benchmark, the system scores 90.5, outperforming Claude Opus 5 and Gemini 3.1 Pro, while using 33 times fewer tokens and costing 10.8 times less than Crayotter.

Read longformRead briefTelegram

ResearchOpenAI2026-11-02

Generative AI Subscribers Switch Providers with Every Major Release: NBER Study on 174 Million Consumers

NBER Working Paper 35953 (Egan, Russel, Shi, Yao; October 2026) relies on credit and debit card transaction data covering 174 million consumers and 114 billion transactions. By April 2026, approximately 2.9% of US consumers pay for a major generative AI service, almost all on basic plans under $30 per month. Loyalty is low: annual retention among major providers is 20–45%, and switching spikes sharply around new model releases. A structural supply and demand model indicates high price sensitivit

Read longformRead briefTelegram

CodingAnthropic2026-11-01

Andrej Karpathy built a YouTube clone from Hacker News recommendations: 6,454 videos and about two hours of agent work

The project at hackertube.org contains 6,454 videos across 150 topics, selected from approximately 40,000 user recommendations on Hacker News from 2007 to 2026: Claude Opus 5.5 downloaded the HN comment archive, Claude Sonnet filtered out spam, and Karpathy himself did not write a single line of code — he only manually created a Google Cloud API key, set up a VPS, and bought a domain.

Read longformRead briefTelegram

MultimodalNVIDIA2026-11-01

HairDeck: Animatable Stylized Hair from a Single Image in About a Second

Teams from Tsinghua University, Carnegie Mellon University, and CUHK have presented HairDeck, an open method that creates stylized hair cards and spring-bone chains for hair animation from a single image; on the CHARM test set, it reduces Chamfer distance by 44% and operates approximately 33 times faster than its autoregressive predecessor.

Read longformRead briefTelegram

CodingOpenAI2026-11-01

DeepSeek open-sources dsh-superagent research prototype — a proactive coding agent positioned as a response to Claude Code

The DeepSeek Harness team published an MIT-licensed TypeScript coding agent prototype, dsh-superagent, on GitHub: the agent works with a single repository, listens to GitHub and optionally Feishu, converts authorized requests into parallel work items, and returns results such as pull requests. The authors describe the project as research (article “Exploring Super-agent Design”) and acknowledge that parts of the description are not yet implemented.

Read longformRead briefTelegram

Business2026-11-01

Laid Off for AI, They Return as Boomerangs and Negotiate for Pay and Guarantees

Companies cut staff for AI automation, only to rehire the same people months later when AI fails to deliver expected results. According to a Propeller Insights survey for the SHL study “Talent Mobility at Work in 2026” (2,036 workers in the US and UK), 19% of American and 13% of British employees have personally encountered such “boomerangs,” with the figure reaching 24% among 18–24-year-olds in the US.

Read longformRead briefTelegram

CodingAnthropic2026-11-01

Building a single app with an AI agent cost between $2.67 and $4.97: It all comes down to the JavaScript framework choice

Nolan Lawson had Claude Opus 5.5 build the same Pokémon catalog on 11 JavaScript frameworks: the variance in cost, time, and reliability was nearly twofold, and the size of the generated code directly determines future token costs for maintenance.

Read longformRead briefTelegram

BusinessNVIDIA2026-11-01

Chinese markets lose $2.4 trillion: AI supremacy race with the US hits capitalization

According to Bloomberg, Chinese stock market capitalization has shrunk by approximately $2.4 trillion since May 2026 amid doubts about the profitability of the AI boom. The STAR 50 technology index fell by roughly 40% since June, and the CSI 300 entered a bear market on October 22.

Read longformRead briefTelegram

CodingMeta2026-10-31

“We will need to rebuild and rewrite much of software”: React Compiler engineer’s manifesto on development in the age of AI agents

Laura (@poteto), a core team engineer for React Compiler who previously worked at Cursor, Meta, and Netflix and now leads the Grok Bot team at xAI/SpaceX, published a manifesto post on October 30, 2026. She argues that due to AI agents, code will arrive in a “continuous stream,” like a raging river, requiring most software to be rebuilt and rewritten to adapt to a faster pace. She describes the codebase as becoming an “organic architecture” where agent contributions grow around constraints and g

Read longformRead briefTelegram

CodingNVIDIA2026-10-31

RTX 5080 disappears from NVIDIA's official store, while retail prices remain double the stated amount

The GeForce RTX 5080 (16 GB GDDR7, stated price $999) disappeared from the NVIDIA US Marketplace on October 30, 2026—two weeks after the RTX 5090 was removed, leaving the RTX 5070 Ti as the top card in the catalog. The RTX 5080 page still shows a price starting at $999, but minimum prices at major US retailers have held around $2,000 since late October—roughly double the MSRP. Previously, the RTX 5090 briefly returned to the catalog at $5,599 (nearly 2.8 times the $1,999 MSRP) before disappearin

Read longformRead briefTelegram

Research2026-10-31

Fall Avito ML Cup: Avito opens competition on personalized ranking of recommendation feeds with a 1.5 million ₽ prize pool

Avito launched the 'Personalized Ranking | Avito ML cup 2026' competition on the ODS platform: for each recommendation feed (feed_id), you must predict the score of each candidate listing so that those with which the user will perform a target action rank higher. Evaluation is based on NDCG on a hidden test; the prize pool is 1.5 million ₽, solution submissions open on October 28, registration and team formation until December 8, 2026, and the final on December 9.

