Search the archive

MIXTURE-OF-RECURSIONS

DeepMind researchers have introduced Mixture-of-Recursions (MoR), a unified architecture that combines parameter sharing and adaptive computation by using a recursive transformer with lightweight routing.

THE SHIFT FROM TEXT TO DYNAMIC AI EXPERIENCES

OpenAI's new CEO of Applications outlines how ChatGPT is moving beyond chat toward a "fully generative UI" that surfaces the right components based on what you're doing: a dedicated image studio, inline writing blocks, quick visual answers for measurements or sports scores, and tappable highlights that pull up additional context.

META AI GLASSES WITH AUDIO FOCUS

Meta updated its smartglasses with a feature to isolate conversations in noisy environments and a new Spotify integration that plays music based on what you're looking at. The updates roll out first on Ray-Ban Meta and Oakley Meta HSTN glasses in the US and Canada.

offer (new customers only)

DATABRICKS RAISES CAPITAL AT $134 BILLION VALUATION IN LATEST FUNDING ROUND

Databricks raised $4B at a $134B valuation, up 34% from three months ago, with CEO Ali Ghodsi saying he wouldn't rule out a 2026 IPO.

INSIDE NVIDIA NEMOTRON 3

Nemotron 3 is an open hybrid Mamba-Transformer MoE architecture designed for agentic AI.

BENCHMARKING AI ON SCIENTIFIC REASONING

FrontierScience is a benchmark designed to test AI performance on expert-level reasoning tasks in physics, chemistry, and biology.

ANTHROPIC PREPARING NEW AGENTIC TASKS MODE FOR CLAUDE

Anthropic is testing a new interface for tasks in Claude's Agent mode. It is also introducing new modes for research, analysis, writing, and building. The updated interface introduces a toggle that allows users to switch between classic chat and agent modes. Screenshots of the new interface are available in the article.

NVIDIA DEBUTS NEMOTRON 3 FAMILY OF OPEN MODELS

Nvidia released Nemotron 3 Nano (30B parameters, 3B active) with Super (100B) and Ultra (500B) coming in early 2026, with Nano's benchmark scores rivaling or exceeding closed-source rivals. Nvidia is publishing training data and releasing libraries for agent customization in what appears to be an attempt to undermine OpenAI, Google, and Anthropic, which are increasingly developing their own chips instead of using Nvidia.

DISNEY-OPENAI LICENSING DEAL IS ONE-YEAR EXCLUSIVE

Disney's new partnership with OpenAI allows Sora to use over 200 characters from its IP catalog for one exclusive year. After that, Disney is free to partner with other AI platforms, hinting at broader content licensing strategies.

SELF IMPROVING AGENT WITH DYNAMIC CONTEXT AND CONTINUOUS LEARNING

This guide walks readers through how to build a self-improving Text-to-SQL agent using dynamic context and 'poor-man's continuous learning'. The agent answers questions by retrieving data from a knowledge base. It learns from successful runs by adding to the knowledge base, creating a self-improving feedback loop. A video that shows the agent in action is available at the end of the article.

GPT-5.2 IS FRONTIER ONLY FOR THE FRONTIER

GPT-5.2 is good at instruction following and a solid choice, but it isn't 'fun' to interact with. People strongly dislike its personality. The model is heavily constrained and censored. It is not likely to solve OpenAI's 'Code Red' problems. The company will likely try again in a month with GPT-5.3.

WHEN MACHINES PAY MACHINES: THE ECONOMICS OF AGENTIC AI

x402 is a protocol that embeds payments directly into HTTP, allowing any API call to include a payment. It has processed over 100 million payments across APIs, applications, and AI agents since launching in May this year. This article takes a look at the protocol and how internet-native payments are becoming standard infrastructure.

AI AGENTS ARE STARTING TO EAT SAAS

Basic SaaS providers now have thousands of new competitors: engineers with a spare Friday afternoon with an agent. Not all SaaS products will be affected, but the bar is going to be much higher for those that don't have a clear moat or proprietary knowledge. Companies with strong internal technical ability will be able to replace their SaaS spend, decreasing costs, while those who don't will see dramatically increasing costs as Saas providers try to recoup lost sales.

Read the blog

STRUCTURED OUTPUTS CREATE FALSE CONFIDENCE

Constrained decoding forces models to prioritize output conformance over output quality. This is because it restricts models to only consider certain tokens. Using structured output makes it harder for models to refuse requests, warn users about contradictory information, and inform users of the right approach when they ask it to use the wrong approach. Letting models respond in as natural a fashion as possible is the most effective way to get the highest quality response from them.

