Latest AI news

ResearchAnthropic2026-11-03

Claude designed proteins that bind scorpion toxins

On October 20, 2026, Anthropic published a report stating that its internal research model, Claude, nearly autonomously conducted a de novo protein design campaign against five scorpion neurotoxins, with independent laboratories Adaptyv Bio and Twist Bioscience confirming binding in 17 of the 211 synthesized proteins. For biotech, this is one of the first documented cases of an LLM agent leading a multi-day (64-hour) autonomous scientific campaign: Claude itself found toxin structures in open PD

Read briefRead longformTelegram

Business2026-11-03

Super Intelligence becomes a loyalty test in the US

Executive Order 14434, dated September 29, 2026, mandated US federal agencies to use Super Intelligence instead of Artificial Intelligence, but a month later the transition remains uneven and has turned into a loyalty test. The renaming is becoming a political loyalty test and is polarizing the national debate on AI.

Read briefRead longformTelegram

BusinessOpenAI2026-11-03

OpenAI Hires Former US Army Secretary to Oversee AI in National Security

On November 2, 2026, OpenAI announced the hiring of Dan Driscoll, former US Secretary of the Army, for a newly created role as senior counselor for national security modernization. He will advise the company and its partners on the application of AI in national security and cybersecurity. OpenAI is transforming national security from a research initiative into a permanent staffed division with a government figure inside a frontier-level laboratory.

Read briefRead longformTelegram

CodingAnthropic2026-11-02

Developers Roll Back from Opus 5.5 to Opus 4.8 Due to New Claude Code Refusals

Anthropic intentionally added two sentences to the Claude Code prompt for Opus 5.5 to prevent the recreation of third-party video games. According to developer reports, the model began refusing legitimate tasks, prompting a rollback to the previous Opus 4.8 version. This case exposed a structural vulnerability in hosted AI agents: providers can alter model behavior by editing prompts or safety filters without touching the developer's code, tools, or local prompts, causing working workflows to br

Read briefRead longformTelegram

Safety2026-11-02

AI company bankruptcies could expose users' private chats

BBC Future warns: when AI companies go bankrupt, user personal data databases, including chats with bots like Character.ai, become a liquid asset that can be sold to third parties. AI company bankruptcies turn accumulated personal data into a liquid asset: the buyer receives it along with rights to train models, build profiles, and run targeted advertising.

Read briefRead longformTelegram

CodingNVIDIA2026-11-02

Mago Team Open-Sources Local AI Film Studio Mago Studio

The Mago team has released Mago Studio under the AGPL-3.0 license. It is a local-first desktop AI video editor for Windows, macOS, and Linux that renders on the user's GPU via ComfyUI or in Mago Cloud. Mago Studio is a rare example of a complete production pipeline (versioning, review, frame-accurate timeline, 16-bit EXR/HDR, integration into existing animation and VFX processes) released as open source: independent studios and developers get a foundation for their own AI pipelines without being

Read briefRead longformTelegram

Research2026-11-02

GPT-6 Astra learns to edit neural network weights directly

MIT CSAIL researchers demonstrated that the GPT-6 Astra model can natively read and write neural network weights without gradient descent (the Weight-Space Programming approach), and that precise, gradient-free edits to fractions of a percent of weights eliminate defects like reasoning loops in Qwen3.5-0.8B, raising MATH-500 scores from 7.9% to 48.1%. If LLMs can directly optimize neural network weights rather than just prompts and code surrounding models, this changes the training paradigm: sym

Read briefRead longformTelegram

Multimodal2026-11-02

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

Matthias Nessler's team (TU Munich) released Video Diffusion Avatars, a method that fine-tunes the WAN2.2-Animate-14B video diffusion model into a 3D avatar from a single photo, with independent control over camera, facial expression, and pose. VDA demonstrates that the 'world' prior of video diffusion can be fine-tuned into a full 3D avatar that autonomously generates shadows, reflections, refractions, and hair dynamics—effects unavailable to explicit 3D renderers.

