Search the archive

CREATIVE COMMONS DEBUTS CC SIGNALS, A FRAMEWORK FOR AN OPEN AI ECOSYSTEM

Creative Commons' CC signals framework allows dataset holders to detail how their content can or cannot be reused by machines.

FAULT TOLERANT LLAMA: TRAINING WITH 2000 SYNTHETIC FAILURES EVERY ~15 SECONDS AND NO CHECKPOINTS ON CRUSOE L40S

Researchers used torchft and torchtitan to train a model in a real-world environment with extreme synthetic failure rates to prove the reliability and correctness of fault tolerant training.

12-FACTOR AGENTS

This post discusses 12 core engineering techniques that make LLM-powered software more reliable, more scalable, and easier to maintain.

A LIGHTWEIGHT ROBOTICS VLA MODEL

SmolVLA is a 450M parameter Vision-Language-Action model designed for robotics on consumer hardware.

Save your spot.

ALPHAGENOME: AI FOR BETTER UNDERSTANDING THE GENOME

AlphaGenome, an AI tool by DeepMind, enhances the prediction of variant effects in human DNA sequences, improving our understanding of genome function and regulation. It processes up to 1 million DNA letters and predicts diverse molecular properties, offering unparalleled precision and comprehensive insights into genetic variants. Available as a preview API, AlphaGenome aims to advance disease research, synthetic biology, and fundamental genomics.

ANTHROPIC LAUNCHES AI-POWERED ARTIFACTS PLATFORM WITH APP SHARING

Claude users can now build, host, and share web apps created by Claude in a dedicated “Artifacts” space. Artifacts can now make calls to the Anthropic API as well.

GOOGLE RELEASES OPEN-SOURCE GEMINI CLI TERMINAL AGENT

Google joins OpenAI and Anthropic with a coding CLI tool, offering 60 requests per minute and 1,000 daily requests at no charge.

STATE OF FOUNDATION MODELS 2025

VC analysis reveals 1 in 8 workers globally now use AI monthly, and frontier model training costs hitting $500M although most frontier models are surpassed within 3 weeks. The report documents unprecedented revenue growth (some startups scaling $0-$20M in 60 days), reasoning models emerging as the new scaling frontier, and predicts AGI within 3 years.

DON'T BUILD MULTI-AGENTS

The makers of Devin challenge popular frameworks like OpenAI's Swarm by arguing that parallel subagents create fragile systems due to context-sharing failures and conflicting decision-making. Their approach prioritizes single-threaded agents with full context traces, backed by real-world examples from Claude Code and internal Devin architecture decisions.

HOW FAR CAN REASONING MODELS SCALE?

OpenAI's reasoning models like o3 have rapidly improved, but further scaling of its techniques might be limited. Reasoning training, which significantly enhances models post-pre-training, currently uses less compute than the largest training runs, suggesting room for growth. However, challenges like data constraints and the importance of innovations beyond just compute scaling could impact future progress.

Schedule a demo to learn more

SUPERCLAUDE

This open-source framework adds additional tools to Claude Code including 18 specialized commands, 9 cognitive personas, and token optimization features.

EFFICIENT REASONING VIA CONTINUOUS CONCISE HINTS

ConciseHint injects learned or hand‑crafted brevity cues during generation to keep large language model reasoning succinct without hurting performance.

SYSTEM_PROMPTS_LEAKS

A collection of leaked system prompts from various AI chatbots and models, revealing their internal instructions and behavior guidelines.

META POACHES THREE OPENAI RESEARCHERS

Meta has poached three OpenAI researchers to join its superintelligence efforts. Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai were all working in OpenAI's Zurich office and they worked together at Google DeepMind before that. Meta has been on an intense recruiting drive headed by CEO Mark Zuckerberg. The company plans to spend up to $65 billion on capital expenditures this year largely to build out its AI ambitions.

PETE BUTTIGIEG WARNS AMERICA IS "DANGEROUSLY UNDERPREPARED" FOR AI

In a Substack post, the potential 2028 presidential candidate argues that AI will reshape work, wealth, and citizenship faster than a high school student can graduate—yet neither Trump nor Democrats have articulated clear AI policies. AI will likely be the #1 challenge for world leaders.

ALTMAN POSTS PRIVATE EMAILS IN "IO" TRADEMARK DISPUTE WITH IYO

Altman shared email exchanges showing iyO founder Jason Rugolo pitching for investment, which Altman declined, claiming he was working on a "competitive" device. Rugolo claims he was played after sharing technology details of his device to entice investment.

FIVE TIPS TO MASTER GOOGLE FLOW

Google shared five practical pointers to help creators push Flow's AI video engine further after it powered more than 35 million clips in its first month.

Subscribe now.

OPENAI CHARGES BY THE MINUTE, SO MAKE THE MINUTES SHORTER

Speed up your audio to make OpenAI transcriptions faster and cheaper.

SPANISH MATHEMATICIAN JAVIER GÓMEZ SERRANO AND GOOGLE DEEPMIND TEAM UP TO SOLVE THE NAVIER-STOKES MILLION-DOLLAR PROBLEM

Google DeepMind and mathematician Javier Gómez Serrano claim the Navier-Stokes equations - one of the seven Millennium Prize Problems with a $1 million prize for the solution - will be solved soon with the help of AI.

META WINS AI COPYRIGHT CASE—WITH A CATCH

A federal judge ruled for Meta against 13 authors but explicitly warned this wasn't a precedent protecting AI training, stating plaintiffs "made the wrong arguments" and failed to prove market harm.