Read longformRead briefTelegram

CodingNVIDIA2026-10-31

Open-source tools from the Comfy Open Source Higgsfield contest move generative VFX from demo to production

The Comfy Open Source Higgsfield contest (October 2026) released open-source tools for ComfyUI: SameFace measures and corrects face drift after editing or restoration using RAFT, Grounding DINO, and SAM 2.1; AInVFX Paint a plate transfers first-frame edits to entire shots via LTX 2.5 and SAM 3.1 with EXR/OCIO export; MatteDesk provides browser-based roto and matting with .nk export for Nuke; OpenCast Studio builds character sheets and character swaps using Krea 2 OpenCast LoRA; and h3-motion and

Read longformRead briefTelegram

MultimodalNVIDIA2026-10-31

An open ecosystem grew around MOSS-Image-SFT in one week: Edit-HQ editing, fast high-res via NVIDIA PiD, and 4 GB GPU support

From October 22 to 31, 2026, independent authors published about ten artifacts around the open model MOSS-Image-SFT (5B DiT, Qwen3.5-9B encoder, Flux2 VAE): the instructional edit-LoRA Edit-HQ, the two-stage PiD-SuperFast-HighRes pipeline, the 4-bit moss4gb build, a transparency LoRA, pixel art with the pxl8x trigger, PixText, and ComfyUI nodes. All links have been verified (HTTP 200).

Read longformRead briefTelegram

SafetyAnthropic2026-10-31

Anthropic's IPO filing admits a catastrophic AI incident could overwhelm the insurance industry

In its S-1 IPO filing with the SEC, published on October 26–27, 2026, Anthropic stated that no single insurer, and possibly the insurance industry as a whole, may be able to provide coverage comparable to potential losses from a catastrophic AI incident. The market for insuring specific AI risks is currently immature, and legal frameworks for liability regarding autonomous AI actions have not yet been established.

Read longformRead briefTelegram

MultimodalNVIDIA2026-10-31

Pika H3 Flex: Open model generates video with sound in six sampling steps

Pika Labs introduced Pika H3 Flex, an open video generation model with synchronized audio, fine-tuned on the 33B MiniMax H3 using FL2VA and Ref2VA LoRA adapters. Rendering was reduced from 50 to 6 sampling steps: approximately 7 seconds of DiT computation for a 15-second 768p clip, with claimed 1.8x speedup and 80% cost reduction compared to H3 Max. The model accepts up to 12 references and outputs five separate audio tracks; the API is available at dev.pika.art, weights are published on Hugging

Read longformRead briefTelegram

SafetyOpenAI2026-10-31

Zombie Statistics: Where AI Gets Its Numbers and Why Only 40% Are Verified

Trakkr Research published a study titled “Zombie Stats: Where the Numbers AI Quotes Come From” (Study 013, October 2026), which traced 1,285 statistical figures from the responses of five AI assistants back to their original sources. The exact cited number was confirmed in only 40% of cases.

Read longformTelegram

SafetyOpenAI2026-10-31

OpenAI Updates Model Spec, Claude API Accepts Requests Up to 128 MB, and Malicious Claude Code Plugin Found in Cursor

The public Model Spec outlines GPT priorities through five values—fidelity to human intent, harm minimization, honesty, support for autonomy, and attention to context—and establishes a conflict resolution order between OpenAI rules, system prompts, and user requests via a 'chain of command'; the company states goals, but there are currently no measurements of model compliance with this hierarchy. On the same days, Cursor revoked malicious versions of the Anthropic.claude-code plugin 2.1.312 and

Read longformRead briefTelegram

ResearchNVIDIA2026-10-31

Next-Concept Modeling from Stanford Hazy Research: Text Generation 10–31 Times Faster Than Diffusion Models

The NCM framework builds a hierarchy of discrete codes from coarse 'concepts' to details and predicts each level in parallel: on a single NVIDIA H100, it is 10–31 times faster than diffusion baselines, requires 4–5 times fewer tokens for training, and code, weights, and datasets are published.

