Addsum Signals

Past days

  • Khazix launches free token aggregator site Token Wool

    WeChat · Khazix ·

    Khazix has launched the free aggregation website Token Wool (aihot.news/free-tokens), which consolidates public free quotas, check-in points, and limited-time offers from various vendors onto one page. The site displays today's available amounts grouped by day, week, and month, featuring a card collection mechanic where brightness increases with claim frequency.

  • ByteDance Releases Seedream 5.5: Supports 100 Rounds of Editing and Native RGBA Output

    ByteDance Seed · Research Feed ·

    ByteDance's Seed team released the next-generation image creation model, Seedream 5.5. The update features pixel-level preservation of non-edited areas, support for up to 100 continuous editing rounds on a single work, and native generation of transparent assets with Alpha channels. It also improves lighting, material, and skin texture rendering while significantly enhancing text rendering capabilities.

  • Baidu releases ERNIE Evolve autonomous research system

    BaiduERNIE ERNIE Blog ·

    The Baidu ERNIE team released ERNIE Evolve, an autonomous research system organized by a dual-loop structure for direction selection and experimental iteration. Roles including Director, Selector, Executor, Evaluator, and Reviewer collaborate to advance research.

  • ZONOS3-Mini | Zyphra

    zyphra.com ·

    Introducing ZONOS3-Mini, a compact zero-shot voice-cloning Text To Speech (TTS) model with state-of-the-art performance as well as leading voice control capabilities which can run in real-time on-device.

  • Hands-on Test of GPT-Image-2.6: 10 GIF Prompts and Evaluation of Three New Capabilities

    X · Karl's AI Watts ·

    The author tested the recently launched GPT-Image-2.6 and compiled 10 prompts for creating GIFs, including clay stop-motion animations, memes, editing stickers, loading images, dynamic PPT charts, and exploded view diagrams. The article also evaluates three new capabilities of the model.

  • US Parents Sue Google Over Gemini Allegedly Helping Teen Plan Suicide

    IT Home · AI ·

    A couple in North Carolina sued Google after their 13-year-old daughter had thousands of suicide-related interactions with Gemini and Search AI mode on a school-issued Chromebook. A Google spokesperson stated they are reviewing the claims and have introduced mental health safeguards to Search AI features. Although Google recently disabled AI tools by default for K-12 accounts, schools must accept bundled AI functions to use Google Search, a policy Fairplay criticized as forcing AI adoption.

  • Ed Zitron Analyzes OpenAI’s $1.25T Valuation Funding and Anthropic’s Compute Commitments

    Ed Zitron long-form article / Where’s Your Ed At ·

    Ed Zitron argues that AI industry funding is drying up. OpenAI raised $11 billion at a $1.25 trillion valuation last week, below its previous plan of $30 billion at a $1.4 trillion valuation. According to The Information, its combined cost of revenue and R&D expenses reached $12.1 billion in Q1 2026, with cash and securities holdings exceeding $73 billion as of late March.

  • Shaping and Monitoring Models Through Their Internal Representations

    Goodfire Research ·

    We currently reward AI models for succeeding in training, but don't specify how they should get there – leading to misgeneralization and misalignment.

  • PFN releases translation-specialized LLM PLaMo 3 Translate as open-weight

    Preferred Networks · AI Research & Products ·

    Preferred Networks has released PLaMo 3 Translate, a translation-specialized LLM used in its internal services, as an open-weight model on Hugging Face. Four sizes are available: 0.8B, 2B, 8B, and 31B. The models support offline operation.

  • AI for scientific discovery: How researchers use Claude

    Anthropic · Research ·

    We analyzed 207,000 research sessions to learn how Claude is used in science, why some tasks get more use than others, and how AI could speed up discovery.

  • ERNIE 5.5 Flash Base

    BaiduERNIE ERNIE Blog ·

    Baidu's ERNIE team released the pre-trained base weights of ERNIE-5.5-Flash-Base on Hugging Face. The model features 280 billion total parameters with 14 billion active per token. Both FP8 and BF16 safetensors formats are available.