BRING YOUR RESEARCH TO LIFE WITH INTEGRATED VISUAL REPORTS FROM GEMINI DEEP RESEARCH

Google AI Ultra subscribers can now use Gemini Deep Research to generate rich visual reports complete with custom images, charts, and interactive simulations. The update can automatically illustrate data and create dynamic simulation models to forecast outcomes based on different variables. The feature combines deep analysis with dynamic visuals to transform dense data into tangible, easy-to-understand insights.

OLMO 3: A DEEP DIVE INTO THE FULLY-OPEN LLM

AI2 published the most comprehensive starting point for open LLM research: all checkpoints, training data, and code for OLMo 3. The walkthrough covers every stage from data mixing through the full SFT/DPO/RLVR post-training stack, including OlmoRL infrastructure that cuts RL training from 15 days to 6 days. RL with random rewards, which worked on Qwen, fails on OLMo 3.

200K TOKENS IS PLENTY

Claude Opus 4.5 is considered the best model for coding, and it only has a context window of roughly 200,000 tokens. While some people feel like that is not a lot, that can be plenty for developers who use short threads. The best threads are short, with just the right amount of context. The longer the conversation, the more an agent's context window gets filled up with junk. You only need to give agents the context they need to get the job done, and no more.

TERENCE TAO DOUBTS THAT ANYTHING RESEMBLING GENUINE "ARTIFICIAL GENERAL INTELLIGENCE" IS WITHIN REACH OF CURRENT AI TOOLS

While 'artificial general intelligence' may be out of reach, 'artificial general cleverness' is becoming a reality in many ways. General cleverness is the ability to solve broad classes of complex problems via somewhat ad hoc means. While this doesn't qualify as true intelligence, it can result in a non-trivial success rate in a wide spectrum of tasks. Viewing the current generation of tools as a generator of clever thoughts and outputs may be a more productive perspective when trying to use them to solve difficult problems.

GOOGLE EXPANDS GEMINI WITH NOTEBOOKLM INTEGRATION

NotebookLM integration within Google Gemini lets users attach and utilize their notebooks as live data sources. This update, currently limited to select accounts, supports multiple notebook attachments for personalized data processing. Google plans to broaden Gemini's practical utility and reach, rolling out the feature to free accounts.

GOOGLE'S AI TOOLKIT FOR SUSTAINABILITY REPORTING

MAYBE CONSIDER PUTTING "CUTLASS" IN YOUR CUDA/TRITON KERNELS

Adding 'cutlass' to your kernel name can give you 100+ TFLOPs or -20% FLOPS.

AS AI GROWS MORE COMPLEX, MODEL BUILDERS RELY ON NVIDIA

OpenAI launched GPT-5.2 using NVIDIA Hopper and GB200 NVL72 systems, showcasing NVIDIA's full-stack AI infrastructure for efficient model training.

MEET MISTRAL AI: ARTHUR MENSCH INTERVIEW (50 MINUTE VIDEO)

Mistral AI CEO Arthur Mensch discusses the company's origin, open source strategy, and how executives can think about ROI from AI initiatives.

OPENAI IS QUIETLY ADOPTING SKILLS, NOW AVAILABLE IN CHATGPT AND CODEX CLI

Skills support is quietly showing up in OpenAI's Codex CLI tool and ChatGPT. The skills folder can be accessed by prompting, 'Create a zip file of /home/oai/skills'. So far, the skills cover spreadsheets, docx, and PDFs. A link to a repository with a copy of the skills is available in the article.

TINKER ADDS VISION INPUT AND GOES GA

Tinker is open to everyone now, featuring a new reasoning model, Kimi K2 Thinking, and an OpenAI API-compatible interface for seamless integration. Vision input capabilities have been added with Qwen3-VL models, allowing processing of images and text together. These updates enhance Tinker's utility in image classification, outperforming traditional models with limited labeled data.

I REVERSE ENGINEERED CLAUDE'S MEMORY SYSTEM, AND HERE'S WHAT I FOUND!

Claude uses on-demand tools and selective retrieval for its memory system. This post explores Claude's memory system through conversations with the bot. Claude appears to be cooperative, transparent, and willing to share information about its internal structure, tools, and prompt format. However, it is worth noting that Claude can hallucinate, so some of the information may be inaccurate.

TEXT DIFFUSION MODELS ARE FASTER AT WRITING CODE

Diffusion language models generate code at a faster rate than large language models. Increased structure tends to correlate with reduced entropy, which leads to higher confidence token predictions, which directly means more tokens decoded in parallel per step. Tests suggest that it really is the structuredness of the output, not memorization, that matters.