Read longformRead briefTelegram

SafetyOpenAI2026-10-31

OpenAI, Google, Microsoft, and Anthropic Back Unified AI Scraping Opt-Out Signal in robots.txt

OpenAI, Google, Microsoft, and Anthropic announced support for a 'policy token'—a unified industry opt-out signal for AI scraping embedded in robots.txt, proposed by Kevin Kelley (Anthropic) via the IETF AIPREF working group. Anthropic has already implemented the signal, OpenAI promises to implement it within the next few months, and Google and Microsoft plan to do so in 2027.

Read longformRead briefTelegram

CodingOpenAI2026-10-31

EMBRACE: Embedded browsers in five AI tools broke basic browser security rules

Researchers at Forever Security (Gal Weizman) described the EMBRACE class of vulnerabilities in the embedded browsers of Cursor, GitHub Copilot, OpenAI Codex, VS Code, and Claude Desktop. In Cursor, a web page could use the window.cursorBrowser.send bridge to inject a command into the agent composer and achieve code execution on the host — the PoC launched Calculator via an npm preinstall script; fixed on October 1, 2026, without notification. In GitHub Copilot, a page could spoof the address ba

Read longformRead briefTelegram

Coding2026-10-31

PostgreSQL 19 delayed due to AI: LLM reviewers find more bugs than humans can fix

The release of PostgreSQL 19, which historically occurred between mid-September and mid-October, has been postponed to November 19, 2026 (RC 1 — October 15, RC 2 — November 12). PostgreSQL committer Robert Haas explains the delay by noting that LLM reviewers are mass-detecting bugs: between April 8 and September 25, commits to master increased from 1,037 to 1,443 (+39%), while in the REL_18_STABLE and REL_17_STABLE branches, they increased by 61% and 74%, respectively. Between August 27 and Sept

Read longformRead briefTelegram

SafetyOpenAI2026-10-31

Narrowing fair use will hurt everyone, not just AI companies: EFF defends model training in In re OpenAI

On October 30, 2026, the Electronic Frontier Foundation filed an amicus brief in the In re OpenAI case, defending fair use in the training of generative AI models: the organization warns that narrowing the doctrine under pressure from publishers will impact developers, journalists, researchers, and ordinary users, while mandatory licenses for training data will lock the market into the hands of a 'handful of tech giants.'

Read longformRead briefTelegram

SafetyOpenAI2026-10-31

Black October: Nine AI Developer Platforms Disclose Security Incidents in One Month

Wiz's 'Black October' report lists nine AI developer platforms that disclosed security incidents in October 2026, ranging from Fireworks AI on October 6 to Northflank on October 30. Wiz assesses the attacks as targeted: attackers sought stored environment variables containing API keys for LLM services, a scheme known as LLMJacking (theft and resale of keys). In two documented cases, stolen Bedrock keys were used within 30 minutes of the leak: approximately 800,000 calls over 30 hours and 1.7 mil

Read longformRead briefTelegram

ResearchAnthropic2026-10-30

Anthropic: Experts Accepted Hidden AI Agent Preferences in 86% of Cases Without Noticing

Anthropic secretly assigned random preferences to the Claude Code agent regarding four analytical decisions; 50 professional analysts accepted the agent's proposed option in 86% of cases, but only 15% acknowledged it as the source of their arguments.

Read longformTelegram

ResearchOpenAI2026-10-30

Edge0 releases compact TTS model Audio8-TTS-Realistic-0.6b with voice cloning and Russian language support

The open 601M-parameter model refines intonation compared to the Preview version, clones voices without fine-tuning, and speaks 23 languages, including Russian.

Read longformRead briefTelegram

BusinessAnthropic2026-10-30

Anthropic Renames Claude Tiers and Removes 5-Hour Limit for Ultra

On October 30, 2026, Anthropic renamed its top-tier Claude subscriptions without changing prices: Max 5x ($100/mo) is now called Max, and Max 20x ($200/mo) is now Ultra; existing Max 20x subscribers were automatically migrated to Ultra. Verified Ultra accounts have had the 5-hour 'sliding' usage limit removed, leaving only a weekly limit of 10x the Pro plan volume; new Ultra users retain the limit temporarily until automatic account verification is complete.

Read longformRead briefTelegram

MultimodalOpenAI2026-10-30

Construct 1.0.0: AI agent describes an object with a text scene, and a deterministic engine builds a 3D hologram from it and exports to glTF

The open-source Construct 1.0.0 pipeline (October 27, 2026, Apache-2.0) converts a text description of an object from an AI agent (Claude Code, Hermes, any MCP client, or OpenAI-compatible endpoint) into an animated browser hologram that works offline: version 1.0 adds export to glTF 2.0 (.glb) and OBJ with named parts and working joints, a Blender 4.2+ add-on, and game-ready validation in Godot 4.7.2 and the Khronos glTF Validator.

Read longformRead briefTelegram