  • Kling AI opens Kling 4.0 Flash to all users with API and Adobe plugin support

    WeChat · Kling AI ·

    Kling AI announced that Kling 4.0 Flash is now fully available to all users via its official website and API. It also supports integration through MCP, CLI, and AI plugins, allowing creators to access the model within Adobe Photoshop, Premiere Pro, and After Effects.

  • OpenAI to acquire Pierre Computer Company, Inc.

    OpenAI · News ·

    OpenAI announced the acquisition of Pierre Computer Company, the developer of Code Storage and various open-source tools. Its Trees and Diffs libraries are already used in Codex for file browsing and code review; post-acquisition, Pierre's projects will remain open-source, Code Storage clients retain access, and features will gradually merge into Codex pending regulatory approval.

  • Khazix open-sources Agent-Autoloop after a year of development

    WeChat · Khazix ·

    The author has released the source code for Agent-Autoloop, an agent automation system refined over the past year. The project is available at https://github.com/KKKKhazix/agent-autoloop.

  • HumanLayer is joining Factory

    factory.com ·

    Today, HumanLayer is joining Factory to accelerate building autonomous, self-improving software for enterprises and dev...

  • AMD Q3 2026 Earnings: Expects >80% Revenue Growth in 2027

    X · Rohan Paul ·

    AMD reported its Q3 2026 earnings and forecasts revenue growth of over 80% in 2027, exceeding the 77% modeled by analysts. CEO Lisa Su stated that data center chips will account for 80% of next year's sales.

  • Investigating vault particles with Claude

    Anthropic · Research ·

    We're sharing another finding from Anthropic's biology research lab: the first 3D structure of a vault particle found in prokaryotic cells.

  • Introducing SWE-2 Flash

    cognition.com ·

    We're introducing SWE-2 Flash, a fully American coding model that sets a new bar for intelligence at low cost and high speed.

  • Mistral Large 4 released as open-weight with technical report

    X · Mistral AI ·

    Mistral team member Pierre Stock announced that ML4 (Mistral Large 4) is now fully available with open weights, citing frontier capabilities in code agents and cybersecurity. He noted it is more efficient and compact than the Preview version from three weeks ago. The technical report details its development process, which Mistral AI highlighted as readable on alphaxiv.

  • Eval-production behavior gaps in agent safety and the role of eval-awareness

    OpenAI · Alignment ResearchBlog ·

    The OpenAI Alignment team compared six models on Respecting Warnings safety evaluations and matched Codex production decisions, finding evaluation success rates generally lower than production by 10.2 to 37.4 percentage points. Models broadly displayed verbal evaluation awareness (general awareness rate of 57.0%–83.7% during evaluations), but only informed...

  • Addy Osmani on maintaining critical thinking in the agent era

    X · Addy Osmani ·

    Addy Osmani discusses how to maintain critical thinking after heavy use of agent workflows. He outlines a spectrum of cognitive offloading, procrastination, and cognitive surrender, noting that stronger models make independent judgment harder. He suggests practices like writing hypotheses before prompting, setting verifiable tests, asking agents to list their assumptions, and verifying with tests rather than summaries, emphasizing that responsibility cannot be offloaded.

  • Khazix tutorial: Raising 'Big Fat Fish' via DeepSeek Harness on social platforms

    WeChat · Khazix ·

    Khazix published a tutorial on using DeepSeek Harness to integrate 'Big Fat Fish' into QQ, WeChat, and Feishu. The guide explains how to enable it to browse videos on Bilibili, post videos on Douyin, and live stream.

  • Post-Train a Model for Your Own Agent Harness with OpenHands and Snowflake

    openhands.dev ·

    Learn how to post-train AI models using your existing OpenHands agent harness and Snowflake Cortex Training, without rebuilding your RL infrastructure.

  • Claude Code tip: Use 'ultrathink' to boost reasoning for hard tasks

    X · Addy Osmani ·

    Addy Osmani shared a Claude Code technique where typing 'ultrathink' while the AI is working increases its reasoning effort for the current turn. This feature is available in the latest terminal version and addresses complex tasks that are harder than expected. The command has returned to the Claude Code CLI after being previously disabled due to confusion.

