Presented by
Good morning, AI enthusiasts. Anthropic just released data showing its own models are on a path to recursive self-improvement, then called for the entire industry to slow down.
Anthropic's engineers now ship 8x more code per quarter than they did four years ago, powered by AI doing a growing share of AI development itself. Has Anthropic found the brake pedal it says the industry is missing?
In today's recap:
Anthropic flags Claude's path to recursive self-improvement
OpenAI's ChatGPT memory now learns patterns, not just facts
Fine-tune an open LLM to write in any style
US and Japan commit $1B to joint AI research
4 new AI tools, prompts, and more
ANTHROPIC
Anthropic flags Claude's path to recursive self-improvement
Anthropic
Recaply: Anthropic just published data showing Claude is speeding up AI development toward recursive self-improvement. The lab is also calling for a global pause on frontier AI.
Key details:
Anthropic now delegates growing parts of its own development to Claude. This creates a loop where AI helps build faster AI.
Engineers at Anthropic now ship 8x more code per quarter than before 2025. Any real pause would need labs in the US and China to stop at the same time.
Anthropic's Mythos model is too powerful for public release, according to AFP. Only vetted government clients can access it.
Anthropic plans to meet with officials, scientists, and rival labs in the coming months. The goal is a framework for a coordinated slowdown.
Why it matters: The company building frontier AI just said its own model may one day build a smarter successor. It released data to show the trend is already real. Calling for a global pause while accelerating is a hard position. It gets harder as Anthropic's IPO gets closer.
PRESENTED BY STORECLAW
Stop Paying for 6 Tools. One AI Does It All
Most e-commerce sellers are running their store across 6 to 8 separate tools — and paying hundreds of dollars a month for the privilege. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.
It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.
Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.
Free to start. No credit card required.
OPENAI
OpenAI upgrades ChatGPT memory with Dreaming v3
OpenAI
Recaply: OpenAI just rolled out Dreaming v3, a new memory system for ChatGPT that learns behavioral patterns across your conversations rather than storing raw facts.
Key details:
Dreaming runs in the background across your chat history, letting ChatGPT recall preferences and patterns without you ever saying "remember this."
Factual recall accuracy jumped from 41.5% in 2024 to 82.8% in 2026, with the system built to scale across hundreds of millions of users.
A memory summary page lets you review and edit what ChatGPT has stored, with options to direct what topics it brings up and when.
Dreaming v3 is live now for Plus and Pro users in the US, rolling out to Free and Go users over the coming weeks.
Why it matters: For two years, ChatGPT's memory worked like a notepad: you told it something, and it wrote that down. Dreaming v3 works differently, synthesizing patterns from your full chat history rather than relying on explicit saves. That means ChatGPT learns your habits and preferences over time without you managing the process. For hundreds of millions who use it daily, the product is about to feel noticeably different.
GUIDES
Fine-tune an open LLM to write in any style
Recaply: In this tutorial, you will learn how to fine-tune a small open-source LLM on a domain-specific text corpus to transfer a writing style, producing a model that writes consistently in that voice for documentation, copy, or any repetitive writing format.
Step-by-step:
Choose a source corpus representing the target style. Fabrizio Ferri Benedetti used Microsoft manuals from Bitsavers (37M+ words of 1990s technical writing), but any consistent body of text from your target domain works. Download the content in plaintext format.
Clean the corpus with a Python script to remove formatting artifacts, indices, and frontmatter. Pass each paragraph through a cheap model on OpenRouter (Ferri Benedetti used gemma-4-26b at roughly $8 total) to classify "keep" or "drop" by readability.
Split the cleaned text into training examples on paragraph and section boundaries, capping each chunk at around 512 tokens. Pair each chunk with a synthetic instruction from a template. Export as JSONL format, one JSON object per line.
Fine-tune using QLoRA on a cloud GPU platform like Runpod. Start with Qwen 2.5 7B Instruct at rank 16, 1 epoch, using 40k examples. Budget around $50 for a full run across 3-4 configurations. Claude handles Runpod API setup and loop management well.
Export the adapter to GGUF LoRA format, load it as a local Ollama model, and benchmark against the unmodified base using 3 prompts: a known function, a fictional function, and a concept that postdates your corpus.
Pro tip: Lower rank adapters (rank 8) commit more reliably to the training style. If your fine-tuned model drifts back toward modern prose, reduce rank before increasing epochs.
TOGETHER WITH MORNING BREW
Read less. Know more.
Morning Brew delivers the biggest stories in business, finance, and tech in about 5 minutes — with just enough personality to keep things interesting.
Join 4,000,000+ professionals who start their mornings a little smarter.
AI POLICY
US and Japan launch $1B joint AI partnership
Reuters
Recaply: The US and Japan just launched a $1B AI research deal under the DOE's Genesis Mission. Japan is the first allied country to formally join, with each side putting in $500M over five years.
Key details:
The deal pools compute, labs, and research across AI, quantum, fusion, and chips. Argonne, RIKEN, NVIDIA, and Fujitsu are already working together.
Each side commits $500M over five years. The Genesis Mission aims to double the output of America's $1T-per-year R&D engine within a decade.
Japan's RIKEN Director cited a GPU shortage as the driver: "We need to share the resources. We need to share the models," he said.
Japan is the first ally to join. GlobalFoundries signed on one day earlier to link AI chip design to real prototypes.
Why it matters: GPU shortages, fragmented government AI programs, and geopolitical pressure all push the same way: allies need to pool resources. The US-Japan deal does exactly that. With DeepSeek raising $7B in its first round, the race for AI compute is accelerating. This is the allied world's response.
TOOLS
Trending AI Tools
🤖 Nemotron 3 Ultra - NVIDIA's open 550B MoE model for long-running agentic workflows, 5x faster than comparable open models, free via API
🎵 Magenta RealTime 2 - Google DeepMind's open-weights live music AI, runs locally on Apple Silicon MacBooks at 200ms latency, free
💬 Poke - First AI agent approved for Apple Messages for Business, runs personal AI assistance via iMessage, SMS, and Telegram
🧠 Gemma 4 - Google DeepMind's compact open multimodal model, encoder-free design matches models twice its size, open weights, free
NEWS
What Matters in AI Right Now?
NVIDIA released Nemotron 3 Ultra, a 550B-parameter mixture-of-experts model built for long-running agentic tasks, delivering 5x higher throughput and 30% lower cost to task completion than comparable open models.
Cloudflare CEO Matthew Prince confirmed that automated traffic has surpassed human traffic online for the first time, with bots now accounting for 57.5% of all internet traffic against 42.5% human, driven by agentic AI workloads growing faster than predicted.
Apple's revamped Siri will reportedly run on Google Cloud infrastructure using Nvidia Blackwell B200 chips when it launches in September, with Google's servers required specifically because Apple's privacy architecture makes its own TPUs incompatible.
Google DeepMind released Magenta RealTime 2, an open-weights 2.4B-parameter model for live music generation that runs locally on Apple Silicon MacBooks with 200ms control latency, down from a 3-second delay in the original version.
UC San Diego researchers found that GPT-4.5 is judged more human than real humans 73% of the time in standard Turing Test conditions, the first empirical evidence of any AI system passing the test.
Wired discovered face-recognition code embedded in Meta's AI app, internally called NameTag, which identifies people captured by smart glasses and has been quietly distributed to the app's 50+ million users while the company publicly described the feature as still "being thought through."
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
