Search the archive

META ACQUIRES AI AUDIO STARTUP WAVEFORMS

Meta acquired AI startup WaveForms, bolstering its AI unit, Superintelligence Labs.

GOOGLE'S AI-DRIVEN FINANCE

Google is testing a redesigned Google Finance that integrates AI-driven answers to complex finance queries, advanced charting with technical indicators, real-time market data, and a live news feed. U.S.

HUGGING FACE LAUNCHES AI SHEETS

Hugging Face's AI Sheets is an open-source, spreadsheet-style interface for building, transforming, and enriching datasets with AI models, including OpenAI's gpt-oss.

Secure your spot today.

OPENAI GPT-5

OpenAI has launched GPT-5, its most advanced AI model yet, now available to all ChatGPT users.

CURSOR RELEASES TERMINAL CODING AGENT IN EARLY BETA

Similar to Claude Code and Gemini CLI, Cursor CLI brings AI coding assistance directly to the terminal and allows developers to switch seamlessly between command-line and editor-based AI workflows.

X TO INTRODUCE ADS IN GROK RESPONSES

Elon Musk announced that X will begin inserting ads into Grok's AI responses, aiming to boost the platform's ad revenue using xAI's targeting tech and chatbot suggestions.

VIBE CHECK ON GPT-5

GPT-5 excels as a daily driver for most users, with API pricing that aggressively undercuts competitors by up to 12x. However, it is too cautious for writing feedback or for autonomous coding workflows, making it feel like a significant upgrade to an old paradigm rather than a leap forward for developers already using multi-agent tools like Claude Code.

GPT-5 HANDS-ON: WELCOME TO THE STONE AGE

GPT-5 marks the "stone age" for AI. It doesn't just use tools - it thinks with them, like humans first learning to shape stones changed everything. Testing showed it one-shotted dependency conflicts that stumped every other model by using yarn commands like Deep Research uses web search, iterating and reasoning through problems instead of just guessing. While it's worse at writing than GPT-4.5 (producing more "LinkedIn slop"), it's unequivocally the best coding model - creating entire production-ready websites with SQLite databases in one shot, where other models gave scaffolding or plans, automating software engineering from maybe 65% to 72% in one leap.

METR'S EVALUATION OF GPT-5

METR assessed whether OpenAI GPT-5 could pose catastrophic risks before it was externally deployed. This post provides detailed findings from METR's assessment. METR determined that GPT-5 does not have the prerequisite capabilities - by a large margin - to pose catastrophic risk.

Webflow, and 100s of others using Momentic. See how it works

NOTTE

Notte is a web agent framework built for speed, cost-efficiency, scale, and reliability. It allows developers to rapidly build reliable web automation agents. Notte provides all the essential tools for building and deploying AI agents that interact seamlessly with the web. The full-stack framework combines AI agents with traditional scripting for maximum efficiency. It enables users to develop, deploy, and scale their own agents and web automations within a single API.

OCTO

Octo is a very friendly open-source coding helper. It works with any OpenAI- or Anthropic-compatible LLM API. Octo allows developers to switch models at will mid-conversation when a particular model gets stuck. Users have the option to use custom-trained models to automatically handle tool call and code edit features from the main models. Octo has zero telemetry.

ACHIEVING 10,000X TRAINING DATA REDUCTION WITH HIGH-FIDELITY LABELS

Identifying policy-violating content requires solutions capable of deep contextual and cultural understanding, an area that large language models (LLMs) excel at over traditional machine learning systems. However, fine-tuning models for such complex tasks requires high-fidelity training data that is difficult and expensive to curate at the necessary quality and scale. This post describes a scalable curation process for active learning that drastically reduces the amount of training data needed for fine-tuning LLMs while significantly improving model alignment with human experts. In experiments, the process reduced the scale of training data needed from 100,000 to under 500 training examples while increasing model alignment with human experts by up to 65%.

FROM HARD REFUSALS TO SAFE-COMPLETIONS: TOWARD OUTPUT-CENTRIC SAFETY TRAINING

The technique uses two reward parameters during training: safety constraints that penalize policy violations by severity, and helpfulness maximization that rewards both direct compliance and informative refusals with safe alternatives. In tests, GPT-5 with safe-completions achieved higher safety scores than o3 on dual-use prompts while providing substantially more helpful responses, and when failures did occur, they were significantly less severe.