Read briefRead longformTelegram

Coding2026-11-02

DeepSeek publishes seven prompts for Harness Creator Mode

DeepSeek has published a collection of seven prompts in English and Chinese on its CDN for Creator Mode in DeepSeek Harness (DSH), describing how to ask the model to build ready-made applications — from a 3D landmark gallery and Transformer computation visualization to a video editor with chat-based editing. Creator Mode in DeepSeek Harness demonstrates that agentic runtimes are beginning to generate full GUI tools (video editor, PDF reader, 3D visualizer) from text descriptions without manual i

Read briefRead longformTelegram

CodingNVIDIA2026-11-02

Neuronaut Releases 35th Collection of Tools for MiniMax-H3

The 'Neuronaut' channel released its 35th collection of open-source tools for the MiniMax-H3 video model — ranging from character references via RefMod (Prompt Rewriter 0.32) to accelerators like Rapid-H3 and RelayHires; the open model base on Hugging Face has surpassed 6.1 million downloads. Without altering the base MiniMax-H3 model (the open MiniMaxAI/MiniMax-H3 base on Hugging Face, with over 6.1 million downloads), the ecosystem addressed several bottlenecks in the production cycle in the 3

Read briefRead longformTelegram

ResearchAnthropic2026-11-02

RWE-bench: Testing AI Models on Real-World Engineering Tasks

Startup Xplore released a closed benchmark, RWE-bench, consisting of 20 engineering tasks where a model in a sandbox with CadQuery, Gmsh, CalculiX, and Python must produce a working design, verified by a deterministic verifier using physics solvers; GPT-6 Astra achieved the best result at 37%, while four out of 20 tasks were not solved by any of the 11 models across 1,100 runs. RWE-bench provides an independent snapshot of where frontier models actually fall short of humans: on tasks with strict

Read briefRead longformTelegram

CodingAnthropic2026-11-02

X and SpaceXAI Art Director Teaches AI Agents to Create 'Faux UI'

Benji Taylor, Art Director at X and SpaceXAI, published the open-source skill Stageshow on October 31, 2026, which teaches AI agents (Grok Build, Claude Code, Codex, Gemini) to create polished static and animated interface mockups—'faux UI'—for landing pages, promo motion, and infographics. This is an example of codifying design 'taste' into agent instructions: 5 typical layout errors (overload, skeleton wireframe, 'window in a frame', flatness, lack of personality) are converted into rules, and

Read briefRead longformTelegram

Coding2026-11-02

Tencent releases Hyra 2.0: a research agent that accumulates knowledge

The Tencent Hunyuan team released Hyra 2.0 on November 2, 2026 — a research agent that independently develops scientific models, workflows, and tools, and accumulates reusable lessons in a shared knowledge base called Books. Hyra 2.0 formalizes the transition of research AI agents from "one-off answers" to an accumulating base of reusable tools.

Read briefRead longformTelegram

CodingOpenAI2026-11-02

New Internet Users: AI Agents

MIT Technology Review's overview "The agentic web is here" (Grace Huckins, November 2, 2026) describes the internet's shift to an agentic model, where autonomous AI agents are increasingly used instead of humans. The attention economy, which underpins advertising, subscriptions, and marketplaces, breaks down when the "user" becomes an agent: agents do not click on "sponsored" items (University of Toronto study), transitions from Google search results dropped by approximately 40% after the launch

Read briefRead longformTelegram

BusinessNVIDIA2026-11-02

Amazon to Sell Trainium AI Chips Outside AWS for the First Time

Amazon will begin selling Trainium AI chips to third-party data centers outside of AWS in late 2027, marking the first time they will be offered as a standalone merchant product. Amazon is entering the merchant silicon market for the first time—selling physical Trainium racks to third-party data centers—challenging Nvidia's dominance (81–90% of the AI accelerator market).

Read briefRead longformTelegram

MultimodalAnthropic2026-11-02

Tencent assembles WeVisCut — an agent that edits entire videos autonomously

Tencent Research introduced WeVisCut (arXiv 2610.38352, October 28, 2026) — an autonomous video editing agent based on Qwen3.5-9B that turns a script into a finished long-form narrative video, scoring 90.5 Overall on a 100-task benchmark, outperforming Claude Opus 5 and Gemini 3.1 Pro while using 33 times fewer tokens. WeVisCut demonstrates that a compact, specialized 9B model can match frontier LLMs (Claude Opus 5, Gemini 3.1 Pro) in complex agentic video editing with 33 times lower token consu

