Search the archive

BAIDU JUST DROPPED AN OPEN-SOURCE MULTIMODAL AI THAT IT CLAIMS BEATS GPT-5 AND GEMINI

ERNIE-4.5-VL-28B-A3B-Thinking is a 28-billion-parameter multimodal AI model that only activates 3 billion parameters during operation. It has a sophisticated routing architecture that allows it to match or exceed the performance of much larger competing systems on several categories of tasks. The model's most distinctive feature is 'Thinking with Images', a capability that allows the AI to dynamically zoom in and out of images to examine details. The feature dramatically elevates the model's ability to handle long-tail visual knowledge. The model has been released under the permissive Apache 2.0 license, which allows unrestricted commercial use.

INSIDE MICROSOFT'S NEW AI ‘SUPER FACTORY'

Microsoft aims to double its total data-center footprint over the next two years. Its newest AI super factory sits within a set of two-story structures in Atlanta. The site will contain hundreds of thousands of Nvidia GPUs and dedicated high-speed connections to other locations. Part of Microsoft's Fairwater network of AI centers, it spans more than 1 million square feet across 85 acres.

FEED THE BEAST: AI EATS SOFTWARE

Software interfaces force their way of thinking on users. AI allows you to bypass all these tools - you just have to feed it the right data, and it can complete complex tasks. Doing this right can result in you getting rid of almost all of your productivity tools. This article shows readers how to get started, along with examples of how AI can completely change a workflow.

the 2025 Production AI Stack Report by Temporal

ELEVENLABS' NEW AI MARKETPLACE LETS BRANDS USE FAMOUS VOICES FOR ADS

The platform connects brands with voice rights owners, ensuring permission and fair compensation, and includes the voices of Michael Caine, Mark Twain, and Maya Angelou.

THERE'S A LOT OF BUZZ ABOUT CODE SANDBOXES

Coding agents are very useful, but building a business around them is quite hard.

PROJECT FETCH: CAN CLAUDE TRAIN A ROBOT DOG?

Anthropic divided eight of its researchers into two teams, one with Claude access, one without, and tasked them with programming quadruped robots to fetch beach balls.

OPENAI PUSHES BACK ON NYT PRIVACY DEMANDS

OpenAI is resisting legal efforts by the New York Times to obtain millions of private ChatGPT conversations, citing user privacy and data security risks.

Access the full prompt library here

Not sure what you can do with Notion Agents? Here are 100 ideas

to get you started...

OPENAI READIES CHATGPT GROUP CHATS WITH CUSTOM CONTROLS

A new Group Chats feature for ChatGPT, currently in development, enables multiple users to join a shared conversation and interact both with each other and with the AI in a single chat feed. Groups will be able to customize the system prompt for discussions and manage when the AI should respond. The feature will likely be rolling out as part of the upcoming seasonal update cycle.

INTRODUCING SCRIBE V2 REALTIME

Scribe v2 Realtime is a low-latency Speech to Text model that delivers live transcription in under 150 milliseconds. It supports over 90 languages and is specifically built for agentic use cases. The model achieves 93.5% accuracy across 30 commonly used European and Asian languages. Scribe v2 Realtime can be accessed through ElevenLabs' API or directly within ElevenLabs Agents.

YANN LECUN REPORTEDLY PLANNING TO LEAVE META

Meta's chief AI scientist, Yann LeCun, is reportedly preparing to leave the company to start his own venture focused on world models. His departure comes amid a broader AI reorganization at Meta, including the formation of Meta Superintelligence Labs, and internal tensions following shifts in AI development strategy.

BYTEDANCE UNVEILS CHINA'S MOST AFFORDABLE AI CODING AGENT AT JUST US$1.30 A MONTH

The Doubao-Seed-Code model is a coding agent priced at 40 yuan a month, with a special launch price of 9.9 yuan (US$1.30) for the first month. The new model set a state-of-the-art record on the SWE-Bench Verified test. It supports popular development tools like veCLI, Cursor, and Cline, and is compatible with Anthropic's API. The model can process up to 256,000 words per query and can handle complex codebases.

BUILD TIMES FOR GIGAWATT-SCALE DATA CENTERS CAN BE 2 YEARS OR LESS

Gigawatt-scale AI data centers are massive undertakings that require extensive permitting, construction, and power infrastructure. Many hyperscalers have concrete plans to build data centers at this scale in two years or less. The time from starting construction to achieving one gigawatt of total facility power ranges from 1 to 3.6 years. The first gigawatt-scale datacenters are expected to come online in early 2026.

