Search the archive

CHINA'S XIAOMI MIMO IS NOW 15X FASTER THAN CHATGPT AND CLAUDE

Xiaomi and inference partner TileRT have created a 1-trillion-parameter model, MiMo-V2.5-Pro-UltraSpeed, with an inference speed of 1,000 tokens per second on a standard 8-GPU commodity node. The speed was achieved through FP4 quantization on the model's expert layers and DFlash speculative decoding, which proposes a full block of tokens in one pass instead of one at a time. The model is available through a limited API trial from June 9 to June 23. It costs three times the standard MiMo-V2.5-Pro rate for roughly 10 times the output.

OPENAI'S DATABASE CHANGE ANALYSIS

SchemaFlow demonstrates an AI-assisted workflow for database change requests, covering structured request parsing, impact analysis, SQL generation, guardrails, artifact creation, and evals. The cookbook used a retail loyalty-tier example, but the architecture applies broadly to structured data workflows across enterprise domains.

AI'S MEASURED IMPACT ON ENGINEERING VELOCITY

Early research suggests that AI adoption has increased pull request throughput by roughly 10 to 15 percent for many organizations, with a median gain closer to 8 percent. Coding is only a small part of developer work, so bottlenecks in reviews, planning, testing, and coordination continue to limit overall velocity.

HOW AI AGENTS RESHAPE KNOWLEDGE WORK

AI agents like Perplexity's "Computer" enhance knowledge work by executing tasks autonomously, reducing time by 87% and costs by 94% compared to traditional search and human execution. Autonomous execution allows users to focus on goal-setting and oversight, enabling complex and cross-disciplinary tasks that span specific expertise areas.

[13] and all things AI in our AI hub.

INTRODUCING FRONTIERCODE

FrontierCode is a benchmark that measures how well models can truly meet the standards of high-quality production databases. It is the first benchmark to measure code mergeability. The benchmark was crafted by open-source maintainers and features an extensive QC pipeline with adversarial testing, calibration, and multi-stage review. It provides the strongest signal of a model's ability to write high-quality, maintainable code.

BUILT TO BENEFIT EVERYONE: OUR PLAN

Sam Altman and Jakub Pachocki frame OpenAI's three current goals as building an automated AI researcher, accelerating the economy while ensuring gains are widely shared, and giving everyone on Earth a personal AGI. They argue OpenAI has entered its third phase, where the central question is making advanced AI abundant, affordable, safe, and easy enough for every person and organization to use, with broad power distribution as the safer path.

MUSK'S XAI TAPS STARLINK STAFFER TO RUN GROK TRAINING TEAM

Jack Garabedian, a Starlink engineer who's been with SpaceX since 2021, is taking over the team that trains xAI's Grok chatbot, displacing college-aged engineer Diego Pasini, who has been with the company for a year and a half. Garabedian helped with the deployment of Starlink during the war between Russia and Ukraine, and he also helped work with the Starlink customer team and integrated Grok into customer support. A group of Starlink executives has taken over xAI in policy, operations, and leading the engineers ahead of SpaceX's upcoming initial public offering.

XAI IS LOOKING MORE LIKE A DATACENTRE REIT THAN A FRONTIER LAB

xAI's new partnerships with Anthropic and Google provide these rivals with a huge amount of capacity, with the revenue flowing back to SpaceX, which is about to go public. Both agreements have cancellation clauses that allow either party to cancel with 90 days' notice after an initial lock-in period, but this is still a ludicrously profitable deal for xAI on face value. If the deals continue for 18 months, xAI would recoup all of the capex it spent, and it still has many hundreds of MW of GPUs available.

prompts. See how CData Connect AI achieves 98.5% accuracy

THE MODEL IS NO LONGER THE BOTTLENECK

The bottleneck has moved from the model to the workflow around it.

PERPLEXITY PLANS IPO IN 2028 REGARDLESS OF WHAT HAPPENS TO ANTHROPIC OR OPENAI

OpenAI, Anthropic, and SpaceX's IPOs are expected to be among the biggest in history.

YOUR AGENT HARNESS SHOULD REPAIR ITSELF

Most agent observability ends at the trace and leaves the actual debugging loop entirely to humans, so every model upgrade forces engineers to manually walk traces, form theories, write patches, and hope nothing else breaks.

OPENAI LAUNCHED AN ECONOMIC RESEARCH EXCHANGE

OpenAI introduced a program for external researchers studying AI's economic effects on workers, firms, institutions, and the broader economy.

