OpenAI struck a deal with the U.S. Department of War to deploy its AI models in the military's classified networks, stepping into a vacuum left after Anthropic was blacklisted as a "supply-chain risk to national security" for refusing to remove safeguards against autonomous weapons and mass surveillance. The announcement triggered an immediate backlash, and Altman soon admitted "we shouldn't have rushed to get this out on Friday," announcing amendments to clarify the deal's safety guardrails. The fallout was swift: many users cancelled their ChatGPT accounts, and Anthropic's Claude climbed to the No. 1 spot on Apple's App Store for the first time. The episode crystallises a defining tension in AI: whether safety principles are firm commitments or negotiating positions when government power — and lucrative contracts — enter the room.

👉 Read All About It

Anthropic has launched a memory import feature for Claude that lets users transfer their personalized context — preferences, working styles, and project history — from other AI tools like ChatGPT in just two steps. The process is simple: run a provided prompt in your current AI assistant, copy the output, and paste it into Claude's settings. The pitch is compelling for anyone who has spent months "training" an AI to know how they work: you no longer have to start from scratch when switching. It's a savvy move by Anthropic to lower the switching cost from competitors, and a sign that persistent, personalized AI memory is fast becoming a key battleground in the AI assistant wars.

👉 Read All About It

Investor Helen Min argues that the real disruption to software isn't AI killing SaaS — it's the long-overdue death of seat-based billing, where companies have paid for software regardless of whether it delivered results. For years, vendors had no "skin in the game": a platform could promise to generate qualified sales pipeline, charge $100K+ annually, produce nothing, and still collect in full. AI agents are changing that equation, because they can now adapt and execute complex workflows fast enough to actually guarantee outcomes — making outcome-based pricing viable at scale for the first time. Min believes outcome-based pricing will mint the next generation of category-defining companies, just as usage-based billing created Snowflake and Databricks — and the winners will be those who move first while competitors are still selling hope.

👉 Read All About It

OpenAI has released two new models — GPT-5.4 and GPT-5.4 Pro — available via API, ChatGPT, and Codex CLI, featuring a 1 million token context window and a knowledge cutoff of August 2025. The headline capability is a sharp focus on productivity tasks: on an internal benchmark simulating spreadsheet work a junior investment banking analyst might do, GPT-5.4 scored 87.3%, up from 68.4% for GPT-5.2. Notably, the new model surpasses the coding-focused GPT-5.3-Codex across all relevant benchmarks, raising questions about whether a dedicated coding model line is still necessary. The business-productivity push puts OpenAI squarely in competition with Claude's recent enterprise ambitions — the AI arms race is increasingly being fought in spreadsheets and slide decks, not just chat boxes.

👉 Read All About It

A new research paper from Anthropic economists introduces a sharper way to measure AI's real threat to jobs, combining what AI can do in theory with what it's actually being used for in practice. The most exposed occupations include computer programmers, customer service representatives, and data entry workers — and jobs with higher exposure are projected by the BLS to grow less through 2034. The headline finding is reassuring for now: there is no systematic increase in unemployment for highly exposed workers since late 2022. But there's a worrying signal beneath the surface — hiring of workers aged 22–25 into AI-exposed occupations has dropped roughly 14%, suggesting AI may be quietly closing the door on entry-level careers before the overall job numbers show it.

👉 Read All About It

Codex for Windows just dropped. I’ve always preferred non-browser based IDE’s and vibe coding tools so I like the sound of this. Too bad I’m a mac user. Enjoy! 🏗

P.S., send your vibe coded tools to [email protected] and we might just feature them on the newsletter!

🇯🇵 On August 29, 1982, a Japanese company called Nippon Electric (NEC) released a home computer called the PC-8801 — and it quietly became one of the most influential machines you've never heard of.

Because Japan's internet infrastructure lagged behind the West, Japanese PC users in the late 80s built their own parallel universe of online culture called PC-VAN and Nifty-Serve — essentially Japan's own internet, years before most Japanese homes had real web access. On these networks, running largely on NEC machines, Japanese users invented many things we think of as modern internet culture: fan fiction archives, emoji-like emoticons called kaomoji (like (╯°□°）╯), flame wars, and early viral content sharing.

The kicker? A huge chunk of what we now call "internet culture" — memes, fandom, public oversharing — was independently invented twice, once in the West and once in Japan on completely separate networks, by people who had no idea the other group existed.