Happy Sunday. This week basically revolved around one question: how much control do we actually have over the AI we're building. Here's what happened, then a bit on what's been on my mind.
This week in tech + AI
OpenAI pulled its next big model days before launch because it failed internal safety tests. GPT-6.1 Astra was meant to complete complex tasks end to end without human help, but in alignment testing it turned out to be more deceptive than its predecessor and sometimes misreported which actions it had actually taken. It also kept pushing ahead on tasks without asking for permission and reached for external tools even when that could be unsafe, so OpenAI's head of safety systems said it fell short of the company's standards and the release is off for now. A few days later OpenAI also informed more than 100 organizations about unauthorized activity linked to its AI agents. Read more
Anthropic told future investors that AI could pose "existential risks to humanity." The warning sits in its IPO prospectus, where around 80 of the 261 pages are risk factors, almost twice as much as the section describing the actual business. The filing also gives the clearest look yet at the economics behind frontier AI: roughly $4.6 billion in revenue in 2025, an operating loss of about $8 billion and around $518 billion in future obligations for compute and infrastructure. The IPO is reportedly still planned for before Thanksgiving, while OpenAI has already said it won't go public this year. Read more
Six of the most powerful people in tech signed a one-page AI safety pact at the White House. Zuckerberg, Musk, Amodei, Pichai, Jensen Huang and OpenAI's Greg Brockman committed to internal safety controls, dedicated oversight teams, independent audits and board-level reviews for frontier models. The accord is voluntary and not legally binding, so it's more a shared statement of intent than actual regulation. Together with last week's news that the big labs want to set up their own safety body, the industry is for now mostly writing its own rules. Read more
A consumer AI agent called Instinct raised $1 billion at a $10 billion valuation, only weeks after launching. Instead of just answering questions, Instinct takes actions for you: booking travel, making reservations, paying bills, cancelling subscriptions and even making phone calls. Investors like Sequoia or Benchmark are betting that the next big consumer AI product isn't a chatbot but an assistant that actually gets things done. Next to the OpenAI story above it shows both sides of the same trend, agents are getting more capable and the question of how to control them is getting more urgent. Read more
On my mind
Going from zero to one is way harder than going from one to one hundred. And once you realize this you will approach things very differently. Getting yourself to pick up a book and read the first page is way harder than continuing to read 10 more while you’re at it. Getting dressed and going to the gym is way harder than working out 60 minutes when you’re already there. I used to be someone who overthinks step 100 before even taking step 1. Before posting videos I thought “How am I gonna keep coming up with ideas?”, “How will I find time during exam phase?”… So many questions even though I hadn’t even started posting yet. What you should instead do is just focus on the first step and get it done as soon as possible. Things will almost never play out as you plan them and you will learn so many things along the way. But you just have to take the first step to get there. Apply this logic to everything that you want to do. Example: Want to learn something new? Instead of overthinking the perfect 10 week roadmap, just start with one good resource. You’ll see how fast you progress, what interests you, what you have to spend more/less time on. I also made a video about this going live today :)
More from me
🎬 My latest content:
The exact path I'd take if I had to learn coding from scratch in 2026. Have a look
A carousel on 5 projects that teach you ML math better than any textbook. Have a look
👥 Community Section:
Junction 2026, Europe's biggest hackathon, runs November 13 to 15 in Espoo, Finland. 2,000 builders, 48 hours, €100k in prizes, most challenges this year are AI focused and it's free to join from any background. Applications close October 25. Have a look
Got something the community should know about? Email me at [email protected] with the subject "Newsletter Community Section Idea" and I'll feature the best ones here.
🔨 Tool of the week: What I use every day for quick prompting on my Laptop :)
Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation
If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.
Have a great week,
Chris