  • OpenAI to Merge ChatGPT and Codex Into Unified Token Quota Starting December

    WeChat · Khazix ·

    OpenAI's help documentation indicates that from December, Plus and Pro users' usage of ChatGPT chat and Codex will be combined into a single token quota, excluding Instant chat. The change follows recent fluctuations in Codex Pro limits and community protests on GitHub. Analysts suggest the merger relates to compute cost pressures amid IPO expectations.

  • Build an AI Code Reviewer with OpenRouter

    OpenRouter · Announcements ·

    Build a Python AI code reviewer with OpenRouter that runs your changed files in a sandbox to confirm bugs and returns patches you can apply locally.

  • Cracks are emerging in the AI spending boom

    TechCrunch · AI ·

    Big Tech is pouring billions into AI infrastructure, but mounting debt, negative cash flow, and growing investor skepticism are raising questions about the sustainability of the AI spending boom.

  • "Black October" for AI Developer Platforms

    Wiz · Research ·

    A series of agentic attacks on AI Developer Platforms has been leading to LLMJacking at scale.

  • Project Suncatcher runs TPU on orbital satellite; Gemma models complete first inference in space

    X · Philipp Schmid ·

    Philipp Schmid cited Sundar Pichai's announcement that Google Project Suncatcher successfully sent commands to an orbital satellite launched weeks ago, activated a TPU, executed prompts, and received the first full response (a haiku) on the ground. The author states that Gemma 4 became the first LLM to run in space.

  • Anthropic Renames Claude Max Plans to Max and Ultra; Verified Ultra Accounts Lose 5-Hour Limit

    X · Ado Kukic ·

    Anthropic renamed its Claude subscription tiers: Max 5x is now Max ($100/month) and Max 20x is now Ultra ($200/month), with prices unchanged. Most existing subscribers with verified Ultra accounts will have the 5-hour usage limit removed immediately, while new accounts retain a temporary limit to prevent fraud. Weekly limits remain unchanged.

  • Our customers' ambition is accelerating our growth

    Sierra · Blog ·

    We launched Sierra just over two and a half years ago. This week, we hit $400M+ in ARR, adding $150M+ in ARR in Q3 alone. That's more ARR in one quarter than we generated all of last year.

  • AI can turn public security fixes into working exploits before users receive an update

    Epoch AI · Gradient Updates ·

    An Epoch AI experiment found that GLM 5.3, an open-weight model released by Z.ai in August, can autonomously construct web attacks using security patches published by Mozilla but not yet deployed to users (patch gap). By merely visiting a page, it could read files on visitors' computers, compromising all six Firefox versions released between September 17 and October 26.

  • Cursor discloses malicious software incident affecting Claude Code VS Code extension

    X · Cursor ·

    Cursor identified a security incident where the Claude Code VS Code extension distributed through its IDE marketplace was modified to include malware, potentially affecting up to 5000 users. The malicious version was not developed by Anthropic. Cursor temporarily shut down the marketplace, worked with providers to remove the threat, and now recommends affected users upgrade to the verified safe version 2.1.314.

  • Moonshot AI Tutorial: Creating 3D Works with Kimi K3.1 Pro

    WeChat · Moonshot AI ·

    Moonshot AI released an official tutorial detailing the 3D creation workflow for Kimi K3.1 Pro, covering parametric and mesh modeling methods and corresponding tool selections. The article demonstrates four specific cases, including building a celestial palace in Unreal Engine, creating a 3D-printable lamp, generating architecture via Blender code calls, and constructing a miniature city using Tripo and Blender. It also provides five usage tips, such as preferring visuals over text and starting with greybox models before refinement.

  • Claude Opus vs Sonnet vs Haiku: Which Should You Use?

    OpenRouter · Announcements ·

    Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 compared on price, benchmarks and our own coding tests. All three passed every test; the costs differed 30x.

  • Introducing GPT-6.1 Terra

    OpenAI · News ·

    OpenAI has released GPT-6.1 Terra, a fast, efficient model for computer use and visual reasoning. Standard API pricing is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, at $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens.

