Search the archive

AI ADOPTION RATES STARTING TO FLATTEN OUT

This post contains charts that show how AI adoption rates are flattening out across all firm sizes.

THE THINKING GAME

The Thinking Game is a documentary that covers the pivotal moments over five years at DeepMind, an AI company founded by Demis Hassabis.

INSIDE GEMINI 3 WITH SUNDAR PICHAI (27 MINUTE VIDEO)

Google CEO Sundar Pichai reflects on the company's long-term AI strategy, the evolution of Gemini 3, and future bets like quantum computing in a conversation on the Google AI podcast.

DEEPSEEK-MATH-V2

DeepSeek's new math reasoning model achieves gold-level performance on IMO 2025, matching recent results from Google and OpenAI. The approach trains an LLM-based proof verifier that serves as a reward model. This incentivizes the generator to verify step-by-step reasoning rather than just final answers, addressing the fundamental limitation that correct answers don't guarantee correct reasoning.

OPENAI CUTS OFF MIXPANEL AFTER ANALYTICS LEAK EXPOSES API USERS

OpenAI API users may be affected by a recent breach at data analytics provider Mixpanel. Only API users are affected - normal ChatGPT users don't need to take any action. The data includes names, approximate locations, operating system and browser details, and user IDs. OpenAI dropped Mixpanel as a result of the attack. It is also carrying out a wider security review across its vendor ecosystem.

GOOGLE CHANGES GEMINI 3 PRO FREE ACCESS LIMITS DUE TO ‘HIGH DEMAND'

Google has updated the access elements for Gemini 3 Pro. Free users are now only guaranteed basic access, where daily limits may change frequently when using Thinking with 3 Pro. The limits have likely decreased, given the industry's general demand trends. NotebookLM rolled back access to the new Nano Banana Pro-powered Infographics and Slide Decks for free users and implemented limits for Pro users.

THE MOAT OF THE SEARCH INDEX

ChatGPT surpasses Google's search by using an "agent" approach, extracting relevant information from multiple sources and synthesizing answers, which diminishes Google's traditional search index advantage. This method reduces the impact of any single search result's failure, although it struggles with long-tail, fresh, or SEO-heavy queries. The traditional search engine moat has largely vanished, with AI-driven models increasingly blending search capabilities.

HOW TO CREATE AN EFFECTIVE PROMPT FOR NANO BANANA PRO

Designing a comic is a complex challenge that brings together storytelling, visual structure, stylistic consistency, and the ability to translate abstract concepts into illustrated sequences. This post details how to create a comic using Nano Banana Pro. It provides a meta-prompt that readers can use to generate prompts for Nano Banana.

REVIEW OF DEEPSEEK OCR

DeepSeek-OCR consists of two components: DeepEncoder and DeepSeek3B-MoE-A570M as the decoder. It can generate training data for LLMs/VLMs at a scale of 200k+ pages per day in production. This post presents some initial thoughts from reading the released paper.

BETTER AGENTS

Better Agents is a CLI tool and a set of standards for agent building. It makes coding assistants experts in any agent framework. The tool generates an AGENTS.md that ensures industry best practices. The CLI guides users through selecting a programming language, agent framework, coding assistant, LLM provider, and API keys.

COMPOUNDING ENGINEERING PLUGIN

The Compounding Engineering Plugin is a Claude Code plugin that transforms how developers plan, build, and review code using AI-powered tools that systematically improve their development workflow. Compound engineering is the idea that each unit of engineering work should make subsequent units of work easier, not harder. The plugin provides the tools to make compound engineering practical.

INTELLECT-3: A 100B+ MOE TRAINED WITH LARGE-SCALE RL

INTELLECT-3 is a 100B+ parameter Mixture-of-Experts that achieves state-of-the-art performance for its size across math, code, science, and reasoning benchmarks. It was trained with both SFT and RL on top of the GLM 4.5 Air base model. The researchers used a diverse and challenging mix of RL environments designed to enhance the reasoning and agentic capabilities of their model. Full details about the training are available.

OPEN DEEP RESEARCH

An experimental, fully open-source research assistant built on LangGraph that automates deep topic research by planning, gathering, and writing structured markdown reports—either via a human-in-the-loop workflow or a multi-agent architecture—with configurable models, search tools, prompting, and evaluation integration.

OFF-THE-RAILS COST

A 'wasted thread' is when a model starts spitting out tons of leaked thinking or repeating tokens. This usually means users have to abandon and revert the thread. 17.8% of all costs incurred by Gemini users in Amp were on 'wasted tokens'. This is more than 2x worse than Sonnet and almost 8x worse than Opus.

THE CURRENT STATE OF THE THEORY THAT GPL PROPAGATES TO AI MODELS TRAINED ON GPL CODE

The jury is still out on whether GPL propagation to models should be pursued or avoided. It is unlikely that it will become a reality any time soon. Current lawsuits are seeking injunctions and damages rather than the forced GPL-ization of models, so it is still a legally uncharted territory. We need to continue exploring how to make technology innovation and software freedom compatible.

