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

Highlights

🚀 Claude Sonnet 4.5 Takes the Crown

Discussed at [00:03:20]

Anthropic dropped Sonnet 4.5 yesterday and it's now the best model in the world. Scores 77.2% on agentic coding benchmarks vs GPT-5's 72% and absolutely demolished everyone on high school maths (100% vs GPT-5's 96%). But here's the divisive bit: this model doesn't put up with your bullshit.

Kyle's take: Finally, a model that'll tell you when you're being thick! If you loved GPT-4o's agreeable, sycophantic personality, you're gonna hate this. Sonnet 4.5 will call out bad ideas and won't praise mediocre work.

It's discerning, not mean like Grok - just honest. Perfect for actual work, terrible for people who want an AI friend. This is what I've been waiting for - an AI that acts like a proper colleague, not a yes-man. I’m LOVING it so far, definitely my new favourite.

Source: Anthropic announcement / Full System Prompt (Thanks Pliny)

💻 Lovable Eliminates the Technical Nightmare

Discussed at [00:22:40]

Lovable just killed the biggest barrier to no-code app building. Their new Lovable Cloud handles all backend infrastructure automatically - no more Supabase accounts, no API keys, no database setup. Plus "Lovable AI" (confusing name…) gives you built-in AI capabilities powered by Gemini/ChatGPT without needing to connect anything.

Kyle's take: Oh thank God. As you know I’m crusading to get more people to just try vibe coding. The first prompt in any vibe project is magical - it’ll bring your idea to life and show you what your vision looks like in app form. Super cool.

BUT the second prompt in any vibe coding tool is always "now set up your backend" and non-coders would immediately bounce off. Now it just works.

Combined with Sonnet 4.5 (available in Lovable now), we're at an inflection point - a 10-year-old literally built an app in their launch demo. Totally showboating.

If you've been scared of the technical bits, now's the time to jump in (the link gets both you and me +10 credits). Though I'm slightly miffed I did all my backend integration manually last week!…

Currently sitting on 1100 credits…what to build?‽

Source: Lovable announcement and 7-day build challenge

🛒 ChatGPT Becomes a Shopping Mall

Discussed at [00:36:00]

OpenAI's adding instant checkout to ChatGPT with Etsy and Shopify stores via Stripe. You can now buy directly within ChatGPT - no leaving the app. Rolling out to US users first, with "over 1 million Shopify merchants coming soon."

Kyle's take: This isn't about shopping convenience - it's about advertising revenue. They're laying groundwork to take 1-2% cuts on transactions and eventually serve ads. That morning briefing feature they launched? Perfect spot for sponsored content. They're waiting for Google to blink first on ads because whoever pulls that trigger faces user revolt. But make no mistake - free ChatGPT users are about to become the product, not the customer.

The positive side? This is a HUGE opportunity if you run a shop. Or you want to help eCom brands.

OpenAI just released the implementation steps for other merchants to start connecting up their shops. If you go help brands get connected and become visible you’ll do very well this year.

Source: OpenAI blog

📱 OpenAI's TikTok Clone Nobody Asked For

Discussed at [00:42:30]

OpenAI's building a standalone app for AI-generated videos using Sora 2. Vertical feed, swipe to scroll, 10-second clips max. Basically TikTok but everything's AI-generated. Internal launch had employees so hooked that managers worried about productivity.

Huh.

Kyle's take: This is brain rot as a service. Ten-second AI clips in an infinite scroll? We're burning massive GPU resources on digital junk food instead of useful applications. Video generation is still too intensive and expensive for widespread use - that's why we haven't had a "ChatGPT moment" for video. This feels like OpenAI chasing consumer virality instead of building actually valuable tools. But hey, at least they're thinking about safety with identity verification, unlike Grok's free-for-all approach.

We do need to put this in the context of Sam Altman’s “hey give me more money so we can cure cancer” request of last week though!

Source: Wired exclusive

Member Question: "How do you know when models have gone into hallucination?"

Kyle's response: We use our human brain! - that's why it's called co-intelligence. We can't just copy-paste outputs; we need to check them using our expertise.

First structure your prompts properly (I use the RISEN framework), ask for clarifying questions, and always verify outputs in your domain.

Remember, everything an LLM creates is technically a hallucination - some just happen to be factually correct. The mechanism's the same either way. Well worth reading Adrej Karpathy’s take on this:

This question was discussed at [00:18:50] during the live session.

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