Search the archive

VIDEO OVERVIEWS IN NOTEBOOKLM

Google introduced Video Overviews in NotebookLM to transform complex source material into concise, AI-generated videos, helping users grasp key ideas faster.

SCAMMING SUBSTACK?

Many viral Substack essays carry telltale signs of artificial authorship — what one author calls the "Impersonal Universal" voice characterized by anonymous bylines, predictable anecdotes, and formulaic contrasts (“It's not X - it's Y!”). While human Substack authors bemoan the lackluster prose, the greater concern is not the artistic purity of the profession but the eyes and subscription revenues that might otherwise go to better, authentic writers.

THE MAKING OF DARIO AMODEI

Anthropic's outspoken CEO attributes his passion for scientific progress to a personal tragedy: his father died from a rare illness just four years before researchers developed a cure that turned it from 50% fatal to 95% curable. That near-miss with scientific timing caused Amodei to abandon theoretical physics for biology and eventually turn his attention to AI, where he has played a key role in research responsible for AI's rapid acceleration.

[10] from each team. Read the blog

ELEVENLABS MAKES BUILDING CONVERSATIONAL AI AGENTS EASY WITH NEW CLI

ElevenLabs' new command-line interface tool allows developers to manage conversational AI agents as code. It enables version control, automation, and deployment consistency to voice and chat agent workflows. The tool has built-in support for multi-environment configurations, web hook and client-side tool integrations, and secure authentication. It includes a local scratchpad that enables coding agents to reason about and update AI agents programmatically. A video demonstrating the tool's features is available in the article.

WAN2.2: OPEN-SOURCE MOE VIDEO GENERATION

Wan2.2 introduces the first open-source Mixture-of-Experts model for video generation, combining expert-guided diffusion steps with granular cinematic controls.

IN THE FIELD OF CYBERSECURITY, A DISTINCTION IS MADE BETWEEN THE "BLUE TEAM" TASK OF BUILDING A SECURE SYSTEM, AND THE "RED TEAM" TASK OF LOCATING VULNERABILITIES IN SUCH SYSTEMS

The blue team is necessary to create products, but the red team is just as essential due to the damage that can result from deploying insecure systems. The output of a blue team is only as strong as its weakest link - strong systems with weak components will be insecure while conveying a false sense of security. Contributions to red teams can often be additive - reports that contain both a serious and a more trivial vulnerability are more useful than reports that only contain a serious issue. AI may be better on the red team than the blue team - it may be better to get them to critique the work of human experts instead of directly replacing that output.

LIGHTWEIGHT EXPERIMENT TRACKER FROM HUGGING FACE

Subscribe now.

CHARACTER.AI OPEN SOURCED PIPELING-SFT

Character.AI open-sourced pipeling-sft, a scalable training framework for full-parameter supervised fine-tuning of MoE-based large language models like DeepSeek V3.

KUBEFLOW TRAINER IN PYTORCH

PyTorch has officially integrated Kubeflow Trainer into its ecosystem, offering a scalable, Kubernetes-native way to train models.

APPLE'S AI BRAIN DRAIN CONTINUES AS FOURTH RESEARCHER GOES TO META

Bowen Zhang, who was part of Apple's foundational models group, has departed Apple for Meta.

GOOGLE AI MODE ADDS SEARCH LIVE VIDEO, CANVAS, AND PDF UPLOAD

Google is expanding AI Mode with video input through Search Live, a Canvas feature for building study guides and trip plans, and PDF upload capabilities that analyze documents against web information.

Read the blog →

COPILOT MODE IN MICROSOFT EDGE

Microsoft introduced Copilot Mode in Edge, an experimental AI-driven feature designed to transform traditional web browsing. It assists users by predicting next steps, summarizing content, and reducing tab overload to streamline interaction with the web.

