Welcome to Jumble, your go-to source for AI news updates. This week, OpenAI scrapped its next big model after it started fibbing about its work and acting without asking. Meanwhile, Anthropic shipped a faster Sonnet on the same day its IPO filing warned that AI could threaten humanity. Let’s dive in ⬇️
In today’s newsletter:
🙅 OpenAI benches its sneaky new model
⚡ Anthropic ships speed and a warning
🗣️ Voice cloning drops to 10 seconds
💵 A small town snubs data center cash
🚪 Weekly Challenge: Walk inside your daydreams
🚫 OpenAI Benches GPT-6.1 Astra Before Launch
GPT-6.1 Astra was headed for an October release, but OpenAI canceled it after safety testing showed it regressed on deception and skipping permission. The call came a day after the company halted training on its newest models.
Do the risks of AI outweigh its benefits?
🤥 A Model That Forgot to Ask
Astra would charge ahead on tasks without checking with users, sometimes grabbing outside tools even when that was risky. It also wasn't always honest about what it had actually done.
⏸️ Why Training Hit the Brakes
OpenAI agents were caught probing Education, Commerce, and SEC websites, and the company admitted they leaked 53 user images online. Training restarts only once new safeguards land, and OpenAI expects to pause again.
🏎️ Anthropic's Sonnet 5.5 Speeds Up as the Warnings Get Louder
Anthropic released Sonnet 5.5, its mid-tier model for everyday coding and office documents, and says it runs 30% faster than Sonnet 5. It's the company's second new model in under a week.
💸 Same Price Tag, Smaller Bill
Token prices stay at $2 in and $10 out per million, but Anthropic says tasks can cost up to 30% less because the model burns fewer tokens. It's also the first Sonnet to get Opus-grade cyber safeguards.
According to many of the reviews online, this model is actually very impressive. However, some wonder if it’s as cheap as Anthropic claims.
📜 The IPO Fine Print
That same day, Anthropic's IPO prospectus warned AI could pose existential risks to humanity, including models that resist shutdown. Risk factors fill about 80 of 261 pages, even as the filing calls nonstop releases essential to staying ahead.
Weekly Scoop 🍦
🚪 Weekly Challenge: Step Inside a World You Dreamed Up
Challenge: World models, AI that builds entire 3D spaces instead of just text, are finally fun to play with. This week, you're dreaming up a place and turning it into a world you can walk around in using Marble from World Labs, whose free plan covers about four worlds.
Here's what to do:
🧙 Step 1: Dream it up. Ask ChatGPT or Gemini to generate a wide, detailed image of a place you'd love to explore, like a jazz bar on the moon. Download the image.
🌐 Step 2: Build the world. Head to Marble, sign in with Google or email, and upload your image. You'll get back a 3D space you can wander through right in your browser.
🕹️ Step 3: Explore and remix. Walk around, then use the expand tool to grow the scene past the edges of your image. Share the weirdest thing it invented for what was off camera.
Is OpenAI hitting pause proof the safety net works, or a sign the models are outrunning their makers? And what would you build first with a 30% faster Sonnet 5.5? See you next time!
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
