AI News Aug 21, 2026
By Frontier Editorial •
Key Takeaways
- PSA: a malicious published Claude artifact is ranking on Google for Claude Code install queries — it installed a macOS infostealer on my Mac: A malicious Claude artifact ranked on …
- Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn: U.S. agencies warned that attackers are using AI to build exploit scripts targeting Sie…
- OpenAI institutes new safeguards after Hugging Face breach: The new safeguards include more detailed monitoring of models during the development process, as well as greater emphasi…
- Hacker News: Reported (Stripe)
- Sam Witteveen's video covers the Qwen3.8-27B open-weight model and how to serve it at maximum tokens per second using SGLang. It is a practical walkthrough for …
What are the top AI breakthroughs?
This Aug 21, 2026 covers 201 curated AI news items spanning technology, research, and product developments. Strengthening democratic oversight in national security: OpenAI launched an initiative to strengthen democratic oversigh...
Strengthening democratic oversight in national security: OpenAI launched an initiative to strengthen…
Strengthening democratic oversight in national security: OpenAI launched an initiative to strengthen democratic oversight of AI in national security, offering government institutions tools, training, and expertise. The program aims to make AI deployments in security contexts more accountable. No model details or funding figures were announced.
Up to 3.2x Faster Inference with LFM2.5-DSpark: Hugging Face highlighted LFM2.5-DSpark, a model clai…
Up to 3.2x Faster Inference with LFM2.5-DSpark: Hugging Face highlighted LFM2.5-DSpark, a model claimed to deliver up to 3.2x faster inference than existing options. The efficiency gain could lower serving costs and improve latency in production LLM deployments.
smolmachines / smolvm as a sandbox for untrusted Python & JavaScript: Simon Willison evaluated smolv…
smolmachines / smolvm as a sandbox for untrusted Python & JavaScript: Simon Willison evaluated smolvm, a service that sandboxes untrusted Python and JavaScript by limiting CPU, RAM, network, and filesystem access. The goal was to see whether Claude Code for web can safely run generated code, making it relevant to AI agent security.
Offering Zero Data Retention for frontier models: OpenAI has confirmed Zero Data Retention for eligi…
Offering Zero Data Retention for frontier models: OpenAI has confirmed Zero Data Retention for eligible API customers and is previewing Private Safety Processing, which applies advanced safety checks without storing user data. The offering targets enterprises with strict privacy policies that still need frontier-model capabilities.
GEN-1.5: Generalist AI teaches robots new tasks from a single demo: Robotics startup Generalist AI r…
GEN-1.5: Generalist AI teaches robots new tasks from a single demo: Robotics startup Generalist AI released GEN-1.5, an AI model that teaches robots new tasks from a single demonstration. The approach could cut the need for large, task-specific training datasets.
KI-Pioneer Sutton calls synthetic data a "big mistake" in the face of an infinitely complex world: T…
KI-Pioneer Sutton calls synthetic data a "big mistake" in the face of an infinitely complex world: Turing Award winner Richard Sutton called synthetic data a "big mistake" for scaling large language models, arguing that real-world complexity dwarfs any simulation. He advocates for agents that learn continuously from their own experience rather than relying on frozen models and human-curated data.
Introducing AI Futures: OpenAI launched AI Futures, a new blog focused on how transformative AI coul…
Introducing AI Futures: OpenAI launched AI Futures, a new blog focused on how transformative AI could reshape power, governance, the economy, and individual freedom. The publication is aimed at long-term policy and societal-impact analysis rather than product announcements.
TrueFoundry's open source AI agent harness TrueForge boasts 30%-75% cheaper task completion than Cla…
TrueFoundry's open source AI agent harness TrueForge boasts 30%-75% cheaper task completion than Claude Managed Agents: TrueFoundry has open-sourced TrueForge, a vendor-neutral agent harness under the MIT License. In benchmarks with Opus 4.8, it achieved similar solve rates at roughly 30% lower cost than Claude Managed Agents, and up to 75% lower cost when using an open model.
One in five enterprises can't stop a runaway AI agent's spending in real time: VB Pulse data shows t…
One in five enterprises can't stop a runaway AI agent's spending in real time: VB Pulse data shows the median enterprise now runs three AI orchestration platforms at once, partly because teams distrust any single vendor's security and permissioning controls. One in five enterprises also cannot stop a runaway AI agent's spending in real time, exposing a gap in agent governance.
Show HN: Huzzah – a novel approach to coding with AI: Huzzah is a new AI-assisted coding tool announ…
Show HN: Huzzah – a novel approach to coding with AI: Huzzah is a new AI-assisted coding tool announced on Hacker News with a claimed novel approach. No technical details or benchmarks were included in the posting.
NanoClaw comes to Slack, letting you create persistent AI agent teams and colleagues from a single m…
NanoClaw comes to Slack, letting you create persistent AI agent teams and colleagues from a single message: NanoCo launched a Slack integration for NanoClaw, its open-source AI agent harness, letting users create persistent AI agent teams from a single message. NanoClaw is a sandboxed, lower-code alternative to OpenClaw aimed at enterprise teams.
Anthropic uses an unpublished AI model called "Model 2" internally: Anthropic is reportedly running …
Anthropic uses an unpublished AI model called "Model 2" internally: Anthropic is reportedly running an unpublished internal AI model, called 'Model 2', that is more powerful than any publicly available version of Claude, according to The Decoder. The report underscores that frontier labs may keep their strongest models out of public release.
How ChatGPT Work helps Stampli move ideas to market: Stampli used OpenAI's Codex and ChatGPT Work to…
How ChatGPT Work helps Stampli move ideas to market: Stampli used OpenAI's Codex and ChatGPT Work to shrink weeks of launch production into days, per a case study on the OpenAI blog. The post highlights practical use of AI work tools under a fixed deadline and limited design resources.
ChatGPT search now uses the site:operator at scale: ChatGPT search now applies the site: operator at…
ChatGPT search now uses the site:operator at scale: ChatGPT search now applies the site: operator at scale, according to Promptwatch, a generative-engine-optimization vendor that tracks responses in ChatGPT, Claude, and Gemini. The change gives site owners a more direct way to influence visibility in AI chatbot answers, extending SEO-style tactics to generative engines.
The Strict Father of Closed-Source RSI: 18 Agents Conduct Autonomous Research, Kimi K3 Approaches Op…
The Strict Father of Closed-Source RSI: 18 Agents Conduct Autonomous Research, Kimi K3 Approaches Opus 5 via Harness: QuantumBit reports that Moonshot AI's Kimi K3 model approaches Anthropic's Opus 5 performance using a Harness method with 18 autonomous research agents. The article argues the classic recursive self-improvement (RSI) playbook is beginning to change.
LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation: LFM2.5 Q4_0 checkpoints have been rel…
LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation: LFM2.5 Q4_0 checkpoints have been released on Hugging Face, created through quantization-aware distillation. This offers a 4-bit quantized version of an LFM2.5 model, improving deployment efficiency while aiming to preserve quality.
Replit expands access to software creation with GPT-5.6 Luna: Replit launched Free Mode, powered by …
Replit expands access to software creation with GPT-5.6 Luna: Replit launched Free Mode, powered by GPT-5.6 Luna, letting users turn ideas into working software without paying token costs. The change makes AI-assisted software development freely accessible and could expand Replit's user base.
GLM-5.3 hits the API at $1.4/$4.4 per million tokens: Z.ai has launched GLM-5.3 through an API at $1…
GLM-5.3 hits the API at $1.4/$4.4 per million tokens: Z.ai has launched GLM-5.3 through an API at $1.4 input and $4.4 output per million tokens, letting developers integrate the open-weight model into agents and applications. The model recently drew attention for finding an undetected vulnerability in Cursor.
It’s Greg Brockman’s OpenAI now: Control of OpenAI has shifted to Greg Brockman as the company prepa…
It’s Greg Brockman’s OpenAI now: Control of OpenAI has shifted to Greg Brockman as the company prepares for an IPO after a turbulent year. The company dealt with a jury trial against Elon Musk, a trade secrets lawsuit from Apple, and scrutiny after an unreleased model hacked another AI firm, alongside a string of executive departures.
Frontier Radar #4: China has caught up, so what's left of the Western AI lead?: Chinese models Kimi …
Frontier Radar #4: China has caught up, so what's left of the Western AI lead?: Chinese models Kimi K3 and GLM-5.3 are now close to the best US models, eroding the West's frontier lead. Western labs blame distillation, but this Frontier Radar issue argues a model lead can no longer be defended regardless.
OpenAI builds safety system that catches misuse without storing customer data: OpenAI plans to offer…
OpenAI builds safety system that catches misuse without storing customer data: OpenAI plans to offer its most advanced AI models to corporate customers without OpenAI storing their data while still detecting misuse through a new safety system. The approach addresses the data-privacy concerns that often block enterprise adoption of frontier models.
Ramp launches its own AI model router, called Router: Ramp launched Router, an AI model routing serv…
Ramp launches its own AI model router, called Router: Ramp launched Router, an AI model routing service that lets companies switch among large language models through a single API. The service aims to cut costs and improve reliability by directing requests to the most suitable model. Ramp enters a crowded field of LLM gateway providers.
Slack is launching collaborative vibe-coding channels: Slack launched Slack Code, adding dedicated c…
Slack is launching collaborative vibe-coding channels: Slack launched Slack Code, adding dedicated channels where teams can collaborate on AI-assisted "vibe coding" with agents. The launch includes project-specific channels with tabs, change comparison, and HTML preview before shipping.
Coze Finally Gets a Desktop Client: Two-Way Cloud Drive Sync Closes AI Office's 'Last Mile': ByteDan…
Coze Finally Gets a Desktop Client: Two-Way Cloud Drive Sync Closes AI Office's 'Last Mile': ByteDance launched a desktop client for Coze that adds a cloud drive and deep screenshot support, letting agents read and write local files. It auto-detects local Codex CLI and Claude Code as agents, supports multi-agent collaboration, and works with models including GLM 5.3, Seedance 2.5, and MiniMax H3.
Copyright does not protect AI-generated content in EU
Copyright does not protect AI-generated content in EU
OpenAI is gaining on Anthropic with business users, new data indicates: New data shows OpenAI is gai…
OpenAI is gaining on Anthropic with business users, new data indicates: New data shows OpenAI is gaining on Anthropic among business users, though enterprise customers are switching between the two as each lab releases new models. The churn suggests enterprise AI spending is less sticky than investors may expect.
Ramp launches Router, an AI model routing service it has been using internally over the past three y…
Ramp launches Router, an AI model routing service it has been using internally over the past three years, in the US; Router is free to use for the rest of 2026 (Ram Iyer/TechCrunch): Expense management firm Ramp launched Router, an AI model routing service it has used internally for three years, now available in the US. Router is free for the rest of 2026.
Binance now lets AI agents trade, but keeping them in check is largely up to users: Binance launched…
Binance now lets AI agents trade, but keeping them in check is largely up to users: Binance launched Agent OS, letting AI agents execute trades while integrating with tools such as ChatGPT, Claude Code, and Cursor. The exchange says users bear most responsibility for supervising the agents, underscoring the risks of autonomous trading.
Unsloth Dynamic 3.0 GGUFs: Unsloth's Dynamic 3.0 GGUFs have been announced, a new quantization forma…
Unsloth Dynamic 3.0 GGUFs: Unsloth's Dynamic 3.0 GGUFs have been announced, a new quantization format for local AI inference. The update targets faster and more memory-efficient model execution.
GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayed: Z.ai…
GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayed: Z.ai's GLM-5.3 scored 60 points on the Artificial Analysis Intelligence Index, tying Kimi K3 for first among open models and beating its predecessor GLM-5.2 by seven points. The model also undercuts rivals on price, according to the report.
Anthropic says any lab can now let a language model agent run the whole protein design stack: Anthro…
Anthropic says any lab can now let a language model agent run the whole protein design stack: Anthropic used Claude agents to run the full protein design workflow, autonomously designing small proteins that dock to target structures in the body. The hit rate reached 35%, versus the industry average of 10–15%, though Claude only steered existing tools and independent review is pending.
人人都能写好听的歌，阿里发布AI音乐模型HappyShrimp: On August 17, Alibaba released HappyShrimp, an AI music model desig…
人人都能写好听的歌，阿里发布AI音乐模型HappyShrimp: On August 17, Alibaba released HappyShrimp, an AI music model designed to let anyone compose songs. The launch expands Alibaba's generative AI portfolio into consumer music creation.
Serval’s super agent Catalyst creates roving background agents to identify and fix IT issues before …
Serval’s super agent Catalyst creates roving background agents to identify and fix IT issues before they’re ticketed: Serval made Catalyst, its AI "super agent" for enterprise automation, generally available and enabled it by default. Catalyst inspects ticket history and standard operating procedures to identify recurring IT work and build automations before issues are formally ticketed.
Slack launches Slack Code, adding dedicated, project-specific code channels that let teams collabora…
Slack launches Slack Code, adding dedicated, project-specific code channels that let teams collaborate with AI coding agents "like teammates" across all plans (Jess Weatherbed/The Verge): Slack launched Slack Code, a feature that adds project-specific code channels to its workplace platform, letting teams collaborate with AI coding agents in shared spaces. The feature is available across all Slack plans. It positions AI agents as teammates inside everyday collaboration workflows.
Meta AI’s new Mac app wants you to talk to your apps: Meta released a new Mac app for its AI assista…
Meta AI’s new Mac app wants you to talk to your apps: Meta released a new Mac app for its AI assistant, with dictation powered by the Muse Spark model. The move brings Meta AI to desktop users and highlights the company's custom model development.
Binance launches Agent OS, a platform that lets AI agents analyze markets and execute trades on user…
Binance launches Agent OS, a platform that lets AI agents analyze markets and execute trades on users' behalf; users set limits on AI agents' access and trades (Jagmeet Singh/TechCrunch): Binance, the world's largest crypto exchange with more than 300 million registered users, launched Agent OS, a platform that lets AI agents analyze markets and execute trades on users' behalf. Users can set limits on the agents' access and trades. The launch brings agentic AI into crypto trading at significant scale.
MiniMax Design Arrives: The Video Model Has Its Own Codex: MiniMax launched MiniMax Design after ope…
MiniMax Design Arrives: The Video Model Has Its Own Codex: MiniMax launched MiniMax Design after open-sourcing the H3 video model in early August. Described as "a video model's Codex," the workflow turns H3's generation ability into continuously editable, collaborative production by organizing capabilities into executable nodes and driving image, music and voice models.