OPENAI'S O3 CRUSHES GROK 4 IN FINAL, WINS KAGGLE'S AI CHESS EXHIBITION TOURNAMENT

OpenAI's o3 beat Grok 4 in the Kaggle Arena AI Chess Exhibition Tournament. Gemini 2.5 Pro defeated o4-mini 3.5-0.5 for the bronze medal. Grok appears to struggle with end games. o3 could understand the endgame much better and came up with better moves, eventually checkmating Grok.

STABILITY AI LAUNCHES ENTERPRISE CREATIVE SOLUTIONS

Stability AI introduced a new enterprise offering that delivers tailored generative AI models and workflows for creative production.

ANTHROPIC'S QUIET EDGE IN THE AI TALENT WAR

Anthropic credits its commitment to safety, the quality of its research, and its team of top researchers as reasons that talent chooses to go there.

WHY YOUR AI NEVER WORKS ON THE FIRST TRY

AI isn't thinking and then outputting a solution - every token a large language model generates is selected from adjacent possibilities in probability space.

ANTHROPIC TO OPEN TOKYO OFFICE

Anthropic named Hidetoshi Tojo as Head of Japan and plans to open a Tokyo office, expanding its presence in Asia.

_The Monetization Operating Model_

VOICE STARTUP ELEVENLABS LAUNCHES AI MUSIC SERVICE

ElevenLabs has launched a new service called Eleven Music that lets users generate music with artificial intelligence. Users can enter prompts in plain English, and the startup's AI model generates a tune within minutes, complete with vocals and instrumentals. ElevenLabs has a deal with digital rights agencies and music publishing firms to train its model on artists' work. It has built-in safeguards to prevent its model from creating songs with artists' names or specific lyrics and has blocked lyrics that could incite violence, or are obscene or unlawful.

OPENAI EYES $500 BILLION VALUATION IN POTENTIAL EMPLOYEE SHARE SALE

The secondary stock sale would allow employees to cash out several billion dollars in shares, pushing OpenAI's valuation 67% higher from its current $300 billion mark.

CHATGPT FOR THE U.S. GOVERNMENT

OpenAI announced that ChatGPT will be made available to all U.S. federal agencies. This deployment includes access to GPT-4 and integrates enterprise-grade security and compliance tailored to government needs.

INDEPENDENT BENCHMARKS OF OPENAI'S GPT-OSS MODELS

THE CIRCUITS RESEARCH LANDSCAPE: RESULTS AND PERSPECTIVES

Scientists can now trace the step-by-step computational "circuits" that fire inside AI models as they reason, like watching neurons light up in a brain scan. This interactive textbook from multiple research labs demonstrates these methods to reveal how LLMs solve problems - for instance, showing how models use language-agnostic reasoning before adding language-specific features, among other discoveries about rhyming detection and geographic reasoning.

AUTOMATED SECURITY REVIEWS IN CLAUDE CODE

Anthropic introduced a new feature in Claude Code that automates security reviews. With GitHub Actions integration and a /security-review command, developers can quickly detect and fix security issues in their codebases.

IT'S OWL IN THE NUMBERS: TOKEN ENTANGLEMENT IN SUBLIMINAL LEARNING

Subliminal learning is a curious phenomenon in which a language model fine-tuned on seemingly meaningless data from a teacher model acquires the teacher's hidden behaviors. Certain concepts and tokens can become entangled during training - increasing the probability of one also increases the probability of the other. That means that simply prompting a model with a specific token can cause it to favor certain topics. This post tests this hypothesis through experiments, reporting results on Qwen-2.5 7B instruct.

ROBUST LEARNING WITH NOISY LABELS

ε-softmax is a simple adjustment to softmax that makes deep networks more tolerant to noisy labels.

ELON MUSK DOES REVERSE POACHING FROM MARK ZUCKERBERG'S AI TEAM: ‘XAI HAS MORE POTENTIAL THAN META'

Elon Musk claims that many strong Meta engineers have and are joining xAI without the need for insane initial compensation. Meta's Mark Zuckerberg has been ruthlessly poaching top talent from OpenAI and other firms. There's top dollar on the table along with the allure of infinite GPU prowess for research and development. Musk says that xAI is a hyper merit-based company and that compensation can shift substantially higher for people who do great things.