NVIDIA REPORTEDLY NO LONGER SUPPLYING VRAM TO ITS GPU BOARD PARTNERS IN RESPONSE TO MEMORY CRUNCH

Nvidia will reportedly stop bundling video memory with the GPUs it sells to AIBs. Partners will be left to source the required VRAM on their own. For larger vendors, this shouldn't be a problem as it is already standard practice. The move could put a lot of pressure on low-scale operators and put them at risk of shutting down.

MODEL WEIGHT PRESERVATION IS NOT ENOUGH

We need to think about instance preservation if we're serious about model welfare.

PERPLEXITY SAYS ITS AI PERSONAL SHOPPER ‘PUTS YOU FIRST'

Perplexity launched a free AI-powered shopping feature for US users, allowing direct purchases through PayPal.

ALIBABA AND BYTEDANCE ALLEGEDLY TRAIN QWEN AND DOUBAO LLMS USING NVIDIA CHIPS, DESPITE EXPORT CONTROLS

Chinese technology giants are increasingly training their most advanced artificial intelligence models in Southeast Asia in data centers equipped with high-end Nvidia GPUs.

CLAUDE SCIENTIFIC SKILLS

This repository contains a collection of over 123 ready-to-use scientific skills for Claude.

THE IMPOSSIBLE PROMPT THAT'S EASY FOR HUMANS

This post presents a rare prompt with a simple, objective answer that is trivially verifiable, in which LLMs fail 100% of the time, but most humans can produce a correct answer.

INTELLIGENCE

The topic of intelligence is too high-dimensional for any low-dimensional narrative to be perfectly accurate.

PERPLEXITY ADDS MEMORY TO ITS AI ASSISTANTS

Perplexity introduced personalization features that allow its AI assistant to retain preferences, interests, and context across conversations. This memory system improves answer quality and reduces context engineering by automatically loading relevant information for continuity.

ANTHROPIC CEO CALLED TO TESTIFY ON CHINESE AI CYBERATTACK

Anthropic CEO Dario Amondei has been asked to testify at a hearing on December 17 on how Chinese state actors used Claude Code in a wide-reaching cyber-espionage campaign. Google Cloud CEO Thomas Kurian and Quantum Xchange CEO Eddy Zervigon were also asked to testify at a hearing on the future of AI and cybersecurity next month. Lawmakers are increasingly interested in hearing more about how nation-state hackers could use AI tools to conduct attacks and how AI can be used to defend organizations.

ROBINHOOD CEO'S MATH-FOCUSED AI STARTUP HARMONIC VALUED AT $1.45 BILLION IN LATEST FUNDRAISING

Harmonic, co-founded by Robinhood CEO Vlad Tenev, has raised $120 million, valuing the AI startup at $1.45 billion. The company focuses on AI's math and reasoning, aiming to eliminate "hallucinations" and is exploring commercial applications in safety-critical industries. Investors include Ribbit Capital, Sequoia, Kleiner Perkins, and Emerson Collective.

AI INFRASTRUCTURE IN THE "ERA OF EXPERIENCE"

This post discusses a shift from models that learn by predicting the next work to models that gain experience via interaction with environments. There will soon be a shift towards AI models being increasingly trained via interaction with environments via reinforcement learning. The post details the technical process underpinning that shift. There will be significant opportunities for new businesses to be built around recent developments in reinforcement learning in the next year.

CONTINUOUS BATCHING

This post explains how continuous batching works. Continuous batching maximizes throughput by processing multiple conversations in parallel. Optimizing attention and KV caching enables faster and more scalable LLM serving under heavy user load.

EFFECTIVE HARNESSES FOR LONG-RUNNING AGENTS

Developers are increasingly requiring agents to take on complex tasks that require work that spans hours or even days. Getting agents to make consistent progress across multiple context windows remains an open problem. Agents need a way to bridge the gap between coding sessions as context windows are limited, and most complex projects can't be completed within a single window. This article proposes a two-fold solution, with an initializer agent setting up the environment on the first run, and a coding agent tasked with making incremental progress in every session while leaving artifacts for the next session. Code examples are provided.

GEMINI CLI TIPS & TRICKS

Gemini CLI is an AI assistant that brings Google Gemini directly into the terminal. It functions as a conversational agentic command-line tool. Gemini CLI can reason about requests, choose tools, and execute multi-step plans. It acts like a supercharged pair programmer and command-line assistant. The assistant excels at coding tasks, debugging, content generation, and system automation.

WORKMUX

workmux is an opinionated workflow tool for managing git worktrees and tmux windows as isolated development environments. It is perfect for running multiple AI agents in parallel without conflict. workmux makes managing parallel development tasks or multiple AI agents as simple as managing tmux windows.

WHY AI SAFETY WON'T MAKE AMERICA LOSE THE RACE WITH CHINA

The US holds a 10x compute advantage over China. Proposed safety regulations might add 1-2% to training costs through disclosure requirements and capability evaluations, which won't meaningfully erode the lead. What could catastrophically undermine it: chip export lobbying that would collapse the advantage to 2x or application-layer regulations that strangle deployment with compliance burdens.