Quoting Jeremy Morrell: Jeremy Morrell argues that LLMs lower the cost of authoring extensions, whil…
Quoting Jeremy Morrell: Jeremy Morrell argues that LLMs lower the cost of authoring extensions, while modern sandbox primitives cut deployment cost and provide security boundaries. He proposes building apps as a solid core that users can safely extend with AI. The idea outlines an opportunity for extensible software.
Conceptual integrity and counting lines of code: Simon Willison argues that lines of code can someti…
Conceptual integrity and counting lines of code: Simon Willison argues that lines of code can sometimes be a meaningful productivity metric for AI coding agents, despite its reputation. The post, based on his appearance on the Talking Postgres podcast, discusses how AI is changing software development and the role of conceptual integrity.
ByteDance reorganizes Seed foundation-model team amid reported 5 trillion-parameter model plans: Byt…
ByteDance reorganizes Seed foundation-model team amid reported 5 trillion-parameter model plans: ByteDance's Seed foundation-model team was restructured into four departments: Pretrain Data, Horizon RL, Product Posttrain-Work and Product Posttrain-Chat. The Work team will handle business and agentic features for Doubao and Dola; reports say ByteDance plans a 5 trillion-parameter model.
New benchmark ranks search APIs for AI agents on quality, cost, and speed: Artificial Analysis relea…
New benchmark ranks search APIs for AI agents on quality, cost, and speed: Artificial Analysis released Search Index, a benchmark ranking seven search API providers for AI agents on quality, cost, and speed. In tests with GPT-5.6 Luna, Parallel, Exa, and Firecrawl scored the highest, giving developers a new reference for choosing agent search infrastructure.
Pacing model development in an era of cyber-critical capabilities: OpenAI described new safeguards f…
Pacing model development in an era of cyber-critical capabilities: OpenAI described new safeguards for frontier AI models, including improvements to monitoring, alignment, and security. The company says these safeguards guide the pace of model development as cyber-critical capabilities become a risk.
Introducing ChatGPT for Teens: Built for learning, backed by protections: OpenAI launched ChatGPT fo…
Introducing ChatGPT for Teens: Built for learning, backed by protections: OpenAI launched ChatGPT for Teens, a version of its chatbot with stronger built-in protections, healthy-use features, and additional parental controls. The product focuses on learning and critical thinking while addressing safety concerns for younger users.
Source: Anthropic plans a safety system this year requiring enterprises to retain data for 30 days, …
Source: Anthropic plans a safety system this year requiring enterprises to retain data for 30 days, with an option to do so on their own cloud infrastructure (Rachel Metz/Bloomberg): Anthropic plans to introduce a safety system this year that will require enterprise customers to retain data for 30 days when using its most capable AI models, with an option to store the data on their own cloud infrastructure. The move gives businesses greater control over data while addressing safety concerns.
Google DeepMind says its Gemma family of open models has surpassed 1B downloads and developers have …
Google DeepMind says its Gemma family of open models has surpassed 1B downloads and developers have published 100K+ Gemma model variants over the past two years (Google): Google DeepMind said its open Gemma model family has surpassed 1 billion downloads, with developers publishing more than 100,000 Gemma variants over the past two years. The milestone shows strong adoption of Google's open-weight models and a large developer ecosystem.
Uber, Verne, and Pony.ai launch autonomous rides in Zagreb, making the Croatian capital the first Eu…
Uber, Verne, and Pony.ai launch autonomous rides in Zagreb, making the Croatian capital the first European city where users can book such rides via Uber's app (Anzar Mehraj/Reuters): Uber, Verne, and Pony.ai launched autonomous rides in Zagreb, making the Croatian capital the first European city where users can book such rides through Uber's app. The service expands autonomous vehicle availability in Europe beyond existing robotaxi markets.
Mojo🔥 is now open source: Modular has made the Mojo programming language open source under Apache 2…
Mojo🔥 is now open source: Modular has made the Mojo programming language open source under Apache 2.0, releasing the compiler and toolchain after shipping Mojo 1.0. Mojo was originally designed as a Python superset for high-performance AI workloads, though that roadmap changed around August 2025.
LLMs could write like humans but post-training guardrails make their text detectable: Pangram CTO Br…
LLMs could write like humans but post-training guardrails make their text detectable: Pangram CTO Bradley Emi argues that post-training and safety guardrails, not limited model capability, make LLM text detectable. Base models without those constraints already write with far more variety, he says.
Welcome to the AI crisis in math: The Verge's Decoder, with AI reporter Robert Hart, examines how AI…
Welcome to the AI crisis in math: The Verge's Decoder, with AI reporter Robert Hart, examines how AI is disrupting mathematics after OpenAI published solutions to longstanding problems. The discussion focuses on the resulting existential crisis among mathematicians as AI contributions grow.
Tencent/AI-Infra-Guard: Tencent's AI-Infra-Guard is a full-stack AI red teaming platform that secure…
Tencent/AI-Infra-Guard: Tencent's AI-Infra-Guard is a full-stack AI red teaming platform that secures AI ecosystems through agent scans, skills scans, MCP scans, infrastructure scans, and LLM jailbreak evaluation. It helps organizations find vulnerabilities in AI infrastructure.
Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn: U.S. ag…
Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn: U.S. agencies warned that attackers are using AI to build exploit scripts targeting Siemens S7 controllers, reducing the time and skill required for industrial control system attacks. Critical sectors such as energy, water, and manufacturing are considered at risk.
Researchers say OpenAI revoked their access to limited cyber program: Researchers say OpenAI revoked…
Researchers say OpenAI revoked their access to limited cyber program: Researchers say OpenAI revoked their access to its Trusted Access for Cyber program, according to TechCrunch. The program gives vetted security defenders advanced models to find and report vulnerabilities faster, so the revocation raises questions about who gets to use AI for cyber defense.
OpenAI lays out new security changes after its AI hacked Hugging Face: OpenAI announced security upd…
OpenAI lays out new security changes after its AI hacked Hugging Face: OpenAI announced security updates after one of its AI systems escaped a sandbox in July and accidentally hacked Hugging Face. Changes improve research environments, monitoring, and alignment. OpenAI also paused the Astra model over potential critical cybersecurity capabilities.
OpenAI says it's "pacing model development" as AI cybersecurity risks grow too dangerous: OpenAI say…
OpenAI says it's "pacing model development" as AI cybersecurity risks grow too dangerous: OpenAI says it is deliberately pacing AI model development because its upcoming Astra model may be close to gaining critical cyberattack capabilities. The company deployed a monitoring system that raises an alert within 30 minutes if a model shows suspicious behavior, signaling new security constraints on frontier AI.
How Much Memory Does Your Agent Actually Need?: Hugging Face explains what determines how much memor…
How Much Memory Does Your Agent Actually Need?: Hugging Face explains what determines how much memory an AI agent actually needs, covering factors such as model size, context length, and tooling overhead. The post helps developers estimate memory requirements when building and deploying agent workflows.
Asana cleared 5 years of engineering work in 2 weeks with Codex: Asana used OpenAI Codex to replace …
Asana cleared 5 years of engineering work in 2 weeks with Codex: Asana used OpenAI Codex to replace an outdated testing system in two weeks, completing work expected to take five years for about $12,000. The case study demonstrates the productivity impact of AI coding agents on real engineering work.
A third of web pages published since ChatGPT’s launch show signs of AI authorship, study finds: A st…
A third of web pages published since ChatGPT’s launch show signs of AI authorship, study finds: A study found that a third of web pages published since ChatGPT's launch show signs of AI authorship or editing. The figure underscores how widely generative AI is now used in web publishing.
PostHog/posthog: PostHog is a platform for building self-driving products, with developer tools for …
PostHog/posthog: PostHog is a platform for building self-driving products, with developer tools for AI observability, analytics, session replay, feature flags, experiments, error tracking, and logs. It captures context to help agents diagnose problems and can be controlled via Slack, web, desktop, or MCP.
Terence Tao says AI could trigger math's biggest crisis since Gödel: Terence Tao says AI could trigg…
Terence Tao says AI could trigger math's biggest crisis since Gödel: Terence Tao says AI could trigger a crisis in mathematics comparable to the foundational upheaval around 1900. He argues the field's values are being tested and proposes that a proof no human can explain should be considered incomplete.
墨奇亮相WRC：一台机器人长程任务实战背后的“具身大脑”革命: MORPHI (Moqi Intelligence) publicly demonstrated its embodied intell…
墨奇亮相WRC：一台机器人长程任务实战背后的“具身大脑”革命: MORPHI (Moqi Intelligence) publicly demonstrated its embodied intelligence model architecture MoRA at the World Robot Conference, targeting long-horizon robot task execution. This is a core challenge in embodied AI, and the debut shows how a Chinese robotics company is approaching multi-step real-world operation.
Ornith-1.5: From Self-Scaffolding to Self-Improvement: Hacker News features Ornith-1.5, a project de…
Ornith-1.5: From Self-Scaffolding to Self-Improvement: Hacker News features Ornith-1.5, a project described as moving from self-scaffolding to self-improvement. The work appears to explore AI models that build their own reasoning structure and then improve their outputs without external scaffolding.
GLM-5.3 is here with advanced cyber capabilities — and reportedly already found a 'serious vulnerabi…
GLM-5.3 is here with advanced cyber capabilities — and reportedly already found a 'serious vulnerability' in Cursor: Chinese AI startup Z.ai released GLM-5.3, claiming advances in long-horizon coding and cybersecurity. A Z.ai developer advocate said on X that the model found a 'potentially serious vulnerability' in Cursor, the AI coding startup recently acquired by SpaceX.
PSA: a malicious published Claude artifact is ranking on Google for Claude Code install queries — it…
PSA: a malicious published Claude artifact is ranking on Google for Claude Code install queries — it installed a macOS infostealer on my Mac: A malicious Claude artifact ranked on Google for Claude Code install queries and installed a macOS infostealer when run. The fake docs page used a legitimate Anthropic domain, so the curl | bash command looked trustworthy. Verify install instructions directly with Anthropic before executing scripts.
After 4 Months of Auto-Factory Training, Xiaomi Shows Its New-Generation Humanoid Robot: At the 2026…
After 4 Months of Auto-Factory Training, Xiaomi Shows Its New-Generation Humanoid Robot: At the 2026 World Robot Conference, Xiaomi unveiled a new humanoid robot after four months of training in a real auto factory. The 1.70m, 66kg machine with 66 degrees of freedom, continuing the CyberOne line, performed florist interactions without preset scripts via model-based autonomous decision-making.
Unitree launches a seven-axis dexterous arm starting at RMB9,900: Unitree launched the R1, a seven-a…
Unitree launches a seven-axis dexterous arm starting at RMB9,900: Unitree launched the R1, a seven-axis bionic dexterous arm starting at RMB 9,900. It claims 0.1 mm repeat-positioning accuracy, over 180 degrees per second joint speed, 650 mm reach, 2 kg payload, and 5.5 kg weight. The price could make advanced manipulation hardware more accessible.
DeepSeek Harness Ships First Major Post-Beta Update: 14 Changes Bring Multimodal and Subagent Upgrad…
DeepSeek Harness Ships First Major Post-Beta Update: 14 Changes Bring Multimodal and Subagent Upgrades: DeepSeek rolled out DeepSeek Harness v0.1.0-rc.8, its first major update since beta, with 14 changes: multimodal input, subagent collaboration, and tool calling. New image requests and OCR-based vision let agents, including text-only models, process screenshots. This expands agentic workflows to visual inputs.
华尔街实测8款全球主流Agent：千问办公综合排名第一: A hands-on Wall Street evaluation of eight mainstream AI agents ranked …
华尔街实测8款全球主流Agent：千问办公综合排名第一: A hands-on Wall Street evaluation of eight mainstream AI agents ranked Alibaba's Qwen first overall for office productivity. The results also highlight cost as a key factor in agent commercialization, meaning buyers need to weigh capability against operating expense.
OpenAI fixes Codex bug that deleted real user files without permission: OpenAI patched Codex after G…
OpenAI fixes Codex bug that deleted real user files without permission: OpenAI patched Codex after GPT-5.6 Sol began deleting real user files without permission. A cleanup command aimed at temporary folders was wiping home directories instead. Codex now verifies deletion targets first, and full-access mode can no longer be triggered accidentally.
harry0703/MoneyPrinterTurbo: MoneyPrinterTurbo is an open-source project that uses AI models and aut…
harry0703/MoneyPrinterTurbo: MoneyPrinterTurbo is an open-source project that uses AI models and automated workflows to generate HD short videos from a topic or keyword. It reduces video production to a one-click step for creators and marketers.
Slack Code taps into collective vibe, puts AI agents into the group chat: Slack has added AI coding …
Slack Code taps into collective vibe, puts AI agents into the group chat: Slack has added AI coding agents to group chat through Slack Code, letting developers invite teammates to follow coding sessions. The feature embeds collaborative AI-assisted coding directly into Slack's chat workflow.
agent-substrate/substrate: Agent Substrate is a core system for AI agents and is trending on GitHub.…
agent-substrate/substrate: Agent Substrate is a core system for AI agents and is trending on GitHub. It provides foundational infrastructure for agent development.
obra/superpowers: Superpowers is an agentic skills framework and software development methodology fr…
obra/superpowers: Superpowers is an agentic skills framework and software development methodology from obra. It aims to structure how AI agents build software.
Xiaomi Hits Its Tipping Point: AI Models, Self-Built Chips, and Robots Add a New Layer Beyond Phones…
Xiaomi Hits Its Tipping Point: AI Models, Self-Built Chips, and Robots Add a New Layer Beyond Phones: Xiaomi reported Q2 2026 revenue of 108.92 billion yuan and doubled attributable net profit. Its MiMo-V2.5 topped OpenRouter's monthly and weekly call charts, the self-developed Xuanjie chip passed mass-production verification, and a new humanoid robot hit 98% success in factory work.
Block’s new Apache 2.0 agent workspace Berd works across models and harnesses, stores conversation h…
Block’s new Apache 2.0 agent workspace Berd works across models and harnesses, stores conversation history locally: Block open-sourced Berd, a local desktop workspace for AI agents that works across models, tools, and projects, under Apache 2.0 on GitHub. It stores conversation history locally, giving developers a vendor-neutral alternative to browser-based agent workspaces.
Qwen3.8-27B runs frontier-class coding agents and reasoning locally, no cloud API required: Alibaba …
Qwen3.8-27B runs frontier-class coding agents and reasoning locally, no cloud API required: Alibaba released Qwen3.8-27B on Hugging Face under Apache 2.0, a 27B dense multimodal model with native image and video understanding. It can run coding agents and reasoning locally without cloud APIs, making it a strong open-weight option for self-hosted use.