HOW TO TRAIN AN LLM: PART 1

This post walks through how a researcher set up basic pre-training infrastructure and trained a 1B Llama 3-style model on 8xH100s. The resulting model is far from state-of-the-art, but it provides a clear working implementation that can be abstracted further.

RL ENVIRONMENTS AND THE HIERARCHY OF AGENTIC CAPABILITIES

Researchers dropped nine frontier AI models into a simulated workplace environment and gave them 150 jobs to do. Most of the AIs were barely coherent, and even the best lacked common sense. The experiment showed how models must acquire more fundamental capabilities before we can even begin to discuss how well they perform common-sense reasoning in real environments. Finding out whether common sense reasoning is a set of identifiable, trainable sub-skills or an emergent property of large-scale real-world training will help shape the next stage of AI development.

WHY SUDOKU VARIANTS REMAIN A GRAND CHALLENGE IN AI REASONING

Modified versions of Sudoku test an AI's creativity and ability to strategize around rulesets that don't appear in its training data, unlike fixed-rule games like Chess or Go. GPT-5 is the first LLM to solve a 9x9 variant by simultaneously coordinating four constraint types (standard rules, region sum lines, XV pairs, and Roman numeral cages), but it still only succeeded 20% of the time.

FROM WORDS TO WORLDS: SPATIAL INTELLIGENCE IS AI'S NEXT FRONTIER

Current LLMs are wordsmiths in the dark—eloquent but inexperienced. They lack a core cognitive function of humans: spatial intelligence. "World models" solve this by generating geometrically and physically consistent simulated worlds, processing multimodal inputs like images and gestures, and predicting next states based on actions. If successful, the technology would be a step change for robotics, video models, and drug discovery.

report

HERE'S WHAT'S NEXT IN AGENTIC CODING

The speed of development in the agentic coding space has been eye-wateringly meteoric. There have been eleven paradigm shifts in ten months. Next year will be when polished harnesses meet new age models. Context management is everything. The degree to which next year's tools will translate intent into working code will be mind-blowing.

NESTED LEARNING REPRODUCTION

This repository contains a high-fidelity reproduction of Google's Nested Learning architecture. It is fully open-source and uv-managed. A copy of Google DeepMind's original Nested Learning Paper is included in the repository.

INTRODUCTION TO AGENTS (3 HOUR READ)

AI is changing from models that excel at passive, discrete tasks to a new class of software capable of autonomous problem-solving and task execution. The new frontier is built around AI agents. This page contains a whitepaper written by Google researchers that introduces readers to AI agents. AI agents are complete applications that plan and take actions to achieve goals. They combine models' ability to reason with the practical ability to act.

THREE AI MEGADEALS ARE BREAKING NEW GROUND ON WALL STREET

Some of the biggest AI infrastructure deals that Meta, OpenAI, and xAI are part of involve innovative, and in some cases risky, funding schemes. These tech titans are offering sweeteners in their deals because they need to offload risk as the cost of the AI arms race soars. While banks and fund managers are writing big checks for now, many are worried about how the complicated deals will perform when the AI frenzy calms down. This article takes a close look at some of the biggest AI megadeals so far.

ANTHROPIC LAUNCHES USE CASE LIBRARY

Anthropic released a searchable library of 45+ practical use cases demonstrating Claude's capabilities across professional work, marketing, education, research, and personal tasks, each with specific prompts and step-by-step guidance.

scale. Read the ungated blog

SOFTBANK'S NVIDIA SALE RATTLES MARKET, RAISES QUESTIONS

For the second time in the last decade, SoftBank sold all of its remaining Nvidia stake worth $5.8 billion.

AMD REVEALS NEW ROADMAP FOR ITS RYZEN CPUS, TEASING ZEN 7 AS THE TRUE "NEXT-GENERATION" LEAP

Zen 6 will leverage TSMC's flagship 2nm process nodes.

EMOTIONAL AGENTS

We know AI will eventually develop the ability to be emotional, but how we will use that emotionality is still unknown.

HAS GOOGLE QUIETLY SOLVED TWO OF AI'S OLDEST PROBLEMS?