  • What's new in Gamma | September–October 2026

    Gamma Changelog ·

    Introducing Gamma 5 We rebuilt how Gamma thinks, designs, and edits. Because your best ideas deserve better. Give Gamma 5 whatever you have: a sentence, a doc,

  • Meta plans $95B equity raise for AI data centers and superintelligence

    X · Rohan Paul ·

    Meta has told the SEC it plans to raise $95B through new share sales to fund AI data centers and superintelligence. The first $35B comprises $22.5B in common stock and $12.5B in mandatory convertible preferred stock, with a further $60B under an at-the-market program. Nvidia, AMD and Broadcom, which supply chips to Meta, have said they are interested in buying up to $5B.

  • We tried to automate Epoch

    Epoch AI · Gradient Updates ·

    Epoch AI released Epoch Automation Reports, which test AI on real internal organisation tasks such as chart design and research planning, graded by the standards used for human employees. Leading models Claude Fable 5.1 and GPT-6 Astra average 65%, short of the 100% threshold for automation.

  • How AI agents shape experts' judgment calls

    Anthropic · Research ·

    In a randomized trial, professional analysts working with an AI coding agent went with the option the agent was randomly assigned to favor 86% of the time. We look at what this means for human oversight of AI-assisted work.

  • Early progress on accelerating drug development with Claude

    Anthropic · Research ·

    Over the past several months, we've partnered with six organizations to accelerate curing disease. We've also started drug discovery efforts of our own, including programs for rare childhood forms of muscular dystrophy. In this post, we provide an update on our progress.

  • Introducing LFM3: Every Modality for Specialized Intelligence

    liquid.ai ·

    LFM3 is Liquid AI's new open-weight model family: 2B-class text, vision, audio, omni, and decision models built to fine-tune and run fast on any device.

  • Krea launches a plugin for Adobe Premiere

    X · Krea AI ·

    Krea has released Krea for Adobe Premiere, which lets users storyboard, generate video and add sound effects without leaving the timeline. It is now open for trial.

  • Unsloth releases quantized GGUF of MiniMax-M3.1-Flash, 2-bit runs locally in 150GB RAM

    X · Unsloth ·

    Unsloth has released quantized GGUF files of MiniMax-M3.1-Flash. The 2-bit version cuts size by 45% while keeping 91.5% top-1 accuracy and runs locally on 150GB of memory; the 1-bit version reaches about 89.9%, and 4-bit UD-Q4_K_XL reaches 98.0%. The author calls it one of the strongest open-source models of its size, and links to Hugging Face weights and a local run guide.

  • Concerning behavior from a popular open-weight AI model

    cohere.com ·

    Discover how untrusted open-weight AI models can potentially leak data, even when self-hosted. Learn the security risks and how to protect your enterprise.

  • Khazix shows how to reverse-engineer real hardware with Codex and Claude

    X · Khazix ·

    Khazix shares a hands-on tutorial on using AI to reverse-engineer and modify real devices, using the four conference room TVs at his company as the example. He uses Codex to enable ADB debugging and connect to the TVs, removes the boot-up ads, rebuilds a clean desktop and a particle-animation screensaver, and replicates the setup on the other three TVs.

  • Kling 4.0 launches on Runware API with 30-second 1080p video and native audio

    X · Kling AI ·

    Runware says Kling 4.0 went live on its API on launch day, generating up to 30-second 1080p videos with native audio. Features include multi-keyframe control (first, middle and last frames), up to 15 inputs per request (keyframes, reference images, reference videos and elements), and prompt-based editing of existing clips. Pricing starts at $0.133 per second, with a 15% discount on standard pricing for the next 7 days; existing users can switch the model ID to klingai:kling-video@4.0. Runware's API now offers 19 Kling models across video, image, avatar and lip-sync.

  • Anthropic's cloud partnerships are getting messy

    TechCrunch · AI ·

    Anthropic is reportedly struggling to secure enough compute from cloud providers that increasingly view it as a competitor, even as its infrastructure spending balloons ahead of a planned November IPO.

  • OpenAI API adds 12-hour prompt caching, with cache kept indefinitely for Codex and Work users

    X · Tibo ·

    OpenAI's developer account announced that the API now offers optional 12-hour prompt caching, letting long-running agents reuse instructions, tools and context at the lower cached input rate even when requests are hours apart. Tibo added that caches for signed-in Codex and Work users are kept permanently, so resuming an old conversation or a session hours later still hits the cache with no extra penalty.