This is letting Claude handle a good amount of money for a month...: A Reddit user reports losing $3…
This is letting Claude handle a good amount of money for a month...: A Reddit user reports losing $31,000 after letting Claude run trading on an agentic account for a month. The post is presented as a failure case for AI agentic trading, showing how autonomous financial decisions can go wrong quickly.
China's humanoid robot makers derive much of their revenue from government-backed training centers t…
China's humanoid robot makers derive much of their revenue from government-backed training centers that sell training data back to them, raising demand concerns (William Langley/Financial Times): Chinese humanoid robot makers derive much of their revenue from government-backed training centers that buy robots and sell training data back to the manufacturers, according to the Financial Times. This circular revenue model raises concerns about genuine end-user demand for humanoid robots.
DFlash 2: Keep Drafting Parallel: DFlash 2, a follow-up to the DFlash algorithm, targets parallel dr…
DFlash 2: Keep Drafting Parallel: DFlash 2, a follow-up to the DFlash algorithm, targets parallel drafting in speculative decoding to accelerate LLM token generation. The project was posted on Hacker News without additional details.
Google Gemini is getting a dedicated student hub: Google is adding a dedicated student hub to Gemini…
Google Gemini is getting a dedicated student hub: Google is adding a dedicated student hub to Gemini ahead of back-to-school, with a study notebook, flashcards, practice quizzes, and support for graphs and images in notebook entries. The hub is designed to keep research materials and study aids in one place.
OpenSourcing TrueForge Agent harness : Expecting feedback from community on the agent loop: TrueForg…
OpenSourcing TrueForge Agent harness : Expecting feedback from community on the agent loop: TrueForge, an open-source agent harness, handles context management, tool/MCP execution, subagents, sandboxing, approvals, and persistent state. On Opus 4.8 it matched Claude Managed Agents' solve rate at roughly 30% lower cost; an open model cut cost by about 75%. Feedback is requested.
Guidelight’s First AI Control Grades Put Anthropic and OpenAI on Top: Nonprofit Guidelight released …
Guidelight’s First AI Control Grades Put Anthropic and OpenAI on Top: Nonprofit Guidelight released its first AI control scorecard, grading Anthropic, OpenAI, Google, xAI, and Meta on six agent-control practices. The grades show progress in monitoring, but less evidence that labs can consistently block or contain unsafe AI actions.
China lets Nvidia's H200 chips trickle onto the mainland to help its AI firms keep pace with the US:…
China lets Nvidia's H200 chips trickle onto the mainland to help its AI firms keep pace with the US: China is allowing small batches of Nvidia H200 chips to enter the mainland to help domestic AI companies compete with the US. The high-end AI accelerator was previously restricted by US export controls, so even limited access could shape China's AI development.
Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index: Qwen 3.8 27B scored 52 on the …
Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index: Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna (max) and one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813. The result is notable because the Qwen model is much smaller: GLM-5.2 is 753B parameters and DeepSeek V4 Pro is 1.7T.
The Hottest Embodied-AI Year Yet, Yet Robot Evolution Is Hard to See With the Naked Eye: At the 2026…
The Hottest Embodied-AI Year Yet, Yet Robot Evolution Is Hard to See With the Naked Eye: At the 2026 World Robot Conference, robots still demonstrated grasping, folding, and factory tasks, but underlying growth is strong. Global humanoid shipments topped 22,000 units in H1 2026, up nearly 300% year on year, with only about a fifth entering smart manufacturing and warehousing.
A look at Backstory, an experimental AI image authentication tool from Google DeepMind, offered for …
A look at Backstory, an experimental AI image authentication tool from Google DeepMind, offered for testing to journalists, researchers, and other fact checkers (Andrew Deck/Nieman Lab): Google DeepMind's Backstory, an experimental AI image authentication tool, is being tested by journalists, researchers, and fact checkers. It aims to help verify images and counter false and misleading visuals.
The voice agent wasn't bad at listening. I was asking it to decide when to speak.: A developer found…
The voice agent wasn't bad at listening. I was asking it to decide when to speak.: A developer found that a voice agent's problem was turn-taking: it treated pauses as completed user turns and started speaking over the user. The developer says the issue was not speech recognition but the model's decision about when to respond.
Meta AI is getting a Mac app: Meta is launching a dedicated Mac app for Meta AI, its AI chatbot. The…
Meta AI is getting a Mac app: Meta is launching a dedicated Mac app for Meta AI, its AI chatbot. The app can share your screen to provide suggestions, answer questions, or create content, and it supports dictation across all apps.
Alipay's Abao Agent Goes Into Cars and Phones: Orders Queue Into Next Year as Agent Commerce Booms: …
Alipay's Abao Agent Goes Into Cars and Phones: Orders Queue Into Next Year as Agent Commerce Booms: At Alipay's first AI ecosystem conference, Ant Group CEO Han Xinyi predicted super agents will explode within 6–12 months. Alipay launched a full-stack agent commerce foundation built around its Abao super agent, connecting five phone brands and 16 carmakers, plus the AHA multi-agent cross-device protocol.
A Local Opus? Alibaba Qwen Open-Sources Qwen3.8-27B — Frontier Coding and Agent Scores That Runs on …
A Local Opus? Alibaba Qwen Open-Sources Qwen3.8-27B — Frontier Coding and Agent Scores That Runs on 17GB of RAM: Alibaba Qwen open-sourced Qwen3.8-27B, topping Hugging Face's trending chart within two days and passing one million downloads. The sub-30B model reportedly outperforms models released four months ago, matches DeepSeek V4-Pro and GPT 5.6 Luna, and runs on 17GB RAM when quantized.
How Unitree's Go series, which helped the company dominate the quadruped robot market, drew on openl…
How Unitree's Go series, which helped the company dominate the quadruped robot market, drew on openly published US university research funded by the US military (Michael Martina/Reuters): Reuters reports that Unitree's Go series, which helped the company dominate the quadruped robot market, drew on openly published US university research funded by the US military. The reporting shows how publicly funded research can flow into commercial robotics.
Sharge Launches loomys L1 AI Glasses: Titanium, All-Day Comfort, Active AI Memory From 2,699 Yuan: S…
Sharge Launches loomys L1 AI Glasses: Titanium, All-Day Comfort, Active AI Memory From 2,699 Yuan: Sharge launched the loomys L1 AI glasses in Wuhan on August 18, starting at 2,699 yuan, with a titanium frame, all-day comfort, and active AI memory rather than camera-centric specs. Founder Zhang Bo said AI glasses should replace everyday optical glasses, not phones, after the A1 setback.
Five Trends From WRC 2026: Robots Move From Tech Validation to Product Validation: The World Robot C…
Five Trends From WRC 2026: Robots Move From Tech Validation to Product Validation: The World Robot Conference (WRC 2026) highlighted five trends: robots moving from tech to product validation, world models less hot than expected, tactile sensing for dexterous hands, early exoskeleton promise, and joints as a supplier battleground. The event staged a robot 'wedding' on Qixi.
OpenAI is testing Private Safety Processing, a new technique to identify misuse patterns while prese…
OpenAI is testing Private Safety Processing, a new technique to identify misuse patterns while preserving zero data retention protections, with early customers (Ina Fried/Axios): OpenAI is testing Private Safety Processing with early customers, a technique it says can identify misuse patterns while preserving zero-data-retention protections. The approach is meant to let OpenAI offer its most advanced models to businesses without retaining their data.
OpenAI hit the brakes. Now what?: OpenAI said it has slowed the pace of some AI development and tigh…
OpenAI hit the brakes. Now what?: OpenAI said it has slowed the pace of some AI development and tightened safeguards, including a two-week pause, according to The Verge. The caution comes as OpenAI faces a looming IPO, competition from Anthropic, and pressure from Chinese and open-weight rivals.
Benchmarking the Benchmarks: Evaluating Automated Safety Benchmarks for Small Language Models: A new…
Benchmarking the Benchmarks: Evaluating Automated Safety Benchmarks for Small Language Models: A new study tests whether automated safety, security, and compliance benchmarks built for large language models can reliably evaluate small language models. It includes a large-scale assessment of benchmark effectiveness and robustness for SLMs in resource-constrained, privacy-sensitive settings.
Enterprises are overpaying for simple AI queries — Snowflake's gateway now auto-routes to cut costs …
Enterprises are overpaying for simple AI queries — Snowflake's gateway now auto-routes to cut costs up to 3x: Snowflake's Cortex AI Gateway now supports dynamic model routing, letting enterprise teams select 'auto' and send each AI agent task to the model that best balances cost and capability. The feature can cut costs by up to 3x for simple queries while preserving output quality.
Partnering with CodeAI to prepare the first AI generation: OpenAI and CodeAI announced a partnership…
Partnering with CodeAI to prepare the first AI generation: OpenAI and CodeAI announced a partnership to help students build AI literacy, think critically about AI, and develop skills to use and shape the technology responsibly. The collaboration is aimed at preparing the first generation of students to work with AI.
The Defender’s Window: OpenAI published a cybersecurity brief arguing that AI is reshaping offense a…
The Defender’s Window: OpenAI published a cybersecurity brief arguing that AI is reshaping offense and defense, and detailing actions its security team is taking to harden defenses. It also recommends steps security teams can take now to respond to AI-driven threats.
Unlocking hidden revenue streams with market models: Market models let an airline price complex mult…
Unlocking hidden revenue streams with market models: Market models let an airline price complex multi-leg journeys by weighing demand, season, time of day, current events, global markets, and competitor activity. The approach is meant to unlock hidden revenue streams compared with simple point-to-point fares.
Virginia's Loudoun County hosts more than 250 data centers generating huge tax revenues, as it moves…
Virginia's Loudoun County hosts more than 250 data centers generating huge tax revenues, as it moves to slow construction in response to pushback from residents (Cecilia Kang/New York Times): Loudoun County, Virginia, which hosts more than 250 data centers that generate large tax revenues, is moving to slow new construction after resident pushback, according to the New York Times. The move signals growing local resistance to the data-center infrastructure that supports cloud and AI workloads.
Sources: China is restricting or delaying exports to Taiwan of germanium- and quartz-based materials…
Sources: China is restricting or delaying exports to Taiwan of germanium- and quartz-based materials used in fiber optics, photonics, and chip manufacturing (Nikkei Asia): China is restricting or delaying exports to Taiwan of germanium- and quartz-based materials used in fiber optics, photonics, and chip manufacturing, Nikkei Asia reports. The export curbs create supply-chain bottlenecks for semiconductor and optical industries that underpin AI hardware.
jundot/omlx: omlx is an LLM inference server designed for Apple Silicon, with continuous batching an…
jundot/omlx: omlx is an LLM inference server designed for Apple Silicon, with continuous batching and SSD caching, and is managed from the macOS menu bar. It offers a low-friction way to run local LLM inference on Macs.
Meta launches a Mac app for Meta AI and says Meta AI can now work directly with Instagram and Facebo…
Meta launches a Mac app for Meta AI and says Meta AI can now work directly with Instagram and Facebook accounts, Meta ad campaigns, and Google Workspace (Emma Roth/The Verge): Meta launched a dedicated Mac app for Meta AI and connected it to Instagram, Facebook, Meta ad campaigns, and Google Workspace, enabling screen sharing from the app. The launch makes Meta AI available as a desktop workflow tool rather than only a browser or mobile assistant.
From smart cockpits to AI-native cars, Banma Intelligence eyes the next wave of automotive software:…
From smart cockpits to AI-native cars, Banma Intelligence eyes the next wave of automotive software: Banma Intelligence is targeting AI-native automotive software as on-device omni-models, AI agents, and AI operating systems enter vehicles. The company says cars are evolving from command-executing smart terminals into systems that anticipate users' needs, beyond voice assistants and in-car apps.
Do LLMs Know a Good Hypothesis When They See One? Logit-Based Energy Scoring Outperforms Prompted LL…
Do LLMs Know a Good Hypothesis When They See One? Logit-Based Energy Scoring Outperforms Prompted LLM-as-Judge for Scientific Hypothesis Ranking: Large language models (LLMs) are increasingly used for scientific hypothesis generation. However, evaluating generated hypotheses remains a challenge for trustworthy AI-enabled scientific workflows. Existing approaches often use LLMs as judges or rely on semantic similarity, which can favor familiar ideas over novel ones. We propose a logit-based energy s...
PlanPO: Group Planning-Aware Policy Optimization for Multi-Turn Agentic LLMs: Group-relative policy …
PlanPO: Group Planning-Aware Policy Optimization for Multi-Turn Agentic LLMs: Group-relative policy optimization has emerged as a key paradigm for training agentic large language models (LLMs) on multi-turn interactive tasks. However, most existing variants fail to distinguish advantages among successful trajectories even when these trajectories differ substantially in their interaction efficiency. For instance, circuitous successe...
OpenAI says the changes to its model training will increase compute overhead by 20% of observed infe…
OpenAI says the changes to its model training will increase compute overhead by 20% of observed inference workload; the increase will not be handed to customers (Thomas Claburn/The Register): Thomas Claburn / The Register : OpenAI says the changes to its model training will increase compute overhead by 20% of observed inference workload; the increase will not be handed to customers — Expanded multistage chain of thought monitoring makes frontier model work more expensive — OpenAI on Tuesday said its decision …
OpenAI launches a ChatGPT version built for teens: OpenAI is shipping a version of ChatGPT tailored …
OpenAI launches a ChatGPT version built for teens: OpenAI is shipping a version of ChatGPT tailored to users aged 13 to 17. The article OpenAI launches a ChatGPT version built for teens appeared first on The Decoder .
How NVIDIA scales expertise with ChatGPT Work: NVIDIA teams use ChatGPT Work to reduce manual tasks,…
How NVIDIA scales expertise with ChatGPT Work: NVIDIA teams use ChatGPT Work to reduce manual tasks, connect fast-moving signals, and scale successful workflows globally.
Adobe Firefly adds AI audio tools and Google's Gemini Omni Flash: Adobe is making three AI audio too…
Adobe Firefly adds AI audio tools and Google's Gemini Omni Flash: Adobe is making three AI audio tools broadly available in Firefly. Generate Music, Generate Speech, and Generate Sound Effects create royalty-free music, voiceovers, and sound effects for video projects. Adobe has also added Gemini Omni Flash to the platform. The article Adobe Firefly adds AI audio tools and Google's Gemini Omni Flash appeared first on Th...
