Search the archive

8-BIT ROTATIONAL QUANTIZATION FOR VECTOR SEARCH

THE PARALLELISM MESH ZOO

There are many parallelization strategies available for scaling training runs to work on more GPUs. This post discusses parallelization strategies in a more schematic way by focusing only on how they affect the device mesh. The device mesh is an abstraction used by PyTorch and JAX that takes GPUs and organizes them into an N-D tensor that expresses how the devices communicate with each other. Being able to explain why a device mesh is set up the way it is is a good check for whether you understand how the parallelization strategy works.

CONTEXT ENGINEERING SERIES EXPLORES BUILDING BETTER AGENTIC RAG SYSTEMS

Coding agents like Claude Code and Cursor “context engineer” by designing tool portfolios, slash commands, and subagent architectures to let AI systems discover the information they need to perform best. Structured tool responses aren't just applicable for coding tools. The same interaction patterns can be applied across industries, but require greater focus on avoiding “context pollution,” where misleading or irrelevant information degrades capabilities.

THE CURIOUS CASE OF FLUNKING MY ANTHROPIC INTERVIEW (AGAIN)

This developer, a big fan of Claude Code, failed an interview at Anthropic for a Developer Relations role despite trying to score extra credit by making a viral post about their positive experiences with Claude and receiving a glowing recommendation from a current employee. The interview involved a secret take-home assignment.

MY AI PREDICTIONS FOR 2027

It's a good idea to be skeptical of fancy academic reasoning. Science doesn't work because of probability distributions, it works because people go out to gather evidence to find out what's true about reality. Statistics is part of finding out the truth, but in the case of AI, there is not enough data to make predictions about what will happen.

Subscribe now.

CHARACTER.AI LOGGING SYSTEM

Character.AI centralized its fragmented logging pipeline to gain real-time observability, cut query latency, and enable better debugging with smart sampling, live tailing, and noise reduction tools.

XAI SUES AN EX-EMPLOYEE FOR ALLEGEDLY STEALING TRADE SECRETS ABOUT GROK

xAI claims that former employee Xuechen Li stole its confidential info and trade secrets before joining OpenAI.

META TO LAUNCH CALIFORNIA SUPER PAC BACKING PRO-AI CANDIDATES

Meta, aiming to influence AI regulation, is launching a California super PAC to support pro-AI candidates.

A BEHIND-THE-SCENES PODCAST COVERING PIXEL 10 AI FEATURES

The Pixel 10 introduces AI features like Conversational Editing and Camera Coach, backed by the Tensor G5 chip and a refined camera system.

BILLIONAIRE AMBANI TAPS GOOGLE, META TO BUILD INDIA'S AI BACKBONE

India's largest conglomerate, Reliance Industries, is betting big on AI infrastructure with a new subsidiary that's secured $100 million from Meta and dedicated cloud infrastructure from Google.

👉 WATCH IT ON-DEMAND

MICROSOFT RELEASES TWO IN-HOUSE MODELS

Microsoft AI has launched MAI-Voice-1, a speech generation model that can produce a minute of audio in under a second, and MAI-1-preview, a mixture-of-experts foundation model trained on ~15,000 NVIDIA H100 GPUs. The technical capabilities of the models are unclear, but they indicate Microsoft's intention to reduce its reliance on OpenAI's model to power its AI offerings.

META PUSHES TO RELEASE NEW LLAMA MODEL BEFORE 2026

Meta plans to release its next version of Llama before the end of the year. The model, internally known as Llama 4.X or Llama 4.5, will be one of the first projects from Meta Superintelligence Labs. Meta recently restructured the unit into four groups, with teams focused on training, research, product, and infrastructure. Meta first announced the Llama 4 models in April.

GROK CODE FAST 1

xAI launched a specialized coding model optimized for speed and cost rather than general intelligence, priced at $0.20 per million input tokens versus $18+ from competitors. The model is temporarily free through partnerships with GitHub Copilot, Cursor, and other coding platforms.