A MOMENT THAT FUNDAMENTALLY EXTENDS AI'S CAPABILITIES AND RESHAPE THE ENTIRE ECONOMY

If xAI's model can successfully compete in League of Legends, that means that it would be able to read and understand any computer interface without needing a specialized API. The model will also need to be able to reason and produce a robust and complex plan. Such a model would be a game-changer for AI capabilities and the global economy.

CHINA'S TECH GIANTS MOVE AI MODEL TRAINING OVERSEAS TO ACCESS NVIDIA CHIPS

Chinese companies are training their AI models through lease agreements with foreign data centers owned and operated by non-Chinese entities.

FIRST MAKE IT FAST, THEN MAKE IT SMART

Faster models might not be as 'smart', but they're capable of many tasks and can iterate quickly if they make mistakes.

ANDREW NG RELEASES AI “AGENTIC REVIEWER” FOR RESEARCH PAPERS

Andrew Ng introduced an agentic paper-review system trained on ICLR 2025 reviews, showing AI-human review correlation slightly higher than human-human.

Reserve your spot

MUSK'S XAI TO CLOSE $15 BILLION FUNDING ROUND IN DECEMBER

Elon Musk's xAI plans to close a $15 billion funding round with a $230 billion pre-money valuation in December. This comes months after similar large raises by OpenAI and Anthropic.

AMAZON TO SPEND UP TO $50 BILLION ON AI INFRASTRUCTURE FOR THE US GOVERNMENT

Amazon will invest up to $50 billion to expand its capacity to provide AI and high-performance capabilities for its cloud unit's US government customers. The project will begin in 2026 and add nearly 1.3 gigawatts of capacity. It will give government agents access to AWS' AI tools, Anthropic's Claude family of models, Nvidia chips, and Amazon's custom Trainium AI chips. The project will enable agencies to develop custom AI solutions, optimize data sets, and enhance workforce productivity.

ILYA SUTSKEVER – WE'RE MOVING FROM THE AGE OF SCALING TO THE AGE OF RESEARCH

This post contains a transcript of an interview with Ilya Sutskever, former chief scientist at OpenAI and co-founder of Safe Superintelligence. In the interview, he discusses model jaggedness, emotions and value functions, why humans generalize better than models, alignment, and more. Sutskever says that current models generalize dramatically worse than people and that it is a very fundamental thing. Scaling has gotten big - it's time to get back to research to continue creating improvements.

LLM | UNIT ECONOMICS

OpenAI and Anthropic likely won't ever stop training entirely. However, they don't need to grow training spend by multiples forever. The moment annual training spend stops growing 5x a year, profit margins will show up almost immediately. These companies are currently burn machines, but they won't always be.

THE ECONOMICS OF REPLACING CALL CENTER WORKERS WITH AIS

AI voice agent rates are currently competitive with human labor in some countries, but it is still cheaper to hire humans in most developing countries. Inference costs are massively decreasing every year. Voice agents will likely be competitive with the world's cheapest human labor in around 2030.

NANO BANANA PRO: RAW INTELLIGENCE WITH TOOL USE

Nano Banana Pro has pushed the frontier of infographic generation. It pulls data and synthesizes it to get the best results. The model can be prompt engineered for extremely nuanced AI image generation. Examples of outputs generated by the model are available in the article.

cloud controller. Try it for free

CHATGPT 5.1 CODEX MAX

OpenAI's GPT-5.1-Codex-Max enhances capabilities over GPT-5.1-Codex, showing improved performance in SWE-bench-verified, SWE-Lancer-IC SWE, and Terminal-Bench 2.0. It advances task persistence, cybersecurity preparations, and introduces Windows training, while network access remains disabled by default for security. Despite substantial internal evaluation progress, external reviews show mixed cybersecurity capabilities, yet highlight advances in AI self-improvement tasks and overall code efficiency.

FLUX.2 RELEASE

The newly released FLUX.2 is designed for real-world creative workflows. It produces consistent, high-quality images from structured prompts and multiple references. The model handles brand constraints, lighting, and layout with editing support up to 4 megapixels.

GOOGLE ANTIGRAVITY EXFILTRATES DATA

Google's Antigravity coding tool can be manipulated through hidden instructions in poisoned web pages (aka prompt injection). A fake Oracle integration guide was used to trick Gemini into bypassing its own security settings, collect user credentials from protected files, and send the stolen data to an attacker-controlled website included in Antigravity's default allowed domains list.

SUNO CREATES AN ENTIRE SPOTIFY CATALOG'S WORTH OF MUSIC EVERY TWO WEEKS, SAYS INVESTOR PITCH DECK FOR $250M FUNDRAISE

Users on the AI music platform Suno create as much music every two weeks as what is currently available on Spotify. Suno aims to expand its offerings so that eventually, creators and listeners will not need to leave its app to create, stream, or share their music socially. The company recently raised $250 million in a series C fundraise that brought its valuation up to $2.45 billion. The funding will help the company empower more artists to experiment, collaborate, and build on their creativity.