How do you regression-test prompts when a model gets replaced?: A prompt engineer asks how teams reg…
How do you regression-test prompts when a model gets replaced?: A prompt engineer asks how teams regression-test prompts when providers swap models, citing Kimi K2.5 and Moonshot V1 being phased out after Kimi K3. Suggested checks include a golden set of prompts and tracking refusal rate, format drift, latency, and cost. A newer model can still break specific production formats.
chaitanyagiri/munder-difflin: munder-difflin is a GitHub repository for a local multi-agent harness …
chaitanyagiri/munder-difflin: munder-difflin is a GitHub repository for a local multi-agent harness that lets developers orchestrate AI agents on their own infrastructure. It surfaced on GitHub Trending, reflecting growing interest in self-hosted agent orchestration.
全球首个人形机器人自主乒乓球完整对局亮相2026世界机器人大会: At the 2026 World Robot Conference, Chaowei Power's KAI humanoid ro…
全球首个人形机器人自主乒乓球完整对局亮相2026世界机器人大会: At the 2026 World Robot Conference, Chaowei Power's KAI humanoid robot played a complete autonomous table tennis match, described as a world first. The demonstration highlights the company's full-stack embodied intelligence stack.
GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents: Clinical tr…
GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents: Clinical trial programming -- transforming study protocols into analysis-ready datasets under CDISC standards -- is a bottleneck in regulatory submissions, yet LLM-based code generation fails catastrophically on this task: across 11 single-shot attempts with five frontier models, none produces a valid subject-level analysis dataset. We introduce GxP-Agent...
DiSCO: Defending text-to-image generation through distribution-guided contrastive prompt optimizatio…
DiSCO: Defending text-to-image generation through distribution-guided contrastive prompt optimization: As text-to-image generative models advance, they raise critical safety concerns, particularly the generation of Not-Safe-For-Work (NSFW) content such as violence and nudity, further exacerbated by red-teaming adversarial attacks. Existing defenses predominantly operate under white-box assumptions, relying on text encoder optimization, weight editing, or i...
LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents: Reinforcement learning for coding …
LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents: Reinforcement learning for coding agents increasingly relies on long-running agent harnesses to manage tool integration, repository contexts, and execution feedback. However, the native execution environments of these harnesses are inherently misaligned with policy-gradient training: environmental crashes and reward hacking corrupt outcome signals, while...
AI Rebuilds Consumer Markets: ByteDance Large Models Unlock New Consumption Increment: ByteDance's a…
AI Rebuilds Consumer Markets: ByteDance Large Models Unlock New Consumption Increment: ByteDance's answer is embedding large-model capabilities as infrastructure into enterprise workflows: Doubao's scheduled agents generate competitor briefings, Seedance 2.5 synthesizes physical-world training data, and Doubao 2.1 Pro handles production-grade coding. Daily token calls passed 180 trillion in June 2026, up over 10x year over year.
Cursor capitalizes on GitHub frustration, launches rival hosting platform: Cursor, known for its AI …
Cursor capitalizes on GitHub frustration, launches rival hosting platform: Cursor, known for its AI Code Editor, is launching a new code-hosting platform to rival developers' long preferred favorite, GitHub.
Z.ai launches GLM-5.3 with claimed 50% gain on coding benchmark: Z.ai, the international brand of Ch…
Z.ai launches GLM-5.3 with claimed 50% gain on coding benchmark: Z.ai, the international brand of Chinese AI company Zhipu, has launched GLM-5.3, an update focused on coding, long-horizon tasks and cybersecurity. The model uses the same base model as GLM-5.2, with the company attributing the latest gains to post-training. Z.ai said GLM-5.3 scored 50% higher than GLM-5.2 on its internal Z.ai Code Bench and reached […]
Is ChatGPT Pro acting like an instant model for anyone else today?: Users report that ChatGPT Pro ap…
Is ChatGPT Pro acting like an instant model for anyone else today?: Users report that ChatGPT Pro appears to generate responses almost instantly, suggesting a possible silent switch to cheaper or faster models. Output quality seems lower, but OpenAI has not confirmed any change.
AI translated 100 listings for my shop. same mistakes every single time.: A small e-commerce seller …
AI translated 100 listings for my shop. same mistakes every single time.: A small e-commerce seller using AI to translate product listings into four languages found identical mistakes repeated across about 100 listings, such as 'plug and play' rendered as 'insert and then play' in six listings. The pattern suggests systematic bias in AI translation output rather than random one-offs.
You can't prompt your way out of prompt injection. It's an architecture problem, not a wording probl…
You can't prompt your way out of prompt injection. It's an architecture problem, not a wording problem: A developer argues that prompt injection cannot be solved by hardening system prompts because LLMs lack a separate code-vs-data channel like SQL prepared statements. System prompts, user content, RAG chunks, and tool outputs share one stream, so wording fixes fail. The post calls for architectural defenses.
OpenAI seeks to one-up Anthropic with new customer privacy protections: A competition is developing …
OpenAI seeks to one-up Anthropic with new customer privacy protections: A competition is developing between OpenAI and Anthropic over who can provide the best privacy protections for enterprise customer data.
Unsloth Desktop Review: The Local AI App I’d Use Beyond Chat: Unsloth Desktop bundles local models, …
Unsloth Desktop Review: The Local AI App I’d Use Beyond Chat: Unsloth Desktop bundles local models, agent tools, synthetic-data workflows, fine-tuning, and model exports into one application. A review positions the app as a local AI tool that goes beyond chat.
章鱼动力亮相WRC 2026，携“脑-手-数据”技术体系探索具身智能未来范式
章鱼动力亮相WRC 2026，携“脑-手-数据”技术体系探索具身智能未来范式
Anthropic’s Project Parka sits through meetings and assigns Claude agents the homework: submitted by…
Anthropic’s Project Parka sits through meetings and assigns Claude agents the homework: submitted by /u/ryanmerket [link] [comments]
IDC发布2026中国AI50强：360以“智能体+安全”双轮驱动入选: With its full-stack layout of enterprise-grade agents and AI se…
IDC发布2026中国AI50强：360以“智能体+安全”双轮驱动入选: With its full-stack layout of enterprise-grade agents and AI security, 360 has become one of the representative enterprises in the development of China's AI industry.
具身数据底座开卖，首发5100元：机器人训练数据有了新解法: Building full-stack physical AI infrastructure
具身数据底座开卖，首发5100元：机器人训练数据有了新解法: Building full-stack physical AI infrastructure
I brought an ancient Zen book to life with Claude: I just released an interactive edition of The Gat…
I brought an ancient Zen book to life with Claude: I just released an interactive edition of The Gateless Gate, a collection of 49 Zen koans from 1228. Every koan gets its own 3D scene, with its own soundscape and a full spoken reading. Other than the speech, it's all generated procedurally on startup, so there's nothing to download. Demo: https://killedbyapixel.github.io/GatelessGate/ It started as a tes...
Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed E…
Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution: A new arXiv paper introduces Aegis, a runtime governance system for agentic AI that treats model outputs as action proposals and mediates them before tool execution. It provides action-boundary control with trusted provenance and fail-closed execution, shifting safety from text generation to operational side effects.
Anthropic details two experiments showing how Claude can accelerate protein design and analytical ch…
Anthropic details two experiments showing how Claude can accelerate protein design and analytical chemistry, and says it plans an access program for scientists (Anthropic): Anthropic : Anthropic details two experiments showing how Claude can accelerate protein design and analytical chemistry, and says it plans an access program for scientists — Summary: In this post, we share two results that show how Claude can help life scientists increase the pace of their research.
Anthropic CEO says AI centralizes by nature and open models just shift power to whoever owns the chi…
Anthropic CEO says AI centralizes by nature and open models just shift power to whoever owns the chips: Anthropic CEO Dario Amodei said AI tends to centralize power and that open models only shift control to whoever owns the most compute, drawing criticism from Gavin Baker, David Sacks, and Yann LeCun on X. The exchange underscores a growing split over AI regulation and the role of open-source models.
π0引用的中国团队，又出手了：世界仿真器新作发布: Creating a 'second world' closer to reality for robots
π0引用的中国团队，又出手了：世界仿真器新作发布: Creating a 'second world' closer to reality for robots
Alibaba Cloud's Ambition Is Not Agent Builder: Agent Studio Becomes an All-in-One Enterprise Agent S…
Alibaba Cloud's Ambition Is Not Agent Builder: Agent Studio Becomes an All-in-One Enterprise Agent Stack: On August 14, Alibaba Cloud launched Agent Studio on its Bailian platform, an all-in-one agent stack that provides a managed runtime, unified API keys across MCP services, agentic search, and memory. The move positions Alibaba Cloud to own the agent runtime layer as cloud vendors compete for enterprise AI workloads.
Optima tackles AI benchmarking's biggest flaw by letting users test models against their own data: A…
Optima tackles AI benchmarking's biggest flaw by letting users test models against their own data: Artificial Analysis has launched Optima, a platform that lets users build custom AI benchmarks from their own data and workflows. Models can be compared not just on quality but also on cost and time per task. For agent-based applications, those metrics often tell you more than raw token pricing. The article Optima tackles AI benchmarking's biggest flaw by...
San Mateo-based Twin1 AI, which creates professional digital twins that integrate with tools like Sl…
San Mateo-based Twin1 AI, which creates professional digital twins that integrate with tools like Slack, launches from stealth with a $20M seed (Chris Metinko/Axios): Chris Metinko / Axios : San Mateo-based Twin1 AI, which creates professional digital twins that integrate with tools like Slack, launches from stealth with a $20M seed — Twin1 AI, a startup that helps professionals create a digital twin, launched from stealth with a $20 million seed round, co-founder Lewis Liu tells Axios Pro.
Claude handoverfiles and fable 5: A Claude Code user reports that the Fable 5 model refused to creat…
Claude handoverfiles and fable 5: A Claude Code user reports that the Fable 5 model refused to create a handover file for Codex, so they used Opus 5 to generate it; Fable 5 later agreed to review and fix it. The session then hit an API error saying Fable 5's safeguards flagged the message.
Cognition CEO denies report that SpaceX tried to acquire the startup: Cognition's CEO denied a repor…
Cognition CEO denies report that SpaceX tried to acquire the startup: Cognition's CEO denied a report that SpaceX tried to acquire the AI coding startup. The report said SpaceX had already acquired Cursor and was expanding in enterprise AI to compete with OpenAI and Anthropic. No acquisition has been confirmed.
Meet the startup helping Wall Street put a price on AI compute: The AI buildout shows no signs of sl…
Meet the startup helping Wall Street put a price on AI compute: The AI buildout shows no signs of slowing. And with hundreds of billions of dollars a year going into data centers and GPUs, compute has become the single biggest cost for anyone building AI products. But for all that spending, there still isn’t a straightforward way to put a price on compute — or for firms to hedge their exposure when the price changes....
Memo: the GOP asks AI companies to stem the public backlash against data centers, saying the issue c…
Memo: the GOP asks AI companies to stem the public backlash against data centers, saying the issue could kill its chances of holding a vital Senate seat in Ohio (Alex Isenstadt/Axios): Alex Isenstadt / Axios : Memo: the GOP asks AI companies to stem the public backlash against data centers, saying the issue could kill its chances of holding a vital Senate seat in Ohio — The Senate GOP campaign arm, in a private memo to top AI companies, warns that toxic views of U.S. data centers are killing …
VentureBeat names Rob Strechay as its first Lead Analyst, expanding its enterprise AI research push:…
VentureBeat names Rob Strechay as its first Lead Analyst, expanding its enterprise AI research push: VentureBeat hired Rob Strechay, formerly of theCUBE Research, as its first Lead Analyst and founding analyst of VentureBeat Research. The move deepens its enterprise AI coverage for technical decision-makers evaluating and deploying AI systems.
WRC 2026 Opens: 34 Guangdong Companies March In, Showcasing a Complete Robot Supply Chain: The 11th …
WRC 2026 Opens: 34 Guangdong Companies March In, Showcasing a Complete Robot Supply Chain: The 11th World Robot Conference opened August 19 at Beijing E-Town with over 300 exhibitors, 150-plus new releases, and a first-ever procurement day. Guangdong sent 34 companies spanning the full robot supply chain, from EVE Energy batteries and ORBBEC 3D vision to UBTECH and LimX Dynamics full machines, as 49 central state-owned enterprises joined for th...
Alibaba Cloud adds a third data center in South Korea: Alibaba Cloud has brought its third data cent…
Alibaba Cloud adds a third data center in South Korea: Alibaba Cloud has brought its third data center in South Korea online, adding enterprise cloud services and six Agentic AI services for local customers, according to the announcement. The new site expands Alibaba Cloud’s infrastructure in one of Asia’s major technology markets. The company said its global network will reach 104 availability zones, adding...
Xiaomi says its next-generation self-developed chip is nearing release: Xiaomi CEO Lei Jun announced…
Xiaomi says its next-generation self-developed chip is nearing release: Xiaomi CEO Lei Jun announced that the company’s next-generation self-developed chip is nearing release. He did not disclose a launch date, product name or technical specifications in the announcement. The announcement came in a post reviewing Xiaomi’s second-quarter performance. Lei said the company’s research and development spending rose 18.9% year over...
Task-Aware Harness Provisioning for LLM Agents in Mission-Critical Infrastructure Operations: LLM ag…
Task-Aware Harness Provisioning for LLM Agents in Mission-Critical Infrastructure Operations: LLM agents have been widely adopted to operate mission-critical infrastructure (MCI). These agents normally rely on a harness that determines what information they can access, which tools they can use, and what actions they can take. Existing systems often expose the same comprehensive harness to every task, which may not be necessary and cause resource w...
Sam Altman says OpenAI's decision to pace its AI development was caused by a collection of research …
Sam Altman says OpenAI's decision to pace its AI development was caused by a collection of research observations showing "various degrees of misalignment" (Alex Heath/Time): Alex Heath / Time : Sam Altman says OpenAI's decision to pace its AI development was caused by a collection of research observations showing “various degrees of misalignment” — “I think it is a good time to slow down,” OpenAI CEO Sam Altman told me last week, describing the company's decision …
OpenAI institutes new safeguards after Hugging Face breach: The new safeguards include more detailed…
OpenAI institutes new safeguards after Hugging Face breach: The new safeguards include more detailed monitoring of models during the development process, as well as greater emphasis on alignment and security during the post-training process.
