Welcome back! Here’s a wild stat to start your day: Anthropic engineers now ship 8x as much code per day as they did a few years ago. And more than 80% of the code merged into Anthropic’s codebase is written by Claude, not humans.
The company is openly mapping the path to "recursive self-improvement.” Basically, AI capable of designing and building its own successor.
Are we just watching tools get better, or are we watching ourselves get written out of the loop? Hit reply and share your thoughts.
Do You Actually Need a Frontier Model for Everything?
Here's a take I think more people need to hear: You should probably have a local AI model on your computer.
Most of the time when you're asking AI to summarize a document, rewrite an email, organize your voice notes, or brainstorm some ideas, you don’t need a state-of-the-art frontier model to do it. Smaller local models are getting good enough to handle a huge percentage of those everyday tasks, and the hardware to run them is finally catching up to the demand.
Here’s an example (not sponsored, just relevant) ➡️
NVIDIA recently unveiled the RTX Spark Superchip, a Windows on Arm platform that combines CPU and GPU capabilities with up to 128 GB of unified memory. Translation: It's becoming genuinely possible to run REALLY capable AI models directly on your own device.
The benefits stack up quickly:
Privacy. Your prompts, files, and conversations don't leave your computer.
Offline access. On a flight, in a remote workspace, in the middle of the desert if that's your vibe—your AI keeps working.
Cost. Likely cheaper per token in the long run, especially if you're building with AI.
Independence. You're not tied to the financial, political, or personal whims of any single AI lab.
And here's the angle I find most interesting given everything happening right now: This ties directly into the data center debate.
A lot of people are concerned (and fairly so) about the energy use, water consumption, and rapid pace of new data center construction. Local models won't eliminate the need for data centers entirely—frontier-class workloads still need them. But every task that runs on your laptop is one less request flowing through a massive server farm.
My POV: This is one of those shifts I think a lot of people are sleeping on. The "AI = ChatGPT in a browser tab" mental model has dominated for the last couple of years, but the hardware is getting cheap enough and the small models are getting good enough that having your own local setup is going to feel completely normal pretty soon. Worth getting ahead of.
So…did I convince you to try out a local model yet?
— Matt
OpenAI Files for IPO, Right Behind Anthropic
Via TechCrunch
OpenAI confidentially filed for an initial public offering on Monday, the company announced in a blog post, just over a week after rival Anthropic filed its own draft registration. The race to the public markets is officially on. 🔥
The money question:
OpenAI is racing to IPO even as it recently missed its internal benchmarks for both user growth and revenue (it was pushing for 1B weekly ChatGPT users by the end of 2025).
It raised $122B in March (the largest funding round in Silicon Valley history) but projects burning $85B in 2028 alone…meaning it's asking investors to back a business that won't be cash-flow positive for years.
Anthropic's contrast: Anthropic says it's close to its first quarterly profit and recently surged past OpenAI to a $1 trillion valuation on secondary market Forge Global. And the fact that Anthropic filed for an IPO first means it gets to set a valuation comp that impacts how OpenAI can price its own offering.
Why it matters: Three of tech's most-watched companies—OpenAI, Anthropic, and SpaceX—could all go public within months of each other, soaking up what's becoming increasingly scarce AI capital. Whoever debuts first likely grabs the most, which is why the order of these filings is as strategic as the filings themselves.
This Conference Season’s Theme? The Agents Are Here
The season's last big keynote belonged to Apple. At WWDC on Monday, it unveiled a rebuilt, conversational Siri that can act inside apps (two years after first promising it). But the biggest takeaway from a packed season of dev conferences: The chatbot era is over, and every giant is racing to sell you agents.
The breakdown:
At Google I/O, Google launched Gemini 3.5 and Gemini Omni, Gemini Spark (a "24/7 personal AI agent"), and Universal Cart for agentic shopping.
At Microsoft Build, Satya Nadella unveiled seven in-house MAI models, plus Scout, an always-on autonomous agent Microsoft dubs an "Autopilot."
Apple's entry: a Siri that actually does what you tell it to, plus a rumored AI agent layer for the App Store.
The bigger picture: Everyone is selling the same future (agents), but they're attacking it from different angles: Google from search and scale, Microsoft from the enterprise stack, Apple from the device in your pocket. The agents are here. But who will you trust to book the reservation or read your inbox?
Run your socials on autopilot
Via Echos
Echos is an AI-powered short-form distribution platform that uses autonomous agents to create, warm, post, and manage fleets of TikTok and Instagram accounts. It discovers viral content, spins up hundreds of persona-driven variations, and optimizes against view goals—no extra headcount required.
How you can use it:
Spin up and manage multiple TikTok and Instagram accounts at once
Auto-generate hundreds of creative variations to test what lands
Discover and adapt trending content before it peaks
Track performance and optimize campaigns toward guaranteed view targets
Pricing: Paid
Architecture docs that keep themselves honest
Via Archyl
Archyl is an AI platform for living software architecture documentation built on the C4 model. It analyzes your codebase, auto-discovers components and relationships, generates interactive diagrams at every level, and flags drift so your docs never quietly fall out of sync with reality.
How you can use it:
Auto-generate interactive C4 diagrams from your actual codebase
Detect architectural drift before docs go stale
Express architecture as code (YAML DSL) for versioning and reviews
Plug into Git workflows so diagrams get reviewed like any other change
Pricing: Paid
Jobs, announcements, and big ideas
Anthropic launches Claude Fable 5, a new Mythos-class model built for broad general-purpose use.
Google adds near real-time voice translation to Gemini 3.5 Live, breaking language barriers in conversation.
Midjourney begins sending invites for its first hardware launch, teasing a move beyond image generation.
OpenAI launches an Economic Research Exchange to study AI's impact on jobs and the economy.
Last week’s Microsoft Build was one for the books! Here’s the breakdown of everything you need to know in 82 seconds ➡
That’s a wrap! See you Friday for more.