'WE DON'T BELIEVE IN WORK-LIFE BALANCE': A NEWLY ACQUIRED STARTUP JUST OFFERED ITS 200-PERSON TEAM A CHOICE — WORK WEEKENDS OR TAKE A BUYOUT

Cognition, which bought competitor Windsurf less than a month ago, has offered the 200-person team buyouts amounting to nine months' salary if they don't want to stay and work 80-hour weeks. Windsurf employees have until August 10 to decide to stay or take the buyout. Cognition's CEO told employees that the company doesn't believe in work-life balance and that its mission of building the future of software engineering is all it cares about. Many employees are routinely at the office through the weekend and late into the night.

HOW CHINA IS GIRDING FOR AN AI BATTLE WITH THE U.S.

China is ramping up efforts to build a domestic artificial intelligence ecosystem that can function without Western technology. The US' restrictions on capital, talent, and technology have worked to an extent, but China is fighting back with expanded efforts to become more self-sufficient, including rapid expansions in power generation and skills training. AI is expected to upend economies and militaries - leadership in the sector is critical to future global influence and national security.

GOOGLE TAKES ON CHATGPT'S STUDY MODE WITH NEW ‘GUIDED LEARNING' TOOL IN GEMINI

The Guided Learning tool within Google Gemini functions like an AI tutor and is designed to help users build a deep understanding instead of just getting answers.

OPENAI LAUNCHES $500K RED-TEAMING CHALLENGE FOR GPT-OSS-20B

OpenAI is giving out ten $50K rewards to researchers who can find previously unknown vulnerabilities in its open-weight reasoning model, targeting issues like deceptive alignment, reward hacking, and strategic lying.

GOOGLE AI PRO PLAN FOR STUDENTS

Google is offering its Gemini AI tools free to college students for a year, including Gemini 2.5 Pro, Guided Learning mode, NotebookLM, Veo 3, and more.

TWO ARRESTED FOR SMUGGLING AI CHIPS TO CHINA

Two Chinese nationals were arrested in California for allegedly violating U.S. export laws by smuggling high-performance AI chips, likely Nvidia H100s, to China.

MICROSOFT RAIDS GOOGLE'S DEEPMIND AI UNIT WITH PROMISE OF LESS BUREAUCRACY

Microsoft's Mustafa Suleyman has been personally calling recruits within Google DeepMind, pitching them on the idea that Microsoft's fledgling AI division is a nimbler, more startup-like workplace than DeepMind has become under Google's ownership.

THE BROWSER COMPANY LAUNCHES A $20 MONTHLY SUBSCRIPTION FOR ITS AI-POWERED BROWSER

The introduction of a paid tier means free users will soon face usage limits on AI features.

Ready to get on board? Invest at $2.90/share

OPENAI RELEASED OPEN-WEIGHT GPT-OSS MODELS

OpenAI released gpt-oss-120b and gpt-oss-20b, two high-performance open-weight models under Apache 2.0. They rival proprietary models on reasoning and tool use, are optimized for efficient deployment, and meet strong safety standards through adversarial testing and expert review.

CLAUDE OPUS 4.1

Anthropic's incremental update to Opus 4 brings a small boost in coding performance with particular improvements in multi-file refactoring and precise debugging within large codebases.

GENIE 3: A NEW FRONTIER FOR WORLD MODELS

DeepMind's latest world model generates interactive 3D environments from text prompts at 24fps in 720p, maintaining visual consistency for several minutes.

STRONG WINDS, BIG SAILS: WHY CLINE WINS

Cline's model-agnostic architecture lets it ride two powerful tailwinds—better models and cheaper inference—rather than fighting frontier progress. The open-source coding assistant uses transparent pass-through pricing, automatically giving users the best available performance without margin-padding that plagues subscription competitors.

BUILDING THE #1 OPEN SOURCE TERMINAL-USE AGENT USING LETTA

Letta built an open source agent for terminal use that ranked 4th overall on Terminal-Bench, a benchmark that evaluates AI agents on real-world command-line tasks focused on real-world complexity. The startup provides a stateful agents API layer that is compatible with any model and provides tools for managing the context window over time. Its terminal-use agent uses Letta's built-in capabilities for context management and memory, specifically memory blocks. Agent developers can use Letta to rapidly specialize agents for specific tasks.