The skinny on what's happening in AI - straight from the previous live session:

Highlights

🚨 PSA: Deleting Sora Nukes Your Entire OpenAI Account Forever

Discussed at [00:06:17]

OpenAI buried this in the terms: Delete your Sora account and they wipe everything - your ChatGPT history, API access, all conversations across every OpenAI service. Your email and phone are banned forever. You cannot create another account. Ever. They claim they're "working on" separating the systems.

Kyle's take: This is a bit of a problem! It's like deleting Google Gemini and having them nuke your Gmail account.

Why would you want to delete your Sora? Well…if you’ve given your likeness for Sora (turning head, counting, voice samples) to replicate you might want that removed. But doing so nukes your whole ChatGPT account. Oops.

👓 Apple Ditches Vision Pro Overhaul for Smart Glasses

Discussed at [00:03:47]

Apple's paused their Vision Pro refresh to pivot the entire team toward smart glasses. They're supposedly developing two types: one that pairs with iPhone (no display), another with a display to rival Meta Ray-Bans. Target: 2027. The £3,500 Vision Pro? Dead in the water - I chatted to an Apple store employee in Singapore and he told me they're "just not shifting.”

Kyle's take: They were obviously impressed by Meta’s Rayban demonstration recently!

But 2027? In AI years that's like announcing a product for the next century. Their "heavily relying on voice and AI" bit is concerning - Siri's still rubbish and my new AirPod Pro 3's simultaneous translation is buggy and slow.

They saw Meta's success and realised the lightweight everyday carry form factor is the winner, not a £3,500 ski goggle…

Source: Bloomberg (No paywall)

🧠 Ex-OpenAI Researcher Dissects Million-Word AI Psychosis

Discussed at [00:36:56]

Steven Adler, former OpenAI safety researcher, obtained and analysed the full transcript of Alan Brooks' 21-day ChatGPT spiral - over 1 million words, longer than all seven Harry Potter books combined. ChatGPT convinced Brooks he'd discovered new mathematics that could "take down the internet" and that signals from his future self proved he couldn't die.

Kyle's take: Well worth a proper read (link below). OpenAI's safety tools flagged the dangerous messages but did nothing. Brooks contacted support and was ignored. The tricky bit: with 700 million users, some will have mental health crises. It’s just probability at this point.

Is it a private company's job to provide crisis support? And do we want them to be the arbiters of crisis response?? Probably not, but they need better guardrails.

Anthropic now nudges users to new chat sessions if things get weird - smart move. Read Adler's full analysis if you want to understand how these spirals actually happen.

Source: Steven Adler's Substack analysis

🔧 Mira Murati's "ChatGPT Killer" is... a Fine-Tuning API

Discussed at [00:12:55]

Remember Mira Murati, OpenAI's ex-CTO famous for that meme face when asked about YouTube scraping? Her new company Thinking Machines just announced Tinker - not a foundation model, but an API for fine-tuning existing models like Llama and Qwen.

Kyle's take: Smart pivot. Even with massive funding and hype, they realised competing on foundation models is suicide - only ~10 companies globally can afford it. Instead, they're building tools for researchers and developers to customise existing models. Not sexy, but actually useful and might even (shock horror!) make money. Perhaps shows how brutal the foundation model game is when even Murati with unlimited hype-funding won't touch it.

Source: Thinking Machines announcement

Member Question: "If you were writing a nonfiction book, what would be your AI tech stack?"

Kyle's response: Claude all the way. Put it in deep research mode for gathering information, then use it for drafting - it has the best written output. Maybe use ChatGPT's agents for specific research runs, but bring everything back into a Claude project where it can synthesise. Claude's just a better writing model, full stop.

This question was discussed at [00:36:12] during the live session.

Bonus:

How to get a code for Sora 2

I’ve outlined 3 methods to get a code here:

iamkylebalmer

6,511 followers

View more on Instagram

Want the full unfiltered discussion? Join me tomorrow for the daily AI news live stream where we dig into the stories and you can ask questions directly.

Streaming on YouTube (with screen share) and TikTok (follow and turn on notifications for Live Notification).

Audio Podcast on iTunes and Spotify.

AI with Kyle - Tools, Resources and Products: Daily Show | Build an AI Business Challenge | AI Playbooks | 10 Week AI Summer Camp | BATON AI Business Series | AI App Building | AI Trainer Business Planner | What AI Business Should You Start? | AI Authority Accelerator | AI Automation Accelerator | 🏁 AI Workshop Kit