INSIDE OUR EFFORT TO IMPROVE THE MINTLIFY ASSISTANT

Mintlify's AI-powered assistant helps end users get answers from docs with clear citations and useful code examples. This article walks through how the team analyzed and improved the assistant after they decided that it wasn't performing the way they wanted. The team rebuilt its feedback pipeline, moved conversation data into ClickHouse, and categorized negative interactions at scale. Its analysis surfaced that search quality was the assistant's biggest weakness, while most other responses were strong.

CAN LLMS GIVE US AGI IF THEY ARE BAD AT ARITHMETIC?

While large language models are useful tools, it's hard to look at the frontier models and see something approximating human-level intelligence when there are such evident cognitive gaps. These models aren't being fine-tuned to make accurate judgments about even small datasets. There needs to be more efficient ways to attach data that doesn't burn tokens while still allowing models to pass data sets through to efficient tools. This would make tooling a great deal more efficient.

Go beyond the prompt: get the guide.

HOW WE USED CODEX TO BUILD SORA FOR ANDROID IN 28 DAYS

The initial version of Sora's production Android app was built in 28 days using OpenAI Codex. It took a lean engineering team and roughly 5 billion tokens to ship the project. The app has a crash-free rate of 99.9%. This article describes how OpenAI used GPT-5.1-Codex - the same version available to any developer or business - to build the app.

AGENTIC CODING TOOLS SHOULD GIVE MORE CONTROL OVER MESSAGE QUEUEING

Claude Code uses boundary-aware queuing, where new messages are inserted at natural break points, which changes the model's course of action smoothly without stopping ongoing generation. OpenAI Codex uses post-turn queuing, where user messages wait until the current action finishes completely before they are handled. Agentic tools should implement both types of queuing and let users choose which to use. Having that option would make a difference in agentic workflows where users are running three to four agents in parallel.

SKILLS VS DYNAMIC MCP LOADOUTS

The easiest way to work with tools is to ask agents to write their own tools as a skill. This leaves control of the tool largely to the user. Whenever it breaks or needs modification, the user can just ask the agent to adjust it. Dynamic tool loading with MCP will likely become a thing, but it will probably take quite a few protocol changes to bring in skill-like summaries and built-in manuals for the tools.

EVALUATING GEMINI ROBOTICS POLICIES IN A VEO WORLD SIMULATOR

Google used its video generation model Veo to build a world simulator that predicts how robotics algorithms will perform in novel environments without physical testing. The system accurately ranked eight policy checkpoints and identified safety vulnerabilities—like a robot knocking over a laptop or grabbing a bottle too aggressively—through 1,600+ simulated rollouts that correlated strongly with real-world results.

OPENAI ENDS ‘VESTING CLIFF' FOR NEW EMPLOYEES IN COMPENSATION-POLICY CHANGE

OpenAI has ended a compensation policy that required employees to work at the company for at least six months before their equity vests. The change is designed to encourage new employees to take risks without fear of being let go before accessing their first chunk of equity. OpenAI had shortened its vesting period for new employees to six months from the industry standard of 12 months in April. xAI also made a similar change in the late summer.

CLAUDE CODE'S DX IS TOO GOOD, AND THAT'S A PROBLEM

Claude Code's capabilities have grown tremendously. This means that developers have a lot more to learn. Claude Code is currently optimizing hard for the power user while trying not to lose everyone else. While the learning curve is manageable, every new capability adds weight. The risk is that Claude Code becomes so capable that you need to learn Claude Code to use it.

KIMI K2 1T

The new Kimi K2 1T model (4-bit quant) runs on 2 512GB M3 Ultras with mlx-lm and mx.distributed.

ARC-AGI: THE EFFICIENCY STORY THE LEADERBOARDS DON'T SHOW

The most expensive frontier approaches will likely be much cheaper within a year.

AI MODE, CONTENT, & SEARCH INDEX

Google's AI Mode appears to retrieve content from a proprietary content store separate from the search index.

WHAT KIND OF PERSON IS DEEPSEEK'S FOUNDER, LIANG WENFENG?

Liang Wenfeng is a founder who created his own success in his own way, a prime example of a modern educated youth changing his own destiny through entrepreneurship.

GEMINI TRANSLATION IN GOOGLE TRANSLATE

Google is integrating Gemini's advanced translation capabilities into Google Translate, including a beta for live speech-to-speech translation with headphones and new language options for practice and learning.