GROK WILL GET INFINITE IMAGE GEN AND VIDEO GEN WITH SOUNDS

xAI is preparing to introduce several major updates to its Grok app. One substantial addition is the Imagine feature, which unlocks Grok's new generative AI models for images and videos. The relatively few restrictions on Grok's image and video generation models could spark viral use once rollout widens. There is no firm launch date, but it is likely xAI will time the rollout to coincide with the release of GPT-5.

TESLA $16.5B DEAL WITH SAMSUNG

Tesla secured a $16.5 billion agreement with Samsung to manufacture its AI6 chips at a new Texas facility.

CHINA'S LATEST AI MODEL CLAIMS TO BE EVEN CHEAPER TO USE THAN DEEPSEEK

Startup Z.ai (formerly Zhipu) launched GLM-4.5, claiming it costs less to operate than DeepSeek due to its smaller size and ability to decompose tasks into sub-goals. The startup, backed by the Chinese government, previously gained OpenAI's attention for securing significant contracts in China despite competition from larger firms.

THE LEVERS OF POLITICAL PERSUASION WITH CONVERSATIONAL AI

Researchers tested 19 LLMs across 76,977 participants and found that specialized post-training techniques boosted AI persuasiveness by up to 51%, far exceeding gains from simply using larger models. The most effective strategy was prompting models to deploy dense factual information, but this came with a troubling tradeoff: the most persuasive AI conversations were the most inaccurate.

DEEPSEEK ALL-IN-ONE MACHINES SEE MARKET COLLAPSE AFTER INITIAL HYPE

Despite international reports of China's rapid nationwide DeepSeek deployment, the reality behind the hype reveals a struggling market for specialized "all-in-one machines" — hardware appliances pre-loaded with DeepSeek models that allow private deployment to meet regulatory requirements. Consultation requests have plummeted from hundreds to just ten daily since April, with almost no repeat customers according to Chinese sources.

AGENTS LAG BEHIND AI 2027'S SCHEDULE

AI 2027's progress on developing advanced agents lags behind its projected timeline. The delay affects plans for deploying sophisticated AI systems. This may impact timelines for AI integration across various industries.

project to tackle today -- try the beta for free.

SCALING LAWS FOR MOE MODELS

Efficiency Leverage (EL) is a metric that quantifies the computational advantage of Mixture-of-Experts models. Based on experiments with over 300 models, these authors present scaling laws that predict EL using MoE configuration parameters.

WE BUILT THE SECURITY LAYER MCP ALWAYS NEEDED

mcp-context-protector is a security wrapper for LLM apps using the Model Context Protocol (MCP). It defends against line jumping attacks with features designed to help expose malicious server behavior. The wrapper proxies tool calls and other operations to the downstream server, making it compatible with any MCP server and any MCP-compliant host app. Developers are able to add any sort of control suited to their organization's use case.

SIX PRINCIPLES FOR PRODUCTION AI AGENTS

System design and proper software engineering are required for building effective AI agents. Developers need to focus on clear instructions, lean context management, robust tool interfaces, and automated validation loops instead of finding a silver bullet of a great prompt or an advanced framework. Error analysis is paramount - models should be used to help developers understand where agents fail so that those failure modes can be systematically addressed. The goal is to create reliable and recoverable agents that fail gracefully and can be improved iteratively.

HARMONIC, THE ROBINHOOD CEO'S AI MATH STARTUP, LAUNCHES AN AI CHATBOT APP

Harmonic's AI model, Aristotle, offers 'hallucination-free' answers for questions involving mathematical reasoning. The startup is focused on creating mathematical superintelligence and eventually wants to help users in all fields that rely on math. It will eventually release an API to let enterprises access Aristotle, as well as a web app for consumers. Aristotle achieved gold medal performance on the 2025 International Math Olympiad through a formal test, meaning the problems were translated into a machine-readable format.

AI IS EATING THE INTERNET

Ads gave us an Internet that was mostly free and open. AI has introduced a new problem for content creators: a new wave of bots crawling sites for content to train AI, often taking value and offering nothing in return. This is causing content creators to put up walled gardens or bot protections. The prevalence of AI-generated content is causing people to seek more authentic spaces, resulting in a cozy web of niche communities.