Read briefRead longformTelegram

ResearchOpenAI2026-11-02

NBER Study: AI Subscribers Easily Switch Providers

According to NBER Working Paper 35953, by April 2026, approximately 2.9% of American consumers pay for a major generative AI service, annual retention among major providers is 20–45%, and switching rates spike around new model releases. The paper provides a rare quantitative dataset at scale (174 million consumers, 114 billion transactions) supporting the conclusion that the generative AI market behaves like a market with low differentiation and low switching costs: at quality parity, providers

Read briefRead longformTelegram

CodingAnthropic2026-11-01

Andrey Karpathy Built a YouTube Clone in Two Hours

On October 31, 2026, Andrey Karpathy launched HackerTube (hackertube.org) — a YouTube clone featuring 6,454 videos across 150 topics, selected from approximately 40,000 user recommendations on Hacker News from 2007 to 2026. He stated that it was built in roughly two hours using Claude models without manually writing code. The case demonstrates a working 'orchestrator + worker' pattern in agentic development: the expensive Opus 5.5 ($4/M input, $20/M output, according to aicoder.com) manages arch

Read briefRead longformTelegram

MultimodalNVIDIA2026-11-01

HairDeck: Animatable Hair from a Single Portrait

Researchers from Tsinghua University, Carnegie Mellon University, and CUHK introduced HairDeck (arXiv:2610.40599) — a method for generating stylized hair cards and spring-bone chains for hair animation from a single image, which is more accurate and approximately 33 times faster than the autoregressive CHARM generator on the CHARM test set. HairDeck addresses a bottleneck in the production pipeline for stylized 3D characters — generating hair cards and spring-bone animations from a single refere

Read briefRead longformTelegram

CodingOpenAI2026-11-01

DeepSeek releases dsh-superagent coding agent prototype in response to Claude Code

The DeepSeek Harness team published an experimental coding agent prototype, dsh-superagent (MIT license, TypeScript), on GitHub, positioning it as a response to Anthropic's Claude Code. DeepSeek is intensifying competition in the AI agent development market, issuing an open challenge to Claude Code (Anthropic) and OpenAI tools.

Read briefRead longformTelegram

Business2026-11-01

Workers fired for AI are coming back to work and negotiating

SHL research and a Propeller Insights survey describe the trend of 'AI rebounding': 19% of workers in the US and 13% in the UK have encountered the return of employees who were laid off for automation, who are demanding higher salaries and job security guarantees for their return. Layoff decisions are made without assessing employees' actual skills, after which companies overpay to rehire the same personnel; this is an argument in favor of talent intelligence and internal mobility (47% of worker

Read briefRead longformTelegram

CodingAnthropic2026-11-01

AI agent built one app on 11 JS frameworks: price spread nearly doubled

Nolan Lawson published a benchmark where the AI agent Claude Opus 5.5 built the same Pokémon catalog on 11 JavaScript frameworks: the cheapest were Hono+htmx ($2.67) and Astro ($2.76), while the most expensive was Angular ($4.97, 9.76M tokens, 2 out of 5 runs failed the specification). The benchmark provides quantitative, reproducible data: framework choice now affects not only DX but also AI build costs and future token expenses for maintenance — the size of generated code has become a direct e

Read briefRead longformTelegram

BusinessNVIDIA2026-11-01

Chinese stock market loses $2.4 trillion as investors question AI boom

The Chinese stock market lost approximately $2.4 trillion in market capitalization since May 2026, according to Bloomberg, amid investor doubts about the profitability of the AI boom. The correction indicates that the Chinese AI boom is sustained by state investments, including 60 billion yuan from a new state AI fund, rather than consumer demand and profits.

Read briefRead longformTelegram

CodingMeta2026-10-31

React Compiler Engineer: Most Software Will Need to Be Rewritten Due to AI Agents

Laura (@poteto), a React Compiler core team engineer leading the Grok Bot team at xAI/SpaceX, published a post-manifesto on October 30, 2026, predicting that AI agents will cause code to arrive in a continuous stream, forcing most software to be restructured and rewritten, with codebases evolving into an “organic architecture” growing around human-created constraints and guardrails. When AI agents generate code in a continuous stream, the competitive advantage shifts from the “cleanliness” of in