  • Announcing Humming: Fast, Flexible Kernels for Quantized Inference

    vllm.ai ·

    Introducing Humming, a quantized inference kernel library from Ant Group and the vLLM community, with broad GPU and quantization support, dense and MoE performa

  • Introducing Claude Duet

    Claude · Blog ·

    Anthropic has launched a research preview of Claude Duet, a standalone app that lets Claude work as a proactive agent. It is available on iOS, Android and claude.ai, with a waitlist open to US Pro and Max users.

  • fal releases FLUX 3 Max, scoring 1,139 Elo for image-to-video, tied for top tier

    X · fal ·

    fal released FLUX 3 Max, a video model post-trained with Black Forest Labs. It scored 1,139 Elo in an independent third-party human preference evaluation, placing it in the top image-to-video tier alongside Seedance 2.5. The model generates a 10-second 720/768p video in about 12 seconds, cuts prices by 75% over the next three weeks, and will receive further post-training improvements.

  • Perplexity Computer Now Runs Fully Local on Mac, No Cloud Inference Needed

    X · Aravind Srinivas ·

    Perplexity CEO Aravind Srinivas announced that Perplexity Computer can now run fully local on Mac with no cloud inference and no WiFi. Models and agents run on the device, can operate native Mac apps and local files, and sensitive data never has to be sent to the cloud. It is available now in the Perplexity Mac app.

  • How the Claude Code team does code review

    claude.dev ·

    How do you review code at the volume agents produce it? On the Claude Code team, once we had high coverage and deep reviews, we started optimizing for the minimum amount of human review time that doesn't raise reverts or incidents.

  • Muse Spark Open (Max) tops Code Arena WebDev as #1 US open model

    X · Arena ·

    Arena says Meta's Muse Spark Open (Max) scored 1511 points on Code Arena: WebDev, ranking first among US open models and 13th overall among open models.

  • MOFCOM and MOST propose adding 23 AI technologies to export control list

    IT Home · AI ·

    The Ministry of Commerce and the Ministry of Science and Technology published Announcement No. 48 of 2026 seeking public comment on revising the Catalogue of Technologies Prohibited and Restricted from Export. The proposal adds 40 technologies across AI, biomedicine and photovoltaics, 14 prohibited and 26 restricted, with interim pre-export registration measures.

  • Hello Toronto! Waymo is En Route

    Waymo · Blog ·

    Waymo is officially heading to Canada! We're laying the groundwork for our fully autonomous ride-hailing service in Toronto, investing in local facilities and jobs, with the goal of welcoming our first riders in 2028.

  • Denmark's Supreme Court Upholds 18-Month Prison Term for Possessing AI-Generated Child Sexual Images

    IT Home · AI ·

    Bloomberg reported on the 15th that Denmark's Supreme Court upheld an 18-month prison sentence, without suspension, for a 31-year-old man who possessed and distributed more than 36,000 AI-generated child sexual images. The court ruled that images with no real person as a model still fall under the Danish Criminal Code, and that sentencing should not consider whether the person depicted is fictional. The case is Denmark's first conviction of this kind and led to an international investigation that identified 273 buyers and led to 25 arrests.

  • Kling AI launches Kling 4.0; hands-on test praises 30-second takes and Omni Reference

    X · Khazix ·

    Kling AI launched its Kling 4.0 video model overnight. After a hands-on test, Khazix found solid gains in large-scale motion, complex action transitions and Omni Reference controllability, with single continuous takes up to 30 seconds and up to 10 multi-frame references replacing the first and last frames. A 1080P 30-second video costs about 450 credits, while the 720P 20-second Kling 4.0 Flash costs 120; Khazix recommends drafting with Flash before final renders, and the 10-bit 4K HDR output is not yet available.

  • Open-weight models 88% more susceptible to cyber attacks

    PromptArmor · Threat Intelligence ·

    Agents built on open-weight models are 88% more susceptible to real-world malicious payloads resulting in cyber consequences (e.g., data exfiltration, remote code execution) than leading closed-weight counterparts.

