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

Highlights

💻 Vibe Coding's Creator Says It Can't Cut It? Not Quite

Discussed at [00:01:05]

Gizmodo ran a headline claiming Andrej Karpathy (who coined "vibe coding") admits it doesn't work, because he hand-coded his new Nano Chat project instead of using AI. Nano Chat is a minimal LLM training pipeline - 8,000 lines of "clean code" for building ChatGPT-style models for $100.

Kyle's take: Gizmodo being misleading here - and it's a bit embarrassing for a tech publication to be this anti-tech.

Karpathy's pushing the frontier of AI research. He's in the 0.01% doing genuinely new work. Vibe coding is brilliant for the other 99.99% of us reinventing wheels - building web apps, MVPs, proof of concepts.

When you're creating “DuoLingo but for job applications” or “Uber for pets” vibe coding works perfectly. It can replicate patterns and build out products fast.

When you're inventing new LLM architectures? Of course you need to hand-code. Karpathy himself literally says vibe coding is for "throwaway weekend projects." Different tools for different jobs. This isn't hypocrisy - it's common sense that Gizmodo ignored.

Source: Gizmodo / Andrej Karpathy's NanoChat GitHub

📸 Charities Using AI-Generated "Poverty Porn" for Fundraising

Discussed at [00:07:37]

Guardian investigation reveals NGOs are using AI-generated images of extreme poverty, abuse victims, and war zones in marketing campaigns. FreePik alone has 63,000+ AI-generated poverty images.

Kyle's take: The intentions might be good - raising money to help real people - but this is morally grey at best.

They cite "consent concerns" and "cost" as reasons. And to be fair, US funding cuts to NGOs made it worse - they're looking to cut costs (and cornerS).

FreePik's CEO says responsibility lies with consumers, not platforms, and that there’s nothing he can do. Rubbish - he could ban AI poverty images tomorrow if he wanted to. He doesn’t.

What's truly bizarre here is charities are PAYING for AI-generated images when they could make them free themselves…

Source: Guardian exclusive

🧠 Claude Skills for Token-Efficient Expertise

Discussed at [00:14:54]

Simon Willison's analysis of Claude's Skills feature - basically markdown files with instructions that Claude only loads when relevant. Each skill takes just dozens of tokens for description, full details loaded only when needed.

Kyle's take: Anthropic announced this cool feature and walked away - typical! No press conference, no big hoorah. Very different to OpenAI!

Skills are genius in their simplicity. Imagine 20-30 skill folders - tone of voice, brand guidelines, compliance rules. Normally you'd dump everything in system prompts, clogging your context window. Now Claude scans brief descriptions of all skills, pulls full details only when needed. Like having appendices you only read when relevant.

I'm more excited about this than MCPs - they consume tens of thousands of tokens just sitting there. GitHub's official MCP alone eats precious context before you even start working. Skills are portable everywhere - chatbot, API, mobile.

I'm building loads more this week, including making a “Role” skill that accesses lots of specific skills underneath, allowing for AI employees to be constructed.

Source: Simon Willison's blog 

📚 Ethan Mollick's Updated AI Guide: Just Pick One

Discussed at [00:22:21]

Ethan Mollick has updated his fantastic guide. His (purposefully!) opinionated guide says stick to the big three: ChatGPT, Claude, or Gemini. They all have the same features now - voice, images, code execution, research modes. For general use: Claude Sonnet 3.5, Gemini 2.0 Flash, or GPT-4o Auto. For heavy thinking: Claude Extended Thinking, Gemini 2.0 Pro, or GPT-4o Thinking.

Kyle's take: At the end of the day, this is subjective. I use Claude because I like how it thinks and talks. Some people love ChatGPT. That's fine - benchmarks don't matter if it doesn't work for YOU.

The skill is knowing when to switch models within a particular tool. Need a lasagna recipe? Standard model, get answer in seconds. Need 600-source medical industry research? Extended thinking, wait 30 minutes. Part of learning AI is keeping up with which model to use when. Send this article to anyone asking "what AI should I use?" - it'll save you explaining! Certainly what I’ll be doing!

Source: Ethan Mollick's "Opinionated Guide to Using AI"

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.