TLDR AI 2025-11-12
ChatGPT Group Chats 💬, growing an RL environment 🌍, ElevenLabs Scribe v2 🗣
Yann LeCun Reportedly Planning to Leave Meta (5 minute read)
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 (2 minute read)
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.
Introducing Scribe v2 Realtime (3 minute read)
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.
OpenAI readies ChatGPT Group Chats with custom controls (2 minute read)
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.
Build times for gigawatt-scale data centers can be 2 years or less (3 minute read)
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 (27 minute read)
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.
Why Sudoku Variants Remain a Grand Challenge in AI Reasoning (15 minute read)
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 (18 minute read)
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.
RL Environments and the Hierarchy of Agentic Capabilities (23 minute read)
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.
👨💻
Engineering & Research
Cut the cost of IT complexity (Sponsor)
IT complexity can drive costs up across your company. Intelligent IT automation is key to taming technology chaos. Get the insights you need from the IBM Institute for Business Value. 👉
Read the IBV reportNested Learning Reproduction (GitHub Repo)
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.
Here's What's Next in Agentic Coding (28 minute read)
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.
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.
Anthropic Launches Use Case Library (20 minute read)
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.
Three AI Megadeals Are Breaking New Ground on Wall Street (10 minute read)
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.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email