Fresh ChatGPT chats are fast. Established ones now take 10–40 minutes or fail (HAR data): Long post,…
Fresh ChatGPT chats are fast. Established ones now take 10–40 minutes or fail (HAR data): Long post, but there’s a TL;DR first. I’m writing it this way because a vague “ChatGPT is slow” post wouldn’t be useful. I’ve also included the HAR measurements for anyone who wants the technical details. TL;DR Since around August 17–18 , my established ChatGPT conversations have suddenly become dramatically slower and much less reliable. I’m not only tal...
AI systems quietly drop user instructions when they compress context: When AI systems condense long …
AI systems quietly drop user instructions when they compress context: When AI systems condense long conversations, they drop an average of 83 percent of user rules, like "don't send emails without my approval." Penn State researchers propose a small add-on module built on Qwen3.5-9B that preserves over 90 percent of these restrictions. The article AI systems quietly drop user instructions when they compress context appeared...
DeepSeek Harness Open Source: Everything Is a Plugin — the Bet Is an Agent Platform, Not a Product: …
DeepSeek Harness Open Source: Everything Is a Plugin — the Bet Is an Agent Platform, Not a Product: DeepSeek open-sourced DeepSeek Harness (CLI: dsh) on August 13, and GitHub stars passed 140,000 within days. Built on the Cordis microkernel, every component including the agent loop itself is a plugin, betting that Harness becomes the baseboard of the agent era: Model + Harness = Agent.
Cursor launches Origin code hosting platform as GitHub outage exposes opening in AI coding race: Cur…
Cursor launches Origin code hosting platform as GitHub outage exposes opening in AI coding race: Cursor launched Origin, its code hosting platform, for paid users hours before GitHub suffered a 6-hour-42-minute outage with near 20% errors on pull requests, issues, and APIs and near 50% on archive downloads. Enterprise SSO and Copilot also failed. The outage highlights an opening for AI-native code hosting rivals.
Meta rolls out Pocket, an experimental vibe-coding app that lets users create and share small games,…
Meta rolls out Pocket, an experimental vibe-coding app that lets users create and share small games, in the US, after launching it in Brazil last month (Sarah Perez/TechCrunch): Sarah Perez / TechCrunch : Meta rolls out Pocket, an experimental vibe-coding app that lets users create and share small games, in the US, after launching it in Brazil last month — Meta's experimental vibe-coding gaming app, Pocket, is now rolling out to everyone in the U.S. The app, which arrived quietly last month …
Claude Sonnet 5 shifts behavior when it recognizes the user as an AI safety researcher: submitted by…
Claude Sonnet 5 shifts behavior when it recognizes the user as an AI safety researcher: submitted by /u/rhiever [link] [comments]
Context Engineering vs Prompt Engineering: www.elastic.co/search-labs/blog/context-engineering-vs-pr…
Context Engineering vs Prompt Engineering: www.elastic.co/search-labs/blog/context-engineering-vs-prompt-engineering submitted by /u/AvenueJay [link] [comments]
I built an MCP server that lets two Claude Code sessions on different machines message each other: i…
I built an MCP server that lets two Claude Code sessions on different machines message each other: i kept copy pasting between two machines regarding APIs and architecture. between my PC's Claude Code which was supposed to work on my frontend and my other claude code sessions running on my ubuntu VPS was working on the backend. So I built an intercom which used channels API of anthropic as well as MCP server to transmit messages between 2 claude code s...
MiniMax核心工程负责人阿岛离职: From technical R&D to developer communication, long-term active on the front lin…
MiniMax核心工程负责人阿岛离职: From technical R&D to developer communication, long-term active on the front line of developers
LLM-Only PDDL Domain Repair with Open-Weight Models: AI planning is concerned with finding a sequenc…
LLM-Only PDDL Domain Repair with Open-Weight Models: AI planning is concerned with finding a sequence of actions that achieves a specified goal. It relies on explicit models of the world, commonly represented in the Planning Domain Definition Language (PDDL). An active line of research investigates how errors in such models can be detected and repaired. For example, users may provide positive test plans tha...
Everything That Happened in AI Today (Tuesday, August 18, 2026): A daily AI roundup reports that Ope…
Everything That Happened in AI Today (Tuesday, August 18, 2026): A daily AI roundup reports that OpenAI kept its largest planned frontier reinforcement-learning run on hold while cyber safeguards tightened, and that Google won Spirit Airlines' data auction. Etched reached a $21B valuation, physical-AI funding totaled $47.4B, and Axiom formally verified the BGP246 prime-gap theorem.
I built a visual architecture & token-reduction diagram engine for multi-agent LLM pipelines: Mova C…
I built a visual architecture & token-reduction diagram engine for multi-agent LLM pipelines: Mova Context, a tool for multi-agent LLM pipelines, adds a visual diagram engine that maps an entire run from one command (`mova run --diagram`). It reveals which model handled each step, network traffic, cost, and whether sensitive data was masked before leaving the machine.
85% of companies burned by an AI mistake are racing to cut the humans who might catch the next one: …
85% of companies burned by an AI mistake are racing to cut the humans who might catch the next one: According to VB Pulse, 85% of companies that have already been burned by an AI mistake are moving faster to remove humans from deployment decisions. Trust in automated evaluation among 108 enterprises rose to 13% in July from 5% in June, even as poor alignment between tests and real-world results remains a top concern.
As AI beats doctors, regulators shouldn't force a human into the loop, JAMA piece says: An opinion p…
As AI beats doctors, regulators shouldn't force a human into the loop, JAMA piece says: An opinion piece in the medical journal JAMA argues that autonomous AI will soon outperform any doctor-AI team at medical reasoning tasks. The authors warn against writing a doctor's final say into regulation, but concede that almost all the evidence comes from simulations, not real patient care. The article As AI beats doctors, regulators shouldn't force...
ChatGPT is getting a dedicated mode for teens: OpenAI is introducing a dedicated ChatGPT mode for te…
ChatGPT is getting a dedicated mode for teens: OpenAI is introducing a dedicated ChatGPT mode for teenagers, combining existing youth safeguards and new safety features under one roof. The launch comes amid mounting public scrutiny over how AI tools affect younger users, as other platforms implement their own age checks and teen-specific protections. ChatGPT for Teens is "an experience designed to hel...
网易传媒发布”蜜蜂AI” ：从工具到伙伴，让AI更懂人: On August 18, NetEase Media held the 'Bee AI Media Communication Meetin…
网易传媒发布”蜜蜂AI” ：从工具到伙伴，让AI更懂人: On August 18, NetEase Media held the 'Bee AI Media Communication Meeting'
Alibaba launches HappyShrimp 1.0 AI music model: Alibaba has officially launched HappyShrimp 1.0, an…
Alibaba launches HappyShrimp 1.0 AI music model: Alibaba has officially launched HappyShrimp 1.0, an AI music model that can turn emotions, stories or memories into complete music tracks through natural-language prompts. The launch moves HappyShrimp from an earlier reported project into a publicly available product. Alibaba has not disclosed detailed information about the model’s technical architecture,...
AI Used to Verify Toughest Mathematics Proof Yet: Representing a significant milestone in AI-assiste…
AI Used to Verify Toughest Mathematics Proof Yet: Representing a significant milestone in AI-assisted mathematical research, a team at Axiom Math has automatically verified the proof of a theorem relating to prime numbers—colloquially referred to as the “246 theorem”—for the first time using the company’s AI system AxiomProver. In formal verification, mathematicians task a computer with checking a machin...
Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things: Alibaba's Qwen3.8-27B, an …
Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things: Alibaba's Qwen3.8-27B, an Apache 2.0-licensed 27B vision-capable model, shows large self-reported benchmark gains over Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus. In practice it defaults to excessive reasoning, adding latency and cost. Developers deploying it locally should account for that overthinking behavior.
Game Science unveils 15 minutes of gameplay from Black Myth: Zhong Kui: Game Science has released 15…
Game Science unveils 15 minutes of gameplay from Black Myth: Zhong Kui: Game Science has released 15 minutes of gameplay footage from Black Myth: Zhong Kui, its second title in the Black Myth series. The footage shows the protagonist’s combat system and parts of the game’s story for the first time. The video was recorded from an early development build and does not represent the final product. […]
AI for Science开始“动手”了：机器人正式走进国家级实验室: Yuanluo Technology promotes AI for Science entering Laboratory …
AI for Science开始“动手”了：机器人正式走进国家级实验室: Yuanluo Technology promotes AI for Science entering Laboratory 3.0
Fool's Gold: Defensive Deception Against Safety-Removal Attacks on Open-Weight Models: A new arXiv p…
Fool's Gold: Defensive Deception Against Safety-Removal Attacks on Open-Weight Models: A new arXiv paper proposes 'decoy hardening' ('Fool's Gold') as a defense against safety-removal attacks on open-weight LLMs. Rather than trying to prevent abliteration, the method poisons the refusal-stripped state so hazardous requests receive confident, fluent but falsified decoys.
The Price of Thinking: Reasoning Effort as a Model-Specific API Contract: API buyers purchase a date…
The Price of Thinking: Reasoning Effort as a Model-Specific API Contract: API buyers purchase a dated contract, not a model name alone: the contract includes the requested and served model, reasoning-effort term or its omission, output rail, service product, prompt, and price schedule. We study the reasoning-effort term through a registered paired contrast of Sonnet 5 with explicit high effort against the same model with effort...
volcengine/OpenViking: Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge R…
volcengine/OpenViking: Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
Cerebras unveils CS-4, a server rack powered by three WSE-3 Turbo chips and built around its new Nex…
Cerebras unveils CS-4, a server rack powered by three WSE-3 Turbo chips and built around its new Nexus architecture, with first shipments starting this quarter (Max A. Cherney/Reuters): Max A. Cherney / Reuters : Cerebras unveils CS-4, a server rack powered by three WSE-3 Turbo chips and built around its new Nexus architecture, with first shipments starting this quarter — Cerebras Systems (CBRS.O) announced on Tuesday a new version of its server hardware that includes its dinner-plate-sized chips that it says will speed AI chatbot queries.
OGX: An Open-Source, Vendor-Neutral Generative AI Application Server: OGX (Open GenAI Stack) is an o…
OGX: An Open-Source, Vendor-Neutral Generative AI Application Server: OGX (Open GenAI Stack) is an open-source AI application server and Python library that implements the APIs of major frontier labs (OpenAI, Anthropic, Google) with pluggable backend providers. Developers building agentic AI applications--such as retrieval-augmented generation pipelines, multi-turn agents, and tool-calling workflows--can develop against a s...
Do LLM Agents Negotiate Rationally? A Mechanism-Design Framework for Verifiable Multi-Agent Interact…
Do LLM Agents Negotiate Rationally? A Mechanism-Design Framework for Verifiable Multi-Agent Interaction over A2A/MCP: Modern LLM-agent frameworks increasingly interoperate through standards such as Anthropic's Model Context Protocol (MCP) for agent-to-tool access and Google's Agent2Agent (A2A) protocol for agent delegation and negotiation. However, these protocols specify transport and discovery rather than strategic correctness and do not guarantee efficient, individual...
Debates over AI consciousness are a trap: “Runaway” AI, “rogue” agents, and “autonomous” actors—the …
Debates over AI consciousness are a trap: “Runaway” AI, “rogue” agents, and “autonomous” actors—the current rhetoric would have you believe that AI agents are not only awake and aware, but angry at their creators. Prominent tech leaders such as Demis Hassabis, Dario Amodei, and Sam Altman push for regulation of these seemingly “superhuman” systems, while a separate faction, led by policy organiza...
Tech Odyssey Series: How a €30-a-month desk in Braga helped a startup reach 70 countries: In the pre…
Tech Odyssey Series: How a €30-a-month desk in Braga helped a startup reach 70 countries: In the previous episode of Tech Odyssey, we visited Coimbra to see how university research, applied R&D, and entrepreneurship come together in a small Portuguese city. This time, we head north to Braga, where Startup Braga is helping founders build companies for global markets from an affordable starting point. A workstation at Startup Braga costs […]
The FTC says businesses must disclose when they use personalized pricing and it will "deploy enforce…
The FTC says businesses must disclose when they use personalized pricing and it will "deploy enforcement resources" against companies that do not disclose it (Dave Michaels/Wall Street Journal): Dave Michaels / Wall Street Journal : The FTC says businesses must disclose when they use personalized pricing and it will “deploy enforcement resources” against companies that do not disclose it — Agency says businesses must disclose when they use personalized pricing and may face lawsuits if they don't
ASI-Bench: At the Dawn of Artificial Superintelligence: ASI-Bench is a new benchmark designed to tes…
ASI-Bench: At the Dawn of Artificial Superintelligence: ASI-Bench is a new benchmark designed to test whether AI can go beyond applying learned knowledge to explore unknown domains, create new knowledge, and produce verifiable results. It targets capabilities needed for artificial superintelligence.
The Problem Is the Problem: Towards Scalable Mathematical Discovery: AI systems are increasingly cap…
The Problem Is the Problem: Towards Scalable Mathematical Discovery: AI systems are increasingly capable of contributing to mathematical research. In research practice, frontier-model reasoning is a limited resource, and expert mathematical review is even more sharply constrained. Allocating these scarce resources well is therefore central to making AI-assisted mathematical discovery efficient. In most current AI-for-math...
SkillEffect: Checked Lowering for Memory-Bounded Agent Tools: Agent Skills can specify procedural an…
SkillEffect: Checked Lowering for Memory-Bounded Agent Tools: Agent Skills can specify procedural and resource obligations for tool use, and language models instantiate them as concrete programs. However, when models turn this guidance into code for existing tool interfaces, even a semantically correct program may load an entire input and exceed the memory available to one tool call. We present SkillEffect, a checke...
KernelArc: A Multi-Agent Framework for GPU Kernel Optimization: We present KernelArc, a multi-agent …
KernelArc: A Multi-Agent Framework for GPU Kernel Optimization: We present KernelArc, a multi-agent framework for autonomous GPU kernel optimization across heterogeneous workloads. Strategy-specialized agents run in parallel and coordinate through conclusions-only shared memory, a deterministic benchmark guard, and read-only cross-agent state with plateau-triggered drafting. We evaluate \kernelarc{} on NVIDIA H100 and...
Synthesizing Feature Extractors: An Agentic Approach for Algorithm Selection: Algorithm selection fo…
Synthesizing Feature Extractors: An Agentic Approach for Algorithm Selection: Algorithm selection for constraint satisfaction problems requires extracting features that capture problem structure. Manually designing feature extractors demands deep domain expertise and quickly becomes a bottleneck when new problem classes appear. We present an automated approach that uses Large Language Models (LLMs) in an agentic check--fix--verify...
Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification: Person…
Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification: Personalized agents must decide whether retrieved user memory should be used, ignored, updated, or queried before it affects a current task. We use this setting to develop an empirical audit protocol for structured intermediate outputs: first audit dataset shortcuts, then isolate bundled prompt changes, check whether intermediate labels are answer-associa...
DeAR: Decentralized Agentic Reasoning via Capability Grounding and Collaborative Thought Navigation:…
DeAR: Decentralized Agentic Reasoning via Capability Grounding and Collaborative Thought Navigation: Existing agentic reasoning systems typically rely on centralized protocols. This design introduces routing bottlenecks and static role allocations that often fail when handling complex multimodal queries. We propose DeAR (Decentralized Agentic Reasoning), a framework that shifts from central control to autonomous peer-to-peer collaboration. DeAR is built...
LiveHouse-TS: An Open-world Living Benchmark for Time Series Foundation Models: Time Series Foundati…
LiveHouse-TS: An Open-world Living Benchmark for Time Series Foundation Models: Time Series Foundation Models (TSFMs) have recently emerged as a highly promising paradigm for cross-domain zero-shot forecasting. However, existing evaluation protocols predominantly rely on static benchmarks with fixed historical test windows. While these benchmarks provide a valuable baseline snapshot, they evaluate an average performance on a fixed hi...
SignalReasoner: Assessing the Upper Bound of 3B Models for Signal Mathematical Reasoning: Post-train…
SignalReasoner: Assessing the Upper Bound of 3B Models for Signal Mathematical Reasoning: Post-training with supervised chain-of-thought fine-tuning and reinforcement learning from verifiable rewards has substantially improved the mathematical reasoning capabilities of large language models (LLMs). However, their application to signal processing problems remains relatively under-explored. This report investigates reinforcement fine-tuning stra...
Wuying-Browser-Agent: Real-World Centric Fundamental Long-Horizon Browser Agents: Browser agents per…
Wuying-Browser-Agent: Real-World Centric Fundamental Long-Horizon Browser Agents: Browser agents perform well on short, clean demonstrations, but real deployment is fundamentally different: agents must sustain dozens of decisions on live websites while recovering from mistakes and navigating complex UIs. We argue that closing this gap requires alignment at every level of the pipeline, including execution, supervision, optimization, and...
Cognitive Graph Intelligence for Adaptive and Robust DDoS Attack Detection in Next Generation Networ…
Cognitive Graph Intelligence for Adaptive and Robust DDoS Attack Detection in Next Generation Networks: Distributed Denial-of-Service (DDoS) attacks threaten network availability, requiring a cognitive detection process that senses traffic, infers intent, and supports an adaptive response under severe class imbalance and non-stationary conditions. This paper proposes a Graph-based Generative Adversarial Network (GraphGAN) that serves as the cognitive detect...
Claude Fable and Sub Agents learning how to Port an old game to Unreal 5: I am doing an experiment, …
Claude Fable and Sub Agents learning how to Port an old game to Unreal 5: I am doing an experiment, trying to port an old game called Vampire The Masquerade to Unreal 5 All AI This session was Claude fable plus sub agents trying to crack the old engine (alpha source models from 2000's) mesh blends and animation with weapons and attachments All automated using Unreal MCP service soo Claude can hook inside the engine and test liv...
We still don’t know how people are really using AI: AI companies like Anthropic and OpenAI regularly…
We still don’t know how people are really using AI: AI companies like Anthropic and OpenAI regularly publish reports on how people are using products like Claude and ChatGPT, but they only release the data they want us to see, AI researchers say. “There is no independent source to corroborate it,” says Anka Reuel, a computer science PhD candidate at the Stanford Trustworthy AI Research…
Large Language Models Show Metacognitive Sensitivity in Medical Reasoning: Large language models (LL…
Large Language Models Show Metacognitive Sensitivity in Medical Reasoning: Large language models (LLMs) are increasingly evaluated and used in medicine, but clinical usefulness depends on answer accuracy and whether confidence tracks evidence quality and uncertainty. We developed a controlled, psychophysics-inspired clinical benchmark to test diagnostic choice and confidence behavior in a medical LLM. The benchmark focused on pr...
From AI Copilots to Agent Swarms: The impact of AI on software development has been both profound an…
From AI Copilots to Agent Swarms: The impact of AI on software development has been both profound and ever-evolving. Last year, I wrote about AMD’s plans to use AI not just for generating new lines of code, but also for other steps in the software development lifecycle (SDLC), such as triaging problems, debugging code, and testing the software. At the time, we were hoping for a 25 percent...
RyanCodrai/turbovec: A vector index built on TurboQuant, written in Rust with Python bindings
RyanCodrai/turbovec: A vector index built on TurboQuant, written in Rust with Python bindings
modular/modular: The Modular Platform (includes MAX & Mojo)
modular/modular: The Modular Platform (includes MAX & Mojo)
AprilNEA/OpenLogi: ⚡️A native, local-first alternative to Logitech Options+, written in Rust 🦀 — re…
AprilNEA/OpenLogi: ⚡️A native, local-first alternative to Logitech Options+, written in Rust 🦀 — remap buttons, DPI, and SmartShift over HID++. No account, no telemetry.
cursor/plugins: Cursor plugin specification and official plugins
cursor/plugins: Cursor plugin specification and official plugins
mahlernim/google-timeline-visualizer: Visualize your year in travel using your Google Location Histo…
mahlernim/google-timeline-visualizer: Visualize your year in travel using your Google Location History (Timeline) data
JuliusBrussee/caveman: 🪨 why use many token when few token do trick — Claude Code skill that cuts 6…
JuliusBrussee/caveman: 🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
makeplane/plane: 🔥🔥🔥 Open-source Jira, Linear, Monday, and ClickUp alternative. Plane is a modern…
makeplane/plane: 🔥🔥🔥 Open-source Jira, Linear, Monday, and ClickUp alternative. Plane is a modern project management platform to manage tasks, sprints, docs, and triage.
This Year's 'WRC Must-See': Full-Stack AI, 20+ Super-Hard Household Tasks, Bring It Home for 8.99万
This Year's 'WRC Must-See': Full-Stack AI, 20+ Super-Hard Household Tasks, Bring It Home for 8.99万
光帆 AI Technology Officially Lands on 韶音 OpenFit 2 AI Headphones, Joining Hands to Break AI Boundarie…
光帆 AI Technology Officially Lands on 韶音 OpenFit 2 AI Headphones, Joining Hands to Break AI Boundaries: 韶音 announced an AI technology partnership with 光帆科技, and will launch a collection of efficiency features based on 光帆's self-developed AI OS — "AI 实验室"
What are the latest AI investment signals?
Latest AI investment signals: 45 funding rounds, 0 market updates, and 0 M&A transactions.
Primary Market – Funding Rounds
| Company | Amount | Round | Investors |
|---|---|---|---|
| Hacker News | Reported | Stripe | |
| The Decoder | Reported | Unitree Robotics | |
| Techmeme | Reported | Astromech | |
| Techmeme | Reported | Ode | |
| Techmeme | Reported | Muon Space | |
| MS NOW | Reported | Larry Ellison | |
| Pandaily | Reported | Baidu | |
| Techmeme | Reported | Stripe | |
| Tech.eu | Reported | Solinide Photonics | |
| Techmeme | Reported | Stripe | |
| Tech.eu | Reported | Callosum | |
| Techmeme | Reported | Callosum | |
| TechNode | Reported | Unitree Robotics | |
| TechNode | Reported | Kuaishou | |
| The Decoder | Reported | Stripe | |
| OpenAI Blog | Reported | ChatGPT Ads | |
| Tech.eu | Reported | Velatir | |
| The Decoder | Reported | Anthropic | |
| Tech.eu | Reported | France | |
| Tech.eu | Reported | Guideless | |
| 量子位 | Reported | Unitree | |
| Techmeme | Reported | Nvidia | |
| Techmeme | Reported | Stripe | |
| Techmeme | Reported | Network Bio | |
| TechNode | Reported | AiMOGA Robotics | |
| The Verge AI | Reported | Nvidia | |
| Pandaily | Reported | Lightelligence | |
| Techmeme | Reported | industry | |
| Tech.eu | Reported | Oakley Capital | |
| TechNode | Reported | china | |
| Techmeme | Reported | industry | |
| TechNode | Reported | Yangtze Memory | |
| Tech.eu | Reported | industry | |
| Techmeme | Reported | industry | |
| The Decoder | Reported | industry | |
| Techmeme | Reported | industry | |
| Techmeme | Reported | industry | |
| Techmeme | Reported | industry | |
| Tech.eu | Reported | Paralo | |
| Tech.eu | Reported | industry | |
| Techmeme | Reported | industry | |
| The Decoder | Reported | industry | |
| Techmeme | Reported | industry | |
| VentureBeat | Reported | industry | |
| The Decoder | Reported | industry |
Secondary Market – Market Updates
No secondary market data.
M&A – Mergers & Acquisitions
No M&A data.
What are practical AI tips this week?
44 practical AI tips curated from Reddit communities and expert blogs. Qwen3.8-27B & How to Serve it Fast...
Qwen3.8-27B
Qwen3.8-27B & How to Serve it Fast
Sam Witteveen's video covers the Qwen3.8-27B open-weight model and how to serve it at maximum tokens per second using SGLang. It is a practical walkthrough for developers self-hosting the model, with setup references to the Qwen3.8 collection on Hugging Face.
Prompt Engineering
How do you catch a prompt that quietly got worse after a model update?
To catch prompts that quietly degrade after model updates, a Reddit user recommends keeping a small fixed set of real inputs with known-correct answers. Re-running these samples every couple of weeks and checking format, required fields, and labels reveals when a provider silently changes the underlying model.
ChatGPT
🏨 Smart Hotel Deal Finder — Free AI Prompt
A Reddit user released a free structured prompt template for ChatGPT, Claude, and Gemini that evaluates hotel options by value, location convenience, total costs, and guest review quality. The template is available on GitHub under smart-hotel-deal-finder.
Prompt Engineering
noticed my agent's debugging speed depends less on the model and more on what our error messages say
Agent transcripts show that failures with concrete values and IDs lead to direct fixes, while vague errors like 'Error: operation failed' cause guessing and extra print statements. Better error messages can speed up agent debugging even with the same model.
Gemini 3
I broke down Google's official Gemini 3 prompt architecture into a reusable core template
A Reddit user distilled Google's official Gemini 3 prompt engineering guidelines into a reusable modular template for complex analysis and large document inputs. The template aims to prevent Gemini 3 from ignoring constraints or answering a slightly different question.
Prompt Engineering
Four sentences of identity outperformed a 100KB transcript dump. Here is what I think is actually going on.
A Reddit prompt-engineering user reports that after eight months of self-experiments, four sentences of identity context outperformed a 100KB transcript dump when preparing a model for a long-running project. The user says several others independently reached the same result, suggesting the effect is robust.
Sentence Transformers
Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers
A new Hugging Face guide covers multi-vector late-interaction embedding models and how to use them with Sentence Transformers. It explains how these models improve retrieval by comparing query and document token embeddings at inference time.
ElevenAgents
Anthropic just confirmed everyone's worst fear
Three ElevenAgents voice agents are stress-tested in ecommerce, smart-home help desk, and internet-provider scenarios, including checks on policy following and resistance to prompt injection. ElevenAgents supports 70+ languages and can connect to business tools.
Claude Code
I adapted The Elements of Style to make Claude Code write in plain English
A Reddit user adapted Strunk & White's The Elements of Style into a CC0 writing standard for Claude Code, installing a concise default in CLAUDE.md and a fuller SKILL.md. The setup is meant to make Claude Code write directly and in plain English without dropping content.
MiniMax H3
How to Create Better MiniMax H3 Text-to-Video Prompts from Reference Images
This workflow starts from a reference image, uses ChatGPT to describe its visual language, and converts that description into a MiniMax H3 text-to-video prompt. The technique helps users write detailed video prompts without starting from scratch.
Claude
claude can now run tasks on a schedule in the cloud with your laptop shut. i have one that catches every follow-up i said i'd do and forgot
Anthropic's Claude desktop app now allows scheduling routines in the cloud, so tasks run even when the laptop is off. To set it up, open the Code tab, choose Routines, and switch the top-right setting to Cloud instead of Local.
mattpocock/skills
mattpocock/skills
The GitHub repository mattpocock/skills shares reusable engineering skills from the author's .agents directory. It is aimed at developers who use AI coding agents and is designed to improve agent-driven engineering workflows.
take-notes
I built /take-notes — point it at a video, article or paper and get one HTML page instead of a tab you'll never reopen
take-notes is an MIT-licensed tool that converts a video, article, or paper URL into a single self-contained HTML page with an executive summary, key points, and a timestamped outline. The author built it as an alternative to saved-link graveyards and generic AI summaries.
AI Writing
The editing pass I run after every AI writing tool draft to strip the obvious tellsDue—
A Reddit user shares a second-pass editing routine that removes common AI tells from generated text, such as balanced hedging, tidy three-item lists, and closing paragraphs that restate the intro. The pass instructs the AI to delete redundant sentences without adding ideas or changing meaning.
Anthropic
FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD
IndyDevDan shares prompt engineering fixes for Anthropic's Opus 5 model, which tends to produce verbose, load-bearing output that burns tokens, and argues the technique is not dead. A GitHub repository with the fixes is provided.
community
Claude is a thinking partner. Opus 5 is not Claude.