Google's newest model appears to be able to read difficult handwritten historical documents just as well as expert humans and also analyze them in deep and nuanced ways.

Plans start at just $49/month. Get a demo

XAI WORKS ON GROK CODE REMOTE TO RIVAL OPENAI'S CODEX

xAI will host a hackathon on December 6-7. Participants will receive early access to experimental models. The company is working on a feature called Grok Code Remote to align with industry trends towards cloud-based code execution. It will allow users to run code in remote environments directly from the web.

ANTHROPIC IS ON TRACK TO TURN A PROFIT MUCH FASTER THAN OPENAI

OpenAI forecasts that it won't turn a profit until 2030, while Anthropic expects to break even for the first time in 2027. OpenAI is also expected to burn roughly 14 times as much cash as Anthropic before turning a profit. The two most valuable AI startups are taking vastly different approaches to growth. OpenAI is investing heavily in chips and data centers. Anthropic is taking a more cautious approach, making sure that costs grow more in line with revenue.

YOU CAN NOW TEST UPCOMING GPT-5.1 ON OPENROUTER

OpenAI's latest code updates reference a general availability date of November 24 for the GPT-5.1 family for Enterprise accounts. Plus and Pro users may gain access to the models even earlier, possibly within the coming week. GPT-5.1 Thinking may currently be testing under the name Polaris Alpha on platforms such as OpenRouter. Polaris Alpha has a 256k context window and quite good performance.

AI DIFFUSION REPORT: WHERE AI IS MOST USED, DEVELOPED, AND BUILT

AI has reached 1.2 billion users in under three years, marking the fastest technology adoption in human history. However, the Global North's adoption rate is nearly double that of the Global South. The next 1–2 billion users may come easily, but adoption is expected to plateau afterward because the remaining ~4 billion people lack electricity or internet access. Low-resource languages create an additional barrier: half the web is in English, even though only 5% of people speak it natively.

THE 4.5 TRILLION DOLLAR ELEPHANT IN THE ROOM

Nvidia, currently valued at $4.5 trillion, is the most valuable company on Earth and the dominant supplier of AI chips. A pattern of fear around the company is impacting policy discussion. There are claims that the major AI companies disagree with Nvidia's policy stances, but are unwilling to speak out because Nvidia has the power to undermine their business. This article looks at the evidence of such claims.

report

OMNILINGUAL ASR

Omnilingual ASR is a speech recognition system that supports over 1,600 languages, with hundreds never before covered by other ASR technology. New languages can be added with just a few paired examples without specialized expertise or large datasets. Omnilingual ASR is designed for broad accessibility - it is aimed at making speech technology more inclusive and adaptable for communities and researchers worldwide.

EGOCENTRIC-10K (HUGGING FACE REPO)

Egocentric-10K is a hand visibility and active manipulation data set with 10,000 hours of video. It contains 192,900 video clips of 2,138 workers completing tasks in real factories. A video showing examples of some of the clips is available in the repository.

AMA WITH MOONSHOT AI, THE OPEN-SOURCE FRONTIER LAB BEHIND KIMI K2 THINKING MODEL (REDDIT THREAD)

Moonshot AI's Kimi K2 is a Mixture-of-Experts model with 32 billion activated parameters and 1 trillion total parameters. It achieves state-of-the-art performance in frontier knowledge, math, and coding among non-thinking models and also has agentic abilities. Kimi-K2-Base and Kimi-K2-Instruct were recently open sourced. This thread contains a discussion with the team at Moonshot AI.

EVERY AI APPLICATION STARTUP IS LIKELY TO BE CRUSHED BY RAPID EXPANSION OF THE FOUNDATIONAL MODEL PROVIDERS

There is now almost no time to build a company and scale it. There are two ways an AI application founder can make money: make a flash-in-the-pan app that generates a ton of cash over a short period, or make a good enough app to get acquired by one of the big players. The situation is highly unstable. The best way to play is to find an application niche in a highly specialized field with extremely unique and specific data barriers.

NEW LEAKED IMAGE SAMPLES BY NANO BANANA 2

Nano Banana 2 access was briefly leaked, and image samples were widely circulated across social media. The model appears capable of remastering low-quality images and following detailed text prompts. There is still no confirmation of the release date, but a launch is likely to be imminent. The leaked samples are available in the article.

Subscribe now.