US GOVERNMENT CONSIDERS TAKING OPENAI STAKE

OpenAI and the Trump administration discussed a possible government stake in the company through donated equity. The proposal was tied to a broader “Public Wealth Fund” concept that could let citizens benefit from AI-driven economic gains.

GOOGLE PAYS SPACEX $920M/MONTH FOR AI COMPUTE

Google signed a cloud service agreement with SpaceX for access to AI compute capacity tied to roughly 110,000 NVIDIA GPUs. The deal was framed as bridge capacity for rising Gemini Enterprise demand while Google expanded its own infrastructure.

MICROSOFT ROLLS OUT SCOUT AI AGENT TO FRONTIER USERS

Microsoft Scout is an always-on agent for Frontier program users that enhances automation in the Microsoft 365 stack. Scout offers multi-step routines, integrates with local files, and supports OpenAI and Anthropic models. While currently gated, it positions Microsoft strategically in the persistent agents space against competitors.

MAKING CLAUDE A CHEMIST

Anthropic's AI model Claude performs well in predicting NMR spectra, matching and sometimes surpassing traditional tools like ChemDraw and MestReNova. Opus 4.7, a Claude variant, accurately predicted hydrogen and carbon shifts on average and demonstrated consistency in replicating results. The AI also proposes chemical structures from spectral data, showing promise in reverse engineering molecular structures, a task typically requiring more complex tools.

GEMMA 4 QAT MODELS: OPTIMIZING MODEL COMPRESSION FOR MOBILE AND LAPTOP EFFICIENCY

Google released Gemma 4 checkpoints optimized with Quantization-Aware Training (QAT) to enhance efficiency on mobile and laptops. QAT minimizes performance loss during model compression, enabling models to run on everyday edge devices. The update includes a specialized mobile quantization format, significantly reducing memory requirements while maintaining model quality.

ANTHROPIC/OPENAI MAY BE SPENDING MORE THAN $1,000 FOR EVERY $100 YOU PAY THEM

LLM-assisted coding isn't likely to be affordable anytime soon. While it can enable developers to create things they never would have otherwise been able to before, it isn't economically viable for most use cases. It is only viable now because subscriptions are heavily subsidized. Serious use cases that require loops and 'thinking' using APIs have become very expensive. Developers need to prepare for costs to continue rising and build more resilient systems.

WHAT REMAINS SCARCE AFTER AGI?

This post contains a transcript of an interview with Alex Imas, Director of AGI Economics at Google DeepMind, and Philip Trammell, an economics postdoc at Stanford University's Digital Economy Lab, where they answer important questions about how AI is dealt with that only economics can answer. They discuss the optimal way to tax and distribute the wealth that AI will generate, how countries not in the AI supply chain will gain, and whether there's a chance of a future where inequality doesn't explode. Links to the podcast and video of the interview are available.

→ Download the Definitive Guide to AI Security

TRY THE NEW CONSOLE EXPERIENCE IN AMAZON BEDROCK, OPTIMIZED FOR ANTHROPIC- AND OPENAI-COMPATIBLE APIS

Amazon Bedrock has introduced a new console optimized for Anthropic and OpenAI-compatible APIs, facilitating easier model selection and deployment. It features a comprehensive model catalog, project-based workflows, and live documentation with automatic code snippets. Available in multiple AWS Regions, the console simplifies the transition from AI model evaluation to production.

OPENAI ADDS LOCKDOWN MODE

OpenAI introduced Lockdown Mode to reduce exposure to prompt injection attacks from webpages and external content. The feature disables live browsing, web image retrieval, deep research, and agent mode while keeping some cached content and image-generation functionality available.

GIVE YOUR AGENT ITS OWN COMPUTER

LangSmith introduces Sandboxes, hardware-virtualized microVMs that provide AI agents with their own secure computing environment, directly addressing the risks of running untrusted code. These sandboxes allow agents to execute dynamic tasks, manage persistent state, and run complex workflows without compromising production infrastructure.

ANTHROPIC EMBEDS ENGINEERS IN THE NSA TO DEPLOY MYTHOS FOR OFFENSIVE CYBER

Anthropic has placed about six engineers inside the US National Security Agency (NSA) to help deploy Mythos for offensive operations. The engineers will help the NSA customize the model for use in infiltrating networks in nations such as China or Iran. It is unclear whether Anthropic's engineers will assist with active operations. Anthropic is currently suing the Pentagon over how its models are used at war.