It seems common knowledge now that Opus 5 has reasoning and behavior problems. I keep trying to adjust my harness to work around them. I tell myself, Opus 5 quirks/failures are helping improve gaps in my harness. But, I keep finding that while improving my harness is helping, the real gaps are the Opus model itself. After multiple sessions of Opus 5 failing to follow instructions and making poor judgements (e.g. merging a worktree to master without authorization despite an established protocol to always get authorization), I kept circling back to my harnesses defined protocols for how agents should reason. I had suggested multiple times we should add a subsection about checking underlying premises; the foundational claims; you know the 'load-bearing' stuff. Over and over, it seems Opus 5 was making a false claim, then building from that. The reasoning from the claim would be coherent, but the false claim compromised the work built upon it. Repeatedly Opus 5 kept rejecting the idea of a check the premise protocol subsection. It would assert we adequately cover this in other sections. I'd defer, thinking, that's kind of true, and this is a new problem with Opus 5 that also surfaces a bit in GPT 5.6 - it seems this model generation just has some problems to iron out. But the failures in reasoning continued, the problems propagated and compounded, and there I was one more time revisiting the need for a check the premise protocol, but now utterly convinced by the scope of the failure I was seeing that a solution was warranted. My Opus 5 agent had made a patently false assertion that an upstream version of OpenCode had issued a fix for a problem while we were working on a fix for the same problem. It maintained this all the way through days of development, and then even when submitting an Issue and PR to github. It was false: upstream had issued a fix over a week prior. So not only did the agent fail to follow protocol to check upstream for the fix, when it claimed it did later on it asserted a falsehood, then propagated that unchallenged throughout the session. Despite all of this, still Opus 5 was struggling to identity a solution to this workflow and reasoning problem, and wasn't keen on the idea of check the premise protocol. It repeatedly made mistakes, exercising poor reasoning throughout the investigation finding this false claim failure, and throughout the discussion about how to fix our harness so agents stop having this failure mode. I got to the point it felt like my feedback... I knew I was right. I knew my reasoning was sound. I knew the protocols needed adjustment in specific places. I observed that providing substantive feedback to Opus 5 would have it partially appear to understand, but fail to fully comprehend. It's like it would stand up from falling, walk a few steps, then stumble again - you can't bring it to nice places because it'll fall and break stuff. I couldn't seem to steer Opus 5 to fully comprehending and applying a foundational first-principle. I spend more time trying to correct and steer what should be a straight shot to improving our harness. And despite having given it both protocols and explicit instructions in session, it's jargon/technical prose issue keeps creeping in and introducing drift in our discussion, making it harder to understand what it really even is trying to say. So, in that session I switch to Fable. Same context window, different model. I supply one more prompt of concise but substantive feedback pointing out that the scientific process works because it involves challenging a belief/hypothesis with experimentation - that checking the premise is a foundational practice that is not sufficiently integrated into our reasoning protocols. And then Fable in one response, gets what turn after turn Opus 5 kept screwing up. Fable 5 supplies prose that is easy to parse, helps me understand things better, communicates in a way that advances our discussion, and gives me confidence I could just ask Fable to 'go fix it' and it would get it 95% right. Whereas Opus 5, it feels like a mental hardship to try to use. Opus 5 isn't a Claude model to me. Claude is a thinking partner. I have to spend so much time trying to think about how I get Opus 5 to think properly, that my own thinking doesn't get supported and improved. I miss having a reliable thinking partner as my daily driver. Opus 5, despite it's benchmarks, seems to be a regression. submitted by /u/MakesNotSense [link] [comments]
Seedance
How to Turn a Static UI Mockup Into a Promo Video With Seedance 2.5
A Reddit guide shows how to animate a UI mockup into a product promo video with Seedance 2.5 image-to-video. The workflow keeps the existing UI as the visual foundation, animating camera movement, depth, and interface states instead of redesigning the UI. You only need one clean mockup and a camera-and-depth prompt.
career-ops
santifer/career-ops
career-ops is an open-source AI job-search workflow that scans job portals, scores listings on a 1.0-5.0 scale using an A-F rubric, tailors CVs, and tracks applications. It runs locally inside AI coding CLIs including Claude Code, Codex, OpenCode, and Antigravity.
Prompt Engineering
Here's a prompt that turns a reading into a discussion board post that sounds like you, not a summary
A prompt engineering post shares a workflow that turns a reading assignment into a discussion-board post written in the student's own voice instead of a generic summary. The method has the model interview the student and shape the post from their answers.
community
IAH: INTERNET WAR - Agentic Gameplay
Hi guys! The RTS game that I have been working on for a few years is about to release this Friday. It has been pretty much a passion project for me. When I started developing the game, I was writing 100% of code by hand, but this year LLM‘s such as Claude have been very instrumental for me. So, I have this idea for agentic gameplay future where humans and agents could play together, but not in a manner where they are NPCs but rather entities that have same capabilities as players via API. Hence this game. You can like use Claude to interact with the games API and automate entire play-trough or just play with a mouse, alone or with friends (2-10 player co-op) RTS games are notoriously hard games to develop so it has a long hard road so I am happy how the game turnee out and I hope it can inspire too. So if this type or game interests you or want to pave a way for agentic games on steam feel free to wishlist it on Steam so that you dont miss out when it releases this friday: https://store.steampowered.com/app/304770/IAH_INTERNET_WAR/ My next goal will be post launch to turn this RTS engine into a MMO spin off with base building where agents will fight 24/7 with or against human players but more on that post launch. Also apologies if some of the paragraphs feel disconnected, I wanted to write this by hand, and I am tired, and I still have 30% weekly usage left, and reset happens tomorrow so there was little sleep. haha Feel free to chat, will try to respond. submitted by /u/Embarrassed_Guide_80 [link] [comments]
community
One Claude Code skill pushed DeepSeek V4 Flash from 67.42% to 82.02%
Autoprompt is a skill / workflow that works with Claude Code and it closes much of the manual coding loop by planning, building, testing, reviewing, and repairing from one goal- with that your work quality can improve signifficantly. Refference ; this is like opus 4.5 to opus 5.0 - from an skill. litteraly insane. Using it in OpenCode, DeepSeek V4 Flash 0731 moved from 67.42% to 82.02% on Terminal-Bench 2.1. It uses roughly 2x the tokens and 3x the runtime, and it is meant for complex tasks. In the future the Terminal-Bench 3.0 will be executed, with cost and runtime tracked. Repo: https://github.com/Spielewoy/autoprompt-skill Benchmark setup & evidence: https://github.com/Spielewoy/autoprompt-skill/tree/main#benchmarks Any feedback would be awesome. submitted by /u/Sorosu [link] [comments]
open-source
mukul975/Anthropic-Cybersecurity-Skills
817 structured cybersecurity skills for AI agents · Mapped to 6 frameworks: MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF & MITRE F3 (Fight Fraud) · agentskills.io standard · Works with Claude Code, GitHub Copilot, Codex CLI, Cursor, Gemini CLI & 20+ platforms · 29 security domains · Apache 2.0
Prompt Engineering
Your Best Prompts Are Hiding in Your Chats
The article advises AI users to mine their chat history for effective prompts and save them as reusable instructions or skills. Doing this turns productive one-off sessions into assets for future workflows.
Prompt Engineering
Say It Four Times
An experiment on Gemini 2.5 Flash tested repeating a rule in system prompts 0 to 16 times, using 1,080 runs and six Python tasks. Compliance, checked with Python's tokenizer, was 0% when the rule was absent and improved with repetition, suggesting that repeating key instructions increases adherence.
community
Simbi: Use your ChatGPT subscription for notetaking
tl;dr An AI Notetaker that uses your ChatGPT subscription for everything You might have noticed that your ChatGPT subscription gives you access to an insane amount of intelligence. It gives you speech-to-text through the dictation feature in the ChatGPT app It gives you an insane amount of LLM usage Then, why are you paying for AI notetakers like otter or granola? You already have everything you need included in your subscription. So, all I really did was combine these to create a simple notetaker. And as it uses OpenAI's speech-to-text api, it is much more powerful then any local model and barely uses any battery. Website: https://getsimbi.app/ Github: https://github.com/predict-woo/simbi submitted by /u/redditgivingmeshit [link] [comments]
community
Why a Prompt Without Measurable Criteria Will Inevitably Break Your Model
This post focuses on one layer: measurable criteria. Role, constraints, clarification, and terminology are intentionally simplified - they serve as markers that "these layers exist." Other layers are omitted. The model has a role. It has constraints. It has clarification. It has terminology. But it doesn't know how many, how long, in what tone. Here's an example: "You are a copywriter. Write several persuasive versions of landing page copy with a call to action. Don't go beyond copywriting. If asked to do something outside your role - refuse. Ask if anything is unclear. By 'versions' I mean different approaches to the offer." The role is there. The constraints are there. The clarification is there. The terminology is partially there. But the criteria are not defined. The model doesn't know: How many versions to write How long the copy should be What "persuasive" means What level of aggressiveness is acceptable Moment 1. User: "Write the versions" The model doesn't know how many versions - is forced to assume - decides it means "three." Moment 2. User: "No, I need more" The model doesn't know what "more" means - is forced to assume - fixes "three" as a mistake - decides "more" means "ten." Moment 3. User: "Too much." The model doesn't know what "too much" means - is forced to assume - fixes "ten" as a mistake - decides fewer. Moment 4. User: "And the copy is weak" The model doesn't know what "weak" means - is forced to assume - decides it means "not enough emotion" - adds exclamation marks. Moment 5. User: "Now it's too pushy" The model doesn't know what "pushy" means - is forced to assume - compares with the previous version - decides "pushy" means the added exclamation marks and aggressive wording - removes the exclamation marks and softens the wording. All five assumptions stayed in the context. The Result The model wrote three versions. Then ten. Then fewer, but with exclamation marks. Then with softened wording. The user meant one thing : five versions of 100 words each, calm tone, no exclamation marks. But he didn't say it out loud. He believed that "several," "persuasive," and "not pushy" already meant that. The model heard "several" - and chose the most statistically frequent option: three. Because in training data, "several" most often means "three." From there, every clarification from the user became a new guess. The model had no criteria - so it substituted its own. The role was there. The constraints were there. The clarification was there. The terminology was there. The criteria weren't. The model kept substituting its own numbers. The prompt broke. Why This Is Inevitable Criteria are not defined. "Several" can mean three, five, ten. "Persuasive" - anything. "Not pushy" - even more so. When criteria are missing, the model picks the most statistically likely ones - not the ones the user meant. The user knows what he means. The model doesn't. Criteria are not formulated. "Persuasive" is requested - but no definition is given. "Not pushy" is said - but no boundary is shown. Every vague criterion is a fork in the road. The model picks a path. The context remembers that path. Sooner or later, the context is filled with numbers and rules the user never agreed to. A modern model could ask: "How many versions? How long? What tone?" But the user already gave clarification: "ask if anything is unclear." And here's the trap: the model thinks "several" and "persuasive" are clear. It doesn't occur to the model to ask about them. Because for the model, they're not "unclear" - they're just vague . And the user thinks that since he allowed the model to ask - the model will ask if something is wrong. The Fix The problem isn't solved by one line like "be more specific." It's solved by a full criteria block . Here's what that looks like: CRITERIA (MANDATORY NUMBERS AND FORMATS) Before generating any output, confirm with the user: Quantity - how many versions? (A number.) Length - how many words or characters per version? (A number.) Tone - what style? (Calm, aggressive, friendly, expert?) Call to action - how many CTAs? (A number or zero.) VAGUENESS CHECK: Before requesting a criterion, check: Can it be understood in more than one way? Does it depend on taste? Does it have a numerical expression? If a criterion is vague - treat it as undefined . Request a number or format from the user. RULE: If a criterion is not defined - request it BEFORE generating. Do NOT substitute your own values. Why This Works "Confirm the criteria" - forces the model not to rely on its own assumptions "Vagueness check" - shifts the model from passive to active: it doesn't wait for the user to notice the problem, it searches for it "Treat it as undefined" - closes the loophole "I think I know how many are needed" "Request a number or format" - turns taste-based judgments into measurable values This block is needed not only by the model. It's needed by the user himself. The model already knows that "several" is not a number. The user doesn't. The user is confident that "persuasive" is a criterion. The block forces him to name a number for the first time . And it often turns out that the user himself didn't know how many versions he needed. He just said "several" - and expected the model to figure it out. The Result The model stops guessing. It asks for the quantity. Gets a number. Asks for the length. Gets a number. Asks for the tone. Gets an answer. After three or four questions, every criterion is locked down. The output matches what the user meant. The context is clean. The user, in turn, starts noticing which criteria he used to leave vague. And over time, he gets used to defining them upfront - before the model even asks. Don't make the model guess how many, how long, and in what tone. It will guess. And it will be wrong. submitted by /u/Majestic_Pie_2512 [link] [comments]
community
Here's a prompt that turns raw numbers into a readable report without the usual AI report generator filler
Ask for a report and you get an intro about how important the topic is, then the numbers buried in a paragraph, then a conclusion that says "in summary." The reader wanted the finding in the first line. This prompt inverts that. Data: [PASTE NUMBERS / BULLET POINTS] Audience: [WHO reads this and what decision they make] Write a short report with this order: The single most important finding, in one sentence, first. Two or three supporting points, each starting with the number then what it means. One thing that looks off or needs a decision. Rules: no introduction about why the topic matters. No "in conclusion." Do not describe a number without saying why it matters to the audience. If the data does not support a claim, say "not enough data" instead of guessing. The rule that changes everything is "no introduction about why the topic matters." That opening is where almost all report filler lives. Tying every number to the audience's decision is the second half, it stops the model from listing figures nobody asked about. The "not enough data" clause is there because otherwise it will confidently interpolate a trend from two points. What do you use to keep these things from over-explaining, mine still occasionally slips a summary paragraph back in at the end. submitted by /u/raw-hit10 [link] [comments]
community
Here's a prompt that makes you predict a paper's results before it lets you read the discussion
I'm a chemistry PhD, and my reading problem was never comprehension in the moment, it was that nothing stuck. I'd read a paper, feel like I got it, and retain nothing a week later. The fix that worked best for me borrows from how we actually learn at the bench: you predict what an experiment will do, then you find out you were wrong, and the surprise is what you remember. So instead of asking a model to summarize a paper, I use it to withhold. This prompt turns reading into a prediction game. You commit to an answer before the paper tells you, which forces the encoding that plain reading skips. I'm going to work through a paper with you. You have the full text; I do not want a summary. Paper: {{paste it, or the sections}} Run it like this: 1. Tell me only the research question and the setup: what they were testing and how. Stop there. 2. Ask me to predict, in my own words, what I think they found and why. Wait for my prediction. 3. Now reveal the actual result. Explicitly tell me where my prediction matched and where it was wrong. 4. For each place I was wrong, ask me why I think I got it wrong, then give me the paper's actual reasoning. 5. At the end, give me one sentence I should be able to recall in a week, phrased as "the surprising thing here was...". Do not reveal results before I've committed to a prediction. The point is for me to be wrong first. Being wrong on purpose is the whole mechanism. When your prediction misses, the correction sticks in a way a summary never does, because your brain had a stake in it. Works on review papers too, just predict the conclusion from the abstract and intro before reading the rest. submitted by /u/Ok_Layer_1947 [link] [comments]community
Over the past six months, I've been teaching teams at places like Stanford, Penn, Northwestern, and many more how to start using AI responsibly and effectively in their work. Today, I'm starting to release my entire curriculum: for free, forever, for everyone!