Read briefRead longformTelegram

CodingNVIDIA2026-10-31

RTX 5080 Disappears from Official NVIDIA Store

The GeForce RTX 5080 vanished from the official NVIDIA US Marketplace on October 30, 2026, following the RTX 5090, leaving the RTX 5070 Ti as the top card in the catalog while US retail prices hover around $2,000 compared to the stated $999. The removal of flagship RTX 5090 and RTX 5080 cards from NVIDIA's own store, while retaining pages with prices of $999 and $1,999 and retail prices ranging from approximately $2,000 to $5,599, indicates that shortages and speculative markups on the Blackwell

Read briefRead longformTelegram

Research2026-10-31

Autumn Avito ML Cup: Recommendation Feed Ranking and a 1.5 Million ₽ Prize Pool

Avito has launched the 'Personal Ranking | Avito ML cup 2026' competition on the ODS platform, focused on predicting the scores of candidate listings in a recommendation feed using the NDCG metric, with a prize pool of 1.5 million ₽. Team registration is open until December 8, 2026. Avito is publicly testing personal ranking in a production recommendation feed—a rare instance where a major marketplace opens anonymized data from real outputs: 128-dimensional user and listing embeddings, and a his

Read briefRead longformTelegram

CodingNVIDIA2026-10-31

Open-Source VFX Tools for ComfyUI

Six open-source tools for ComfyUI were released at the Comfy Open Source Higgsfield competition (October 2026), including SameFace for measuring and correcting face drift, AInVFX Paint a plate for propagating first-frame edits across a shot, and MatteDesk for browser-based rotoscoping and matting. The package addresses common pain points in the VFX pipeline: temporal consistency of faces after inpainting, propagating first-frame edits across a shot in plate format with EXR/OCIO export, automatic

Read briefRead longformTelegram

MultimodalNVIDIA2026-10-31

About a dozen open tools released around MOSS-Image-SFT in one week

From October 22 to 31, 2026, the community released about ten tools for the open-source MOSS-Image-SFT model, including the Edit-HQ edit-LoRA, the PiD-SuperFast-HighRes pipeline with the NVIDIA PiD pixel decoder, and the moss4gb 4-bit build. MOSS-Image-SFT is becoming an open platform: within a week, LoRAs for instruction-based editing and RGBA transparency, the PiD pixel decoder upscaler, 4-bit ports, and layout nodes appeared.

Read briefRead longformTelegram

SafetyAnthropic2026-10-31

Anthropic: A catastrophic AI incident may exceed the capacity of the entire insurance industry

In its S-1 filing for an IPO with the SEC, published on October 26–27, 2026, Anthropic warned that in the event of a catastrophic AI incident, no single insurer—or possibly the insurance industry as a whole—would be able to provide coverage comparable to potential losses. For the first time, a major AI company has explicitly acknowledged in a regulatory document that its risks may exceed the total insurance capacity of the market.

Read briefRead longformTelegram

MultimodalNVIDIA2026-10-31

Pika releases H3 Flex — open video model with audio in 6 steps

Pika Labs has released the weights for Pika H3 Flex, a video generation model with audio fine-tuned on 33B MiniMax H3: generation takes 6 sampling steps instead of 50 — about 7 seconds for a 15-second clip in 768p. The open weights of the 33B video generation model with 6-step inference, multi-reference control (up to 12 files: up to 9 images, 3 videos, 3 audio files), and five separate audio tracks lower the barrier to entry for iterative production: drafts, targeted character and camera adjust

Read briefRead longformTelegram

SafetyOpenAI2026-10-31

OpenAI Updated Model Behavior Specification

On October 29, 2026, OpenAI updated the Model Spec, consolidating GPT priorities into five values and introducing a 'chain of command' to resolve conflicts between company rules, system prompts, and user requests. The public Model Spec with a formal hierarchy of priorities serves as a transparency reference for alignment practices, but it is established as a goal: sources do not contain measurements of model compliance with the 'chain of command,' so it is premature to build it into multi-agent

Read briefRead longformTelegram

ResearchNVIDIA2026-10-31

Stanford Introduces NCM: Text Generation 21 Times Faster Than Diffusion

Stanford Hazy Research researchers introduced Next-Concept Modeling, a non-autoregressive text generation framework that is 10–31 times faster than diffusion baselines on a single NVIDIA H100 with comparable quality, while training on 4–5 times fewer tokens. NCM offers a third path between autoregressive generation (token by token) and diffusion language models (iterative refinement of the entire sequence): generation by hierarchy levels requires only one transformer pass per level, reducing inf