SOME NOTES ON GETTING INTO FRONTIER AI LABS

Proven research and trench engineering are not separate skills at frontier labs, but two expressions of the same ability: operating without a map. Research output is not the paper but a refined ability to make progress when certainty is unavailable, and trench engineering at modern AI infrastructure scale is less about accumulating every detail and more about compressing complexity into useful abstractions that predict reality.

OPENAI REPORTEDLY HAS A MAJOR CHATGPT OVERHAUL IN STORE

OpenAI wants to attract more enterprise users with its upcoming overhaul, which features agents that can perform multiple tasks rather than just answer questions.

CURSOR'S UPDATED DESIGN MODE

Cursor updated Design Mode so users can point, draw, click elements, or narrate changes directly on a running product.

HOW LLMS ACTUALLY WORK

Modern large language models are mostly built by stacking transformer blocks over and over - the differences come from what each one was trained on, the scale and configuration choices, and the post-training done on top.

FIVE LABS, FIVE MINDS: BUILDING A MULTI-MODEL FINANCE DRAMA ON SMALL MODELS

The hackathon's Thousand Token Wood v2 game uses small models from OpenAI, NVIDIA, OpenBMB, and Qwen, creating diverse agent behaviors.

Save your spot

CHATGPT DREAMING V3

OpenAI introduced a new memory synthesis system for ChatGPT designed to improve freshness, continuity, and relevance over longer time horizons. The update began rolling out to Plus and Pro users in the US, with broader availability planned later.

A NEW "CLAUDE-OCEANUS-V1-P" HAS BEEN MADE AVAILABLE TO RED TEAMS

Anthropic appears to be gearing up for the public launch of a new version of Mythos that is better than Mythos Preview. A checkpoint of the model, codenamed Oceanus, was recently made available to red teamers. These programs typically begin a week before a wider launch. The program was apparently paused due to an individual in the program reselling the model via a Chinese API proxy. It is unknown whether this will impact the launch date.

WHEN AI BUILDS ITSELF

Anthropic is expediting AI development by enabling AI systems to autonomously design and develop successors, a concept known as recursive self-improvement. Internal benchmarks show AI-driven processes allow typical engineers to ship eight times more code than in previous years.

HOW WE MADE CONTINUOUS TRACE INTELLIGENCE POSSIBLE AT SCALE

Braintrust founder Ankur Goyal lays out Topics, the intelligence layer for analyzing production agent traces at scale where million-token traces with hundreds of spans break every standard NLP tool that expects uniform document shapes. Inspired by Anthropic's Clio paper, the pipeline runs preprocess to facet to embed to cluster to name to classify, with the LLM summary doing the one job that makes the rest tractable since the raw trace never has to fit in an embedding model's context window.

QWEN-IMAGE-FLASH

A study of few-step distillation for Qwen-Image-2.0 found that data composition, teacher guidance, and task mixture strongly affected student model performance.

Request early access to Crusoe Serverless Fine-Tuning

DEFENDING CODE REFERENCE HARNESS

This repository contains a reference implementation for autonomous vulnerability discovery and remediation with Claude. It can be used to build custom vulnerability pipelines based on general best practices. Anthropic offers a managed option that can find and fix vulnerabilities across multiple projects.

NEMOTRON 3.5 CONTENT SAFETY

NVIDIA released Nemotron 3.5 Content Safety, a unified model for multimodal, multilingual, and customizable enterprise safety enforcement. It supported auditable reasoning and was designed to fit into production moderation pipelines.

OLLAMA MODEL TESTER

Ollama Model Tester is a CLI tool for comparing local Ollama models by running the same prompt multiple times and saving responses for easy comparison.

APPLE'S MESSAGES APP ON IPHONE NOW HAS A THIRD-PARTY AI AGENT

Apple approved the third-party AI service Poke for use in its iPhone Messages app. This integration allows users to chat with Poke directly in iMessage to perform various tasks. Some users report issues with response times, likely due to high demand.

ANTHROPIC SAYS 80% OF ITS NEW PRODUCTION CODE IS NOW AUTHORED BY CLAUDE — HOW YOUR ENTERPRISE CAN CATCH UP

Anthropic reported that 80% of its production code now comes from its AI model, Claude, leading to an 8x increase in code volume per engineer.