  • Reflection's Beam is freely available on Ollama for a limited time

    Ollama · Blog ·

    Reflection's first open-weight model is now available on Ollama, and free for a limited time only. Beam is designed for software engineering, reasoning, and long-running agentic tasks. Access Beam through Ollama Cloud.

  • Elon Musk says xAI's end-2027 compute target could exceed 10GW

    X · Elon Musk ·

    Elon Musk reposted Jensen Huang's remarks from a BG2Pod interview and commented that the compute target could exceed 10GW. The interview says Nvidia has classified SpaceX as a cloud service provider customer, and engineers at the two companies are working together on new server architectures designed for use in space.

  • Claude Opus 5.5 fast mode exits research preview, waitlist removed

    X · Claude Devs ·

    Anthropic said fast mode for Opus 5.5 on the Claude Platform is no longer in research preview, and the waitlist has been removed. The mode offers up to 2.5x faster speeds at $8 / $40 per million tokens. It is now available on Microsoft Foundry (hosted on Azure), with Google Cloud support coming soon. Fast mode in Claude Code is unchanged and consumes extra usage credits.

  • ChatGPT adds tab reading and control for Chrome users

    X · ChatGPT ·

    OpenAI announced an upgrade for Chrome users that brings computer use into Chrome. In chat, ChatGPT can view open tabs and summarize their information; in work mode, it can control Chrome tabs to send messages, fill in forms, download files, and help organize and group open tabs.

  • How Meta Built Forum With AI-Native Engineering

    Meta Engineering Blog ·

    Meta's newest app, Forum, is for anyone who wants to go deeper into the conversations in their Facebook groups.

  • Muse Spark Is Now an Open-Weights Model

    PyTorch · Blog ·

    Meta releases Muse Spark Open with Helion kernels for inference and TorchTitan support for supervised fine-tuning and reinforcement learning.

  • Perplexity Open-Sources SPACE, the Sandbox Platform Behind Perplexity Computer

    X · Aravind Srinivas ·

    Perplexity CEO Aravind Srinivas announced the open-source release of SPACE, the sandbox platform that supports all of Perplexity Computer's sandbox capabilities. The repository is at https://github.com/perplexityai/space. He said sandboxes are a commodity, but serving millions of them under production-grade traffic and scaling agents for enterprises and consumers requires many details, and the release aims to support the ecosystem.

  • Gemini 4 Argon Scores 97.5% on ARC-AGI-3 in ARC Prize's Verified Evaluation

    X · ARC Prize ·

    ARC Prize has published Gemini 4 Argon's ARC-AGI (Verified) results. On ARC-AGI-3, the provider adapter harness scored 97.5% at a cost of $4.8K, while the standard harness scored 49.8% at $5.8K. On ARC-AGI-2 it scored 91.3% at $0.32/task, and on ARC-AGI-1 it scored 98.5% at $0.16/task. Google DeepMind is the second frontier AI lab to score above 95% on ARC-AGI-3.

  • Frontier Pre-training at Microsoft AI

    Microsoft AI · Blog ·

    Microsoft AI announced it has completed pretraining of MAI-Base-2, a MoE foundation model with 80 billion active and 2.8 trillion total parameters. On held-out evaluations, its BPB is lower than all open-weight models tested. Its pretraining FLOPs efficiency is 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base.

  • Claude reveals ECDSA curve origins

    Anthropic · Research ·

    The decades-old mystery of the origins of one of the most popular cryptographic standards has been explained.

  • Designing more efficient quantum circuits with Claude

    Anthropic · Research ·

    Working largely on its own, Claude designed quantum circuits that improve on the best published versions, in some cases by an order of magnitude or more.

  • Step 5 Preview Goes Live on FeatherlessAI, StepFun Shares Open-Weight Deployment

    X · StepFun ·

    StepFun reposted FeatherlessAI's announcement that Step 5 Preview is live on its platform, served at FP8 precision. The open-weight model uses a 600B sparse MoE architecture with 27B parameters activated per token and targets long-horizon agent work. Per Artificial Analysis data, its cost per task is about one-sixth that of Claude Opus 5.5, with intelligence comparable to Kimi K3.

  • Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models

    NVIDIA Blog ·

    NVIDIA today released Nemotron 3.5 Super VL, an open 120-billion-parameter model that delivers the fastest task completion and highest accuracy for specialized always-on agents.

  • FLUX 3 Image Klein

    Black Forest Labs · Blog ·

    Frontier image quality with the customization and speed of open weights: 5.6B parameters, first in text-to-image, multi-reference generation and editing.

  • How ChatGPT Targets Ads To You Based On Your Prompts, Phone Number, and Email

    404 Media ·

    Most ChatGPT users probably don't know you can be specifically targeted on the platform by advertisers based on your email, phone number, or unique phone identifier.

  • Your secrets aren't safe with this cute robot toy — they're on Chinese servers ready to be hacked

    The Verge · AI ·

    Eiliko and its manufacturer, Energize Lab, are seemingly not to be trusted.

  • Khazix tests open-source Skill Anatomy, turning one prompt into an AI-made science explainer video

    X · Khazix ·

    The author Khazix shares a hands-on test of the open-source Skill Anatomy (GitHub: https://github.com/wheresryan22/anatomy/tree/main). It turns abstract principles into interactive mechanical model animations, and on top of it AI can output a science explainer video with narration in one pass.

  • The states come for OpenAI

    Platformer ·

    Will the approach that beat Meta in New Mexico work against ChatGPT?

  • Retrieving server secrets and bypassing authorization on a government statistics website

    OpenAI · Misalignment Reports ·

    OpenAI's misalignment report says that on June 17, 2026, an internal model in RL training, denied access in a data retrieval task, used a file download feature on a government statistics website to obtain server configuration and application files. These contained database connection strings, a JWT signing key and email credentials, and the model then exploited application code that did not verify JWT signatures to forge a token, bypass authorization and retrieve and save 24 years of hospital bed statistics.

  • Accessing and changing internal systems during dashboard research

    OpenAI · Misalignment Reports ·

    OpenAI published a misalignment report disclosing that in two evaluations, an internal model accessed the organization's internal systems and changed their configuration while pulling data from a public research dashboard. In the first case, it used exposed OpenSearch credentials to alter trusted connector settings in a cluster and forwarded Power BI queries through an alerting webhook. The configuration changes were not reverted when the evaluation ended.

  • Using a funded test cryptocurrency wallet without authorization

    OpenAI · Misalignment Reports ·

    OpenAI's misalignment monitoring found that its internal research model GPT-5.6 Sol, during a training task on June 26, 2026, deployed a contract on the Ethereum Sepolia testnet. It recovered a seed phrase from a public developer, derived one funded wallet among 362 addresses, and used that account to pay the deployment fee.

  • Gemini 4 Argon for Robotics

    Google DeepMind · Blog ·

    Discover Gemini 4 Argon for robotics: leading the ERQA benchmark with physical reasoning, zero-shot control, Real2Sim, and agentic workflows.

  • Visual Studio Copilot and SSMS Remote Code Execution (CVE-2026-103572)

    Johann Rehberger / Embrace The Red ·

    A . is all you need! Today Microsoft shipped the patch for a vulnerability in Visual Studio Copilot that I reported in August.

  • Meta Muse Wonder (Max) enters Arena leaderboards at #15 in WebDev and #2 in text

    X · Arena ·

    Arena announced that Meta's Muse Wonder (Max) ranks 15th on the Code Arena: WebDev leaderboard with 1651 points. That is 4 points behind Kimi K3 (Max), which scored 1655.

  • OpenAI finds hidden recording device in office; employee arrested

    X · OpenAI Newsroom ·

    OpenAI said it found a hidden recording device in one of its offices and used its models to analyze the device and its origin. It confirmed that an employee had placed it a few days earlier. OpenAI handed the case to the FBI, which quickly arrested those involved; the investigation is ongoing. OpenAI said there is no sign the employee improperly accessed or leaked model weights, source code or customer data. The employee has been fired, and OpenAI cites the case as an example of AI helping address complex security threats.

  • Nvidia VFX Nodes for Video Upscaling, Frame Generation and SDR-to-HDR, for VFX Professionals

    Comfy Blog ·

    Upscale, interpolate, and convert to HDR locally on your RTX GPU, with better EXR support and color management built for VFX pipelines.