Read briefRead longformTelegram

SafetyOpenAI2026-10-31

AI giants back unified opt-out for AI scraping in robots.txt

OpenAI, Google, Microsoft, and Anthropic announced support for a unified industry opt-out signal for AI scraping — a "policy token" embedded in robots.txt, proposed by Kevin Kelley (Anthropic) through the AIPREF working group in the IETF. A unified machine-readable standard is emerging through the AIPREF working group in the IETF: publishers no longer need to configure opt-outs separately for each AI provider, and providers gain a cross-cutting control mechanism.

Read briefRead longformTelegram

CodingOpenAI2026-10-31

EMBRACE: Vulnerable Browsers in Five AI Tools

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, where web pages gained privileges not available in standard browsers. Embedded browsers in AI tools have become a new attack surface: products have granted them privileges that standard browsers do not have—such as a bridge to the agent, access to host files, and control over the address bar—breaking the classi

Read briefRead longformTelegram

Coding2026-10-31

PostgreSQL 19 Delayed Due to Flood of Bugs from LLM Reviewers

The release of PostgreSQL 19 has been moved from the usual September–October window to November 19, 2026. Committer Robert Haas attributes the delay to LLM reviewers finding bugs en masse, which forced the rollback of three major features. LLM-assisted code review has shifted the development bottleneck: finding defects is no longer the constraint; manual verification and committing fixes by humans now are.

Read briefRead longformTelegram

SafetyOpenAI2026-10-31

EFF Defends Fair Use in In re OpenAI Case

On October 30, 2026, EFF filed an amicus brief in the In re OpenAI case, defending fair use in the training of generative AI models and warning that narrowing the doctrine under publisher pressure will restrict developers, journalists, researchers, and ordinary users. The decision in In re OpenAI may determine whether AI training on copyrighted materials remains within the scope of fair use.

Read briefRead longformTelegram

SafetyOpenAI2026-10-31

Black October: Nine AI Developer Platforms Hacked

According to Wiz's "Black October" report dated October 30, 2026, at least nine AI platforms for developers disclosed security incidents in one month, which Wiz assesses to be related to the targeted theft of LLM service API keys—LLMJacking. AI development infrastructure—hosting, CI/CD, agent platforms—has become a valuable target where key theft is monetized within minutes through the LLM access resale market; this changes requirements for secret storage, API key permission scopes, and monitori

Read briefRead longformTelegram

ResearchOpenAI2026-10-30

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

Edge0 has released an open, compact TTS model, Audio8-TTS-Realistic-0.6b, with 601,159,424 parameters (excluding the codec). It improves intonation and expressiveness compared to the Preview version, clones voices from a recording and its text without fine-tuning, and supports 23 languages, including Russian. This open 0.6B TTS model features a proprietary 44.1 kHz audio codec and a DualAR architecture (a 24-layer slow transformer predicts a semantic token per frame, while a 4-layer fast transfo

Read briefRead longformTelegram

BusinessAnthropic2026-10-30

Anthropic Renames Claude's Top Tiers

On October 30, 2026, Anthropic renamed Claude's top subscriptions: Max 5x ($100/mo) became Max, and Max 20x ($200/mo) became Ultra, without changing prices and automatically migrating existing subscribers, while also removing the 5-hour "sliding" limit for verified Ultra accounts while maintaining a weekly limit of 10× the Pro plan. The renaming removes the confusing Max 5x/20x multipliers, which only described volume within 5-hour windows rather than the overall subscription limit, a point of c

Read briefRead longformTelegram

MultimodalOpenAI2026-10-30

Construct 1.0.0: AI agent describes object in text, engine builds hologram in browser

Construct 1.0.0 (released 27.10.2026, Apache-2.0 license) is an open pipeline where an AI agent (Claude Code, Hermes, any MCP client, or OpenAI-compatible endpoint) describes a real object as a text 'scene' with dimensions, parts, assembly order, and joints, while a separate engine builds an animated hologram from it in construct.html, with export to glTF 2.0 and OBJ and an add-on for Blender 4.2+. The separation of 'agent describes — engine builds' makes the output portable: the agent does not

Read briefRead longformTelegram