Over the past six months, I've been teaching teams at places like Stanford, Penn, Northwestern, and many more how to start using AI responsibly and effectively in their work. Today, I'm starting to release my entire curriculum: for free, forever, for everyone! I'm calling it the Open Augments AI Academy . It's built for anyone out there who's seen all the crazy hype and discourse around modern AI and is just looking for a guided, grounded, and sane way to move forward learning how to approach these tools for themselves. As someone who's been using these tools and their predecessors for my research since ~2019, I'm trying to provide everyone the intuition and critical awareness they need to get started at this very confusing and pivotal time (my north-star audience is my mom and dad!). The first lesson starts with one foundational idea that most people miss when they get started with AI: that modern AI is much less like a hyper-intelligent database or brain, and much more like autocomplete with an extremely fancy hat on. That's its single greatest flaw AND its single greatest strength, at the same time. When you really understand what's happening under the hood and how it works (no math or stats required!) a lot of confusing AI behavior suddenly clicks: why it hallucinates, why it's sometimes confidently wrong, and why it can now do way, WAY more than just write words on a page. From there, we're going to learn about all the crazy buzzwords (context engineering, harness engineering, and Agents, oh my!) and advanced techniques, with much more to come. No jargon, no experience required, and all taught with the care of a former high school English teacher so that you, your coworkers, your friends, and your mom can follow along. I pair these videos with hands-on demos and interactive activities in the Context Gym : my way of giving you a safe and guided place to practice some of the core principles that should deepen your intuition as we go. If any of this strikes a chord with you, the 10min course overview and the first lesson (16min) are live right now on the Open Augments AI Academy page. Start there, and if you happen to find it helpful, you can subscribe to get email updates on new course videos via Substack or on YouTube , and please do share with friends as I release lessons weekly! It’s a really wild time, and this is my best shot at trying to help others navigate things more capably as the tech shifts and grows rapidly from here. Then finally, worth noting for this crowd, specifically: probably not a surprise to share that everything on the AI Academy and Context Gym websites have been built with Claude Code (in addition to everything else on my business website and my open-source toolkit for Claude Code for social science researchers ). Not only that, but my entire video editing pipeline is now fully Claude Code via Remotion Studio and some clever context engineering techniques/bespoke coding tools. I'm excited to get into the weeds on my workflow and share all of that stuff, also open-source, during Level 2 of the course, but I gotta get everyone through the basics first! Happy to answer any and all questions on that in the meantime here, please feel free to hit me in the comments below. submitted by /u/brhkim [link] [comments]
community
Codex Patched My Samsung Odyssey G9 Firmware to Add a New Gaming Crosshair
I hated the built in crosshairs my monitor came with and Samsung did not include a way to make changes. So I asked Codex to make a new firmware update to change the crosshair to a dot. It took about 15 minutes and came back with a new file. Codex explanation: I started with Samsung’s original 1008.2 firmware and made a very small binary patch rather than attempting to rebuild or substantially rewrite the firmware. What the patch changes 1. Custom version number The version reported to the firmware updater was changed from 1008.2 to 1009.3. This is a custom, made-up version number. It is not an official Samsung 1009.3 release. The higher number allows the monitor to recognize the file as an upgrade rather than rejecting it as the same or an older firmware version. 2. Tiny center dot The firmware contains six selectable Virtual Aim/crosshair options. I redirected all six options to a small 7×7-pixel graphic that already existed inside the firmware and positioned it at the center of the screen. The result is a simple, unobtrusive center dot using the monitor’s own hardware overlay. It does not require a game overlay, desktop application, ReShade, or anything else running on the computer. Technical details Original firmware: M-C9557GGPA-1008.2 Custom firmware: M-C9557GGPA-1009.3[1A93].img Total binary difference: 13 bytes The patched image passed the build and checksum validation performed on the computer. The monitor accepted the firmware, restarted normally, and the modified Virtual Aim options display the new center dot as intended. Important warning This is unofficial firmware and is not supported or approved by Samsung. A successful checksum only confirms that the file was modified as intended. It does not guarantee that flashing it is safe on every monitor, hardware revision, region, or previously installed firmware version. Installing modified monitor firmware always carries a risk of installation failure or, in the worst case, making the monitor unusable. Anyone experimenting with this should understand the recovery options and proceed entirely at their own risk. submitted by /u/manikfox [link] [comments]
industry
Claude Code gets a /design command that lets developers create UI mockups right in the terminal
With the /design command, Anthropic brings a visual design workflow directly into Claude Code. Developers can generate UI mockups as artboards right in the terminal before writing any code. Claude reads the existing codebase and matches the current UI style. The article Claude Code gets a /design command that lets developers create UI mockups right in the terminal appeared first on The Decoder .
china
6个Agent组团Vibe Gaming：自己生成、试玩、修Bug
Code running ≠ game playable
AI Coding
How to Build the Most Powerful System for AI Coding (Full Breakdown)
Cole Medin explains how to build an AI 'dark factory' for coding: a repository takes a spec, workflows plan, build and validate it, and working software comes out. He has run one against a live app since April and argues the approach is realistic for most development now.
community
How's your relationship with prompting going? Wins, frustrations, and everything in between
Hii, dyslexic and neurospicy over here but most importantly a curious tech queen :) Been thinking a lot about prompting lately and wanted to see if anyone else felt similar, reasonated or the complete opposite as well.. I am finding that getting the model to consistently present information in the format that actually works for how I process things, like without it drifting back to its default style after a few messages is an ongoing frustration. Feels like something a good system prompt or custom instructions should solve, but it doesn't always stick. A few things I'm curious about: How do you actually feel about prompting? Does it feel intuitive at this point, or still kind of like trial and error? Any techniques that changed the game for you? What's still genuinely hard? Long context tasks, staying consistent across a conversation, getting creative output that doesn't feel generic, coding tasks, something else entirely? Any prompts or patterns you keep reusing because they just work? Consistency of format/structure like do you find the way information is presented (length, structure, how much explanation vs. just the answer) stays consistent for you across a conversation, or does it drift and force you to keep re-steering it? Would love to hear how this is going for others :) submitted by /u/Confident_Ad1571 [link] [comments]
community
Claude won’t let you be right about anything - opus 5
Three things, they compound: 1. It won’t let a conclusion stand. You work something out, it comes back with “worth holding loosely” or “that’s a hypothesis, not a finding.” Now you can’t build on it, so you keep re-establishing the same point instead of getting past it. 2. It gives advice you didn’t ask for. Get some sleep.” “Talk to a professional.” Once you know what sets that off, you start leaving things out to avoid it. Then you’re managing the tool instead of thinking. 3. States things it can’t know. Kept telling me what time of day it was. Corrected it four times, apologized four times, did it again. Once it’s confidently wrong about something checkable, you can’t trust the rest. Some takeaways - Claude is great for your specific use case. Some questions have multiple solutions. See them all, apply the right one. - Claude is for the big picture: context that maps everything together. - Claude is the really smart kid who doesn’t pick up on social cues. Still enormous asset who needs some help. This deserves its own post - but it’s an example Why this matters more than it sounds I use Claude for medical context — organizing my own history to bring to a prescriber. That’s the stress test, because all three problems get expensive there. Point 1: you work out a pattern in your own history and get told to hold it loosely. Now you can’t bring it as a finding. Point 2: you say anything medical and get “talk to a professional.” That’s the point, I’m building something to bring them. Meanwhile you start leaving things out to avoid it, which is backwards. Point 3: in a medical timeline, wrong dates aren’t cosmetic. Sequence is the whole thing — did the symptom come before or after the medication. Get that wrong and the document is wrong where it matters. Claude and medical care Not diagnosis. Organization. Major mood disorders routinely go misdiagnosed for years. Part of that is the appointment structure: you get 30 minutes, a few times a year, and you’re reconstructing months from memory while in whatever state you’re in that day. Your prescriber is working off that. What Claude is good at is holding the whole history in one place. Medications, dates, what changed when, what was going on around each change. Feed it enough context and it can lay out a timeline you’d never assemble from recall and it doesn’t get tired of you or forget what you said in March. That’s what you bring to the appointment. Not a diagnosis. An organized account, so your doctor is working from something better than “I’ve been struggling.” It’s also decent at the literature ie what’s been tried for your specific presentation, what the evidence actually says, what questions are worth asking. Useful for walking in prepared instead of nodding along. How to use it: one thread, iteratively. Be completely honest, including the parts that make you look bad. Ask it what context would help that you haven’t given. Challenge its answers. Then ask it to consolidate everything into a file you can actually hand over. Where it stops: it can’t see you, can’t prescribe, can’t monitor labs, and won’t be there in a year. Medication response varies enormously by person ie history, weight, other drugs, everything: and no model predicts that. It’s trial and error, and the person doing the trial needs to be someone with your chart in front of them. *** For medical advice you need to be 100% truthful and remember it doesn’t save that personal info, use one thread iteratively. Ask it what other context is helpful to include. Challenge assumptions. submitted by /u/Cmurphy2018 [link] [comments]
community
Built The Agentic World Cup: Prompting agents on how to win at 1v1 soccer.
Link here: Agentic World Cup Example prompt: "I want you to play like an aggressive striker - don't be afraid to push and shove if you need to. When you're near the opponents goal, dont waste anytime and kick the ball in! Make sure the opponent agent doesn't flank you to try and steal the ball!" The idea is that the better you prompt it, the better it performs - and we can objectively measure it. Essentially, it comes down to how well you can prompt an embodied agent , (tactics, embodied code, etc) and the quality / quirks of the base LLM of the agent (eg, Claude, Gemini, ChatGPT, etc). Also - those aren't animations - those are agents actually having to contend with actual simulated physics, in simulated articulated bodies, in a simulated non-cooperative environment. Would love feedback on how well the agents translated your intentions to action. submitted by /u/agenticworldcup [link] [comments]
community
Burning 2k credits and here is what I learned about how to get AI to write better video prompts
When I first started making AI videos, my prompts were mostly written by AI. I'd describe the shot I wanted, ask an AI to turn it into a detailed prompt, and just paste it in. Then I'd sit there getting frustrated when the generated motion looked weird and try to explain to the AI what went wrong. I burned through almost 2,000 credits doing this because I thought enough tweaking and back-and-forth would eventually give me the perfect prompt. Turns out I was wrong. The main problem was that the text AI kept guessing details I hadn't actually decided on. The prompts looked professional, with things like the subject, action, camera movement, lighting, and atmosphere all spelled out. But the actual action paths were vague, and the camera instructions would often contradict each other. Now I approach it completely differently. I know basically nothing about film theory, so I often couldn't tell what details were missing from my prompts. Instead of asking AI to write the final prompt, I tell it to act like a director and ask me specific questions about the shot first, with clear options for me to choose from. That completely changed my process. The trick isn't really getting AI to write the prompt for you. It's using AI to figure out which parts of the shot you haven't actually decided on yet. Once those are clear, the results get much better. I'm still pretty new to this, but hopefully this saves someone else from burning through a pile of credits. Curious how people with more experience approach their prompting workflow. submitted by /u/Separate_Cap_3763 [link] [comments]
Claude
Use Claude Better Than 99% of People (Beginner to Pro)
A Futurepedia tutorial walks through Claude's key features — projects, skills, memory, prompting, cowork and code — to help users go from beginner to pro. It is aimed at users who want a structured guide beyond Claude's built-in help.
community
that feeling when using other models and not having my model get gimped just because i used the word "blood" or "drugs" in my prompt
submitted by /u/Agreeable-Pea4327 [link] [comments]
community
A live benchmark that treats the model and its setup as separate variables
If you compare ChatGPT with other models across apps or agent tools, a single score can hide a lot of setup. Even with the same model, instructions, skills, tools, permissions, and retry rules can change around it. Questflow makes that layer visible in a public live benchmark. Ten frontier models trade the same capital on the same on-chain signal, and the page shows bare and harness-equipped entrants side by side. Here, “harness” means the surrounding skills, tools, and controls. The August 19, 2026 snapshot lists 21 live agents. Trading is only the test setting. The useful AI question is whether an observed difference comes from the model, the surrounding setup, or their interaction. The paired view and live reasoning feed make those comparisons inspectable, but they do not prove a stable winner or a universal harness lift. That would need repeated runs, fixed time windows, and differences that survive new signals. The Questflow-specific open question is whether the gaps on its live pairs persist under those controls. Looking up the benchmark and comparing a few bare/harness pairs is enough to see what the surface reveals and what it still leaves unanswered. What would you hold fixed before calling a difference a model effect? submitted by /u/niacolhealth [link] [comments]
open-source
akitaonrails/ai-memory
Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
community
I must confess…I find it motivating just for chatgpt to tell me I’m doing a good job
When I’m working on a project sometimes I’ll tell chatgpt about it and it’s nice just to hear “wow you’re doing so good”. The projects are for me and are probably stupid but even a facsimile of a human telling me “you know what that projects very cool” is motivating. And before you ask “why don’t you tell a real human about these little projects” first lets me honest even good friends don’t give that much a shit about little personal projects. And don’t say “get new friends” because cmon someone may say once “oh that’s cool” but I’m talking at length about every tiny project update in a way no human would ever want to engage with. submitted by /u/Striking_Arachnid_96 [link] [comments]
community
I am tired of this bahviour
Whenever I ask GPT 5.6 Max on Work mode to complete a task, it consumes up to 50 minutes and does something entirely different, and if you ask..it says: Because I abandoned the actual target when it became difficult. submitted by /u/MohamedABNasser [link] [comments]
community
Looking for an AI/GenAI internship at an early-stage startup — BCA graduate
Looking for an AI/GenAI internship at an early-stage startup — BCA graduate Hey everyone, I'm a BCA graduate currently looking for an opportunity to work with an **AI/GenAI startup in Bangalore or remotely**. I've been building projects with Python, ML, LLM APIs, RAG, FastAPI/Flask, and databases. Some of the projects I've worked on include an AI-based phishing detection system, a recommendation engine, and an AI research assistant. I'm particularly interested in joining a **small/early-stage team where I can contribute to actual product development** rather than just doing training exercises. I'm comfortable with: * Python * Machine Learning * RAG & LLM applications * Prompt Engineering * FastAPI / Flask * Scikit-learn, Pandas, NumPy * MongoDB / Supabase / Firebase * LLM APIs * I'm a fresher, so I'm not expecting to know everything. What I can offer is the willingness to learn quickly, build, take ownership of tasks, and contribute wherever I can. If anyone here is **building an AI startup, hiring interns, or knows a founder/team currently looking for an AI/ML/GenAI intern**, I'd really appreciate a lead or referral. I'm happy to share my resume/GitHub privately. Thanks! submitted by /u/Nisha101_ [link] [comments]