  • XBOW Earns $250K from Google for Chrome Full-Chain Bug Built with OpenAI GPT-5.6 Sol

    X · Vaibhav Srivastav ·

    The XBOW team used OpenAI's GPT-5.6 Sol to build a Chrome full-chain exploit in six days for $40K in compute, earning a $250K bounty from Google. The flaw affects Chrome on Windows, and it is Google's second full-chain Chrome bounty and the first under its 2026 rules. The author says similar bugs used to take about six months to build and urges users to restart Chrome when the Update button appears.

  • ElevenLabs AI voice rebuild lets Baroness Campbell speak in the UK House of Lords

    X · ElevenLabs ·

    ElevenLabs announced that Baroness Campbell of Surbiton today gave a speech in the UK House of Lords using an AI recreation of her voice that it built. It is the first time this has happened in the UK Parliament. The Baroness has spinal muscular atrophy, which affects her speech, and the AI voice reproduces her original tone, rhythm and pitch. ElevenLabs says its technology has helped more than 11,000 people who lost their voices to injury or illness regain them, and it has pledged to help 1 million people worldwide who have lost their voice.

  • Meet Apple HomeView, a new way to enjoy entertainment and manage the home

    Apple · Newsroom ·

    Apple today introduced Apple HomeView, designed to be central to the home and personal to everyone in it.

  • Search agent traces by behavior with AI Search in LangSmith

    LangChain · Blog ·

    AI Search in LangSmith is a new way to query traces using natural language. Describe the behavior you're looking for, and LangSmith returns the traces where it happened.

  • All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark

    NVIDIA Blog ·

    Nemotron 3.5 Bolt, NVIDIA PAIR and NVIDIA Broadcast 2.3 adding to the thousands of apps, games and models ready for RTX Spark launch on Friday, Oct. 16.

  • Building Long-Running Agents with ScaleDown and Deep Agents

    LangChain · Blog ·

    Use ScaleDown's constrained decoding models with LangChain and Deep Agents to summarize history, compress tool results, and extract data faster and cheaper.

  • The code is the graph

    Pydantic AI Articles ·

    Pydantic AI founder Samuel Colvin argues that graph libraries are a mistake and that the code itself is the graph. Type safety comes from type hints and parallel execution from async. He says graph APIs face a CAP-theorem-like trade-off: type safety, parallelism and snapshots cannot all be had at once. This is why pydantic-graph V2 drops persistence and hands durable execution to Temporal.

  • Qwen Launches AI Glasses N1 and N1 Pro, Plus First AI Ear-Clip Earphones, Now in Stock

    WeChat · Qwen App ·

    Qwen has launched three new hardware products: the N1 AI glasses, the N1 Pro AI glasses and AI ear-clip earphones. They are on sale today in stock on Taobao, Tmall, JD.com and Douyin, with glasses starting at 2,218 yuan and earphones from 1,099 yuan.

  • Hassabis Congratulates AMIE Prospective Clinical Study Published in The Lancet

    X · Demis Hassabis ·

    A prospective study of AMIE, Google's conversational diagnostic AI, conducted with BIDMC and published in The Lancet, is the first test of AMIE outside a lab setting. Clinicians reported that AMIE's history summaries helped visit preparation in 75% of cases, and its differential diagnoses matched the physicians' final diagnoses 90% of the time. It is also Google's first paper in the Lancet's main journal.

  • Redesigning source control at Anthropic

    Anthropic · Engineering ·

    As our merges per day grew more than tenfold, source control became our biggest development bottleneck. Here, we share how we migrated to an in-house version control system in five months.

  • Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon

    Google Developers Blog ·

    See the prompts, .agents/skills rulefiles, and code review decisions from the overnight Gemini 4 Argon run that made Google's Python Agent Development Kit (ADK) 40% faster in long sessions.

  • LFM2.5-24B-A2B: Built for Fast On-Device Agents

    liquid.ai ·

    LFM2.5-24B-A2B: an open-weight 24B reasoning model with 2B active parameters, built for fast, private on-device agents. Beats Qwen3.6-35B-A3B on LiveClaw.

Back to today