MASS INTELLIGENCE

Over a billion people now use AI chatbots regularly. Recent releases have given even unpaid users access to the most powerful models like GPT-5 or nano banana. Intelligence is no longer scarce or expensive, and will only get cheaper and more abundant, so institutions of every size face the chaos of a billion people wielding AI that outperforms humans at intellectual tasks

BUILDING YOUR OWN CLI CODING AGENT WITH PYDANTIC-AI

CLI coding agents can read code, run tests, and update codebases. This article walks readers through how to create their own coding agent by assembling open source tools and using specific development standards for testing, documentation production, code reasoning, and file system operations. While commercial tools are impressive, they're built for general use cases. Learning how to build a coding agent can provide insight into how systems work and the quality of available tooling.

AI access risks.

INTRODUCING GPT-REALTIME AND REALTIME API UPDATES FOR PRODUCTION VOICE AGENTS

OpenAI's Realtime API is now generally available. It now supports MCP servers, image inputs, and phone calling, making voice agents more capable. The underlying model, gpt-realtime, is OpenAI's most advanced speech-to-speech model yet, showing improvements in following complex instructions, calling tools with precision, and producing speech that sounds natural and expressive. The Realtime API processes and generates audio directly through a single model and API, which reduces latency and preserves nuance in speech, producing more natural, expressive responses.

OLLM

oLLM is a lightweight Python library for large-context LLM inference that enables users to run models like Llama-3.1-8B-Instruct on 100k context using around a consumer GPU with 8GB VRAM. It doesn't use any quantization, only fp16 precision.

R-ZERO: SELF-EVOLVING REASONING LLM FROM ZERO DATA

Self-evolving large language models are models that can autonomously generate, refine, and learn from their own experiences. Existing methods for training such models still rely heavily on vast human-curated tasks and labels. This poses a fundamental bottleneck to advancing AI systems toward capabilities beyond human intelligence. R-Zero is a fully autonomous framework that generates its own training data from scratch. It substantially improves reasoning capability across different backbone models.

XCODE 26 BETA 7 ADDS SUPPORT FOR CLAUDE SONNET AND CHATGPT 5

Claude Sonnet 4 is now available in Xcode. It can be accessed in the Intelligence setting panel. An existing paid Claude account is required to use the integration. Users can now start a new conversation with either GPT-4.1 or GPT-5, with GPT-5 set as the default, when using ChatGPT in Xcode.

THE 100 MOST INFLUENTIAL PEOPLE IN AI

TIME's third annual list of the 100 most influential people in AI includes familiar leaders like Sam Altman, Elon Musk, and Jensen Huang alongside newcomers like DeepSeek CEO Liang Wenfeng and Pope Leo XIV.

chunk (43%) are never using it. Find out why

ANTHROPIC USERS FACE A NEW CHOICE – OPT OUT OR SHARE YOUR DATA FOR AI TRAINING

Anthropic reversed its no-training policy for consumer data, now requiring users to actively choose whether Claude can learn from their conversations by September 28.

NEW AI-POWERED LIVE TRANSLATION AND LANGUAGE LEARNING TOOLS IN GOOGLE TRANSLATE

Google Translate now offers real-time conversation translation in over 70 languages, enhancing live interactions with advanced AI voice and speech recognition models.

GOVERNANCE ENFORCEMENT IS WEAK_._

Security leaders also say they're struggling with:

Read the research and learn how to mitigate AI access risks.

NEW CODEX FEATURES POWERED BY GPT-5

Codex has rolled out updates, including a new extension for editors like Cursor and VSCode, an enhanced CLI for local use, and seamless management of local and cloud tasks. GitHub code reviews can now be driven by Codex, all integrated into existing ChatGPT plans and backed by GPT-5.