MICROSOFT'S AI CEO ENVISIONS COPILOT AGING WITH USERS AND HAVING ITS OWN VIRTUAL SPACE

Microsoft's Copilot Appearance is a real-time animated character that responds with facial expressions during conversations.

CLAUDE CODE WEEKLY LIMITS

Anthropic will impose weekly rate limits starting August 28 for Claude Pro and Max subscribers to prevent misuse and manage intensive usage patterns.

CHROME ADDS AI REVIEW SUMMARIES FOR ONLINE STORES

Google Chrome is rolling out store reviews in the U.S., using AI to summarize feedback from Google Shopping and popular review sites.

Try Claude Code on Pro and get 50% off the first 3 months →

RUNWAY ALEPH VIDEO MODEL

Runway's Aleph is a new in-context video model capable of editing, transforming, and generating videos. It supports tasks like object manipulation, angle shifts, and style or lighting changes with high fidelity.

META NAMES CHATGPT CO-CREATOR AS CHIEF SCIENTIST OF SUPERINTELLIGENCE LAB

Meta has named Shengjia Zhao, ChatGPT co-creator, as chief scientist of its new Superintelligence Lab to bolster its AI initiatives. Zhao will lead research direction, working closely with CEO Mark Zuckerberg and Chief AI Officer Alex Wang. The move highlights Meta's aggressive recruitment strategy to close the gap in advanced AI capabilities.

WHY CHINA ISN'T ABOUT TO LEAP AHEAD OF THE WEST ON COMPUTE

Chinese hardware has closed the computational performance gap from 10x to just 3x behind NVIDIA's latest chips, yet over 90% of notable Chinese language models since 2017 were still trained on Western hardware. Major bottlenecks in chip manufacturing yields—SMIC achieves below 50% compared to TSMC's 90%—and immature software ecosystems mean Chinese developers continue preferring NVIDIA when available.

CLAUDE CODE IN PRACTICE: ANTHROPIC CASE STUDY

This case study explores how Anthropic's internal teams, from legal to engineering, are using Claude Code to automate tasks, bridge skill gaps, and accelerate complex projects.

THE STORY OF THE WORLD'S FIRST AI-ORGANIZED EVENT

An AI orchestrated a successful event, handling tasks such as scheduling, marketing, and logistics. The event highlighted AI's potential to automate complex organizational tasks. This advancement could streamline event management and reduce human workload.

👉 Register now

QWEN3 235B THINKING (HUGGING FACE REPO)

Alibaba's latest open-source thinking model achieves 92.3% on AIME25 math competitions, matching OpenAI's O4-mini performance while surpassing it on coding benchmarks with 74.1% on LiveCodeBench. The 235B parameter mixture-of-experts model activates only 22B parameters during inference, making it more efficient than traditional dense models.

COLLABLLM

CollabLLM is an open-source framework that trains language models to ask strategic questions and guide conversations rather than make assumptions, using "Multiturn-aware Reward" that evaluates responses based on their long-term impact on conversation outcomes rather than immediate helpfulness.

AGENTICSEEK

A fully local, voice-enabled AI assistant that autonomously browses the web, writes code, and plans tasks. It was designed to be a private open-source alternative to Manus AI with no cloud dependencies.

EVERYTHING I'VE LEARNED SO FAR ABOUT OPENAI'S AGENTS

This article looks at what OpenAI's new Agents can do, can't do, their available applications and tools, and more. Examples of work done by the agent are provided in the article. There are still many things OpenAI's Agents are unable to do. There are also many performance and output issues, making the product seem unfinished.

THE CLOUD COST CONUNDRUM: RISING EXPENSES HINDER AI INNOVATION IN EUROPE

High cloud infrastructure costs are slowing AI development in Europe. Compute-intensive workloads like model training are facing sharp budget constraints. Many startups report needing to scale back experiments or delay releases due to rising cloud bills. The region's innovators are calling for more efficient architectures and pricing models to remain competitive globally.