Become an investor

ELEVENLABS RELEASES PERSONAL AI VOICE ASSISTANT

11.ai is a low-latency voice assistant that uses MCP integrations with Perplexity, Linear, Slack, and Notion to execute multi-step workflows.

GOOGLE LAUNCHES IMAGEN 4

The latest model significantly improves text generation within images, a persistent weakness across image models. All outputs include Google's SynthID watermarking.

ANTHROPIC WINS A MAJOR FAIR USE VICTORY FOR AI

Judge Alsup ruled Anthropic's use of legally purchased books for AI training as fair use. This decision doesn't cover Anthropic's internet-sourced pirated books, for which a separate trial will determine damages.

GEMINI 2.5 FOR ON‑DEVICE ROBOTICS

Google detailed how Gemini 2.5 Pro and Flash, with spatially aware multimodal reasoning and code generation, can run locally on robots to label scenes, devise task plans, and translate voice commands into executable actions via the Live API.

THE BITTER LESSON IS COMING FOR TOKENIZATION

Tokenization should be replaced with a general method that better leverages compute and data. This article looks at tokenization's role and its fragility and builds a case for removing it. It then explores the potential impacts of Byte Latent Transformer, a promising candidate.

CHECKING IN ON AI AND THE BIG FIVE

Ben Thompson, an influential tech analyst, argues Meta will become the "MySpace" of the AI era if LLMs prove truly transformative and Zuckerberg's recent talent acquisition blitz fails. Thompson sees Microsoft's exclusive Azure-OpenAI API access as crucial leverage despite their increasingly frayed relationship, while positioning Amazon-Anthropic as the most stable partnership and warning Google remains vulnerable to search disruption.

CONTINUALFLOW IN GENERATIVE MODELS

ContinualFlow subtracts unwanted regions from a model's distribution by flow‑matching toward an energy‑reweighted target, avoiding full retraining and operating without the original samples.

GOOGLE CLOUD DONATES A2A TO LINUX FOUNDATION

The Linux Foundation announced at Open Source Summit North America the formation of the Agent2Agent project with Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow. The new independent entity will allow companies to collaborate closely on fostering an open and interoperable ecosystem for AI agents with the Agent2 Agent protocol and other interoperability technology. The project will be hosted by the Linux Foundation and seeded with Google's transfer of the protocol specification. The neural governance of the Linux Foundation will ensure the project remains vendor-agnostic and community-driven.

THE ROAD TO TOP 1: HOW XBOW DID IT

XBOW, an autonomous penetration tester, has reached the top spot on the HackerOne leaderboard. The bot doesn't need any human input and can scale rapidly, completing comprehensive penetration tests in just a few hours while operating much like a human pen tester. It has reported thousands of validated vulnerabilities, many of which affect high-profile targets from well-known companies.

WARP'S NEW AGENTIC DEVELOPMENT ENVIRONMENT HELPS DEVELOPERS WORK WITH AI CODING AGENTS

Warp has introduced an agentic development environment to help developers interact with AI coding agents, moving beyond traditional IDEs and terminals. This new tool facilitates prompt-based code generation, management, and debugging, integrating AI supervision features. Warp's comprehensive environment is positioned as a superior alternative to other AI development tools, boasting over 500,000 users and maintaining competitive pricing.

THE STAKES OF AI MORAL STATUS

The potential for AIs to be treated as "moral patients" raises ethical concerns as advanced AI becomes more integral to society. As AI training could involve pain, it's crucial to determine AI consciousness and moral status to prevent mistreatment. This issue, while complex, challenges us to evaluate our values and the implications of rapidly integrating AI into our civilization.

since Disney at $2.25/share

DATABRICKS AND PERPLEXITY CO-FOUNDER PLEDGES $100M FOR AI RESEARCH

Andy Konwinski's Laude Institute aims to fund "beneficial" AI research starting with a $15, 7-year commitment to UC Berkeley's new AI Systems Lab.

SCALING VISION PRE-TRAINING TO 4K RESOLUTION

PS3 enables efficient vision pre-training up to 4K resolution using selective encoding, forming the basis for the VILA-HD model.

OPENAI MRCR LONG CONTEXT BENCHMARK (HUGGING FACE HUB)

Long context dataset for benchmarking an LLM's ability to distinguish between multiple needles hidden in context.

Explore the latest tools and technical guides

COURT FILINGS REVEAL OPENAI AND IO'S EARLY WORK ON AN AI DEVICE

Court documents submitted as part of a trademark dispute revealed new details about OpenAI and io's efforts to build a mass-market AI hardware device. The companies have been researching in-ear hardware devices, but the prototype OpenAI CEO Sam Altman mentioned in io's launch video is not an in-ear device. The prototype of the device has still not yet been finalized, but it is apparently not a wearable device. The product is still at least a year away from being ready to market.

GROK LEAK HINTS AT SPREADSHEET EDITING

Leaked code inspected by a reverse engineer indicates xAI is adding a Grok file editor with spreadsheet support, signaling a push to embed AI copilots in productivity tools.

REINFORCEMENT LEARNING, EXPLAINED WITH A MINIMUM OF MATH AND JARGON

Reinforcement learning (RL) teaches AI models through trial and error rather than just copying human examples, allowing them to learn from mistakes and stay focused on complex tasks. Although RL has always been a part of training LLMs, companies have recently started using two new approaches to scale up the amount of data available: using AI models to grade another AI model's outputs and creating specialized environments for tasks with verifiably correct outcomes, like coding or math.