THE TOP 100 [GEN AI] CONSUMER APPS

The Gen AI ecosystem is starting to stabilize. There were only 11 new names on Andreessen Horowitz's most recent Top 100 Gen AI Consumer Apps list, compared to 17 newcomers in the March ranking. There were significantly more mobile newcomers as app stores have cracked down on ChatGPT copycats, opening up room for more original apps. ChatGPT is still the leading general assistant, but Google, Grok, and Meta are closing the gap. More on the top Gen AI consumer apps is available in the article.

NVIDIA REPORTS RECORD SALES AS THE AI BOOM CONTINUES

Nvidia, the world's most valuable company, reported $46.7 billion in revenue this quarter, a 56% increase compared to the same period last year. AI-dominated data center business largely fuelled that growth.

BUILDING AGENTS FOR SMALL LANGUAGE MODELS: A DEEP DIVE INTO LIGHTWEIGHT AI

Small language models are models ranging from 270M to 32B parameters that run efficiently on CPUs or modest GPUs. They offer immense potential: privacy through local deployment, predictable costs, and full control thanks to open weights. They also present unique challenges that require a shift in how agent architectures are designed. This article reviews lessons learned from hands-on experimentation, debugging, and optimizing of inference pipelines for small language models.

MALLEABLE SOFTWARE WILL EAT THE SAAS WORLD

The winners in the AI era will be the tools that adapt to users. Large language models shift the focus from designing the solution to defining the problem. They can handle the 'how' aspects of the problem, so you just need to describe what you want in plain language. The future belongs to software that bends without breaking.

HOW CLOUDFLARE RUNS MORE AI MODELS ON FEWER GPUS: A TECHNICAL DEEP-DIVE

Omni is an internal platform at Cloudflare built for running and managing AI models on its edge nodes. It can spawn and manage multiple models on a single machine and GPU using lightweight isolation. Omni makes it easy and efficient to run many small and/or low-volume models. This article looks at how Cloudflare uses Omni to run more models on every node in its network, improving model availability, minimizing latency, and reducing power consumed by idle GPUs.

STOP “VIBE TESTING” YOUR LLMS. IT'S TIME FOR REAL EVALS

Stax is an experimental developer tool designed to streamline the LLM evaluation lifecycle. Evals give developers clear metrics to help understand what's actually better, so developers don't have to waste hours 'vibe testing' every time they try a new model or tweak a prompt. Stax empowers developers to rigorously test their AI stacks to help make data-driven decisions. It allows developers to easily define their own criteria and build a custom autorater.

ENVIRONMENTS HUB: A COMMUNITY HUB TO SCALE RL TO OPEN AGI

A new open-source platform is providing dozens of reinforcement learning environments to support “open AGI” and inviting contributors to add more.

OH GOODY, THE 'FIRST KNOWN AI-POWERED RANSOMWARE' HAS BEEN DISCOVERED, AND IT 'MAY EXFILTRATE DATA, ENCRYPT IT, OR POTENTIALLY DESTROY IT'

PromptLock is a newly discovered proof-of-concept ransomware that uses OpenAI's gpt-oss-20b model to generate code to exfiltrate data, encrypt it, and potentially destroy it. It uses the Ollama API to generate malicious Lua scripts on the fly, which it then executes. The ransomware runs locally, which allows it to avoid detection more easily as it doesn't use any online resources.

OPENAI AND ANTHROPIC SHARE SAFETY ACCESS

OpenAI and Anthropic allowed each other internal API access for joint AI safety testing, aiming to uncover blind spots in their model evaluations.

Explore OpenSearch

GOOGLE IRONWOOD TPU SWINGS FOR REASONING MODEL LEADERSHIP AT HOT CHIPS 2025

Ironwood is the first Google TPU explicitly designed for large-scale AI inference.

CRITIQUES OF WORLD MODELS

The primary goal of a world model is to simulate all actionable possibilities of the real world for purposeful reasoning and acting.

AN UNUSUAL CONSEQUENCE OF AI CODING

This developer found that they stopped listening to music when they replaced their